I’ve figured out something about my $200-a-month AI subscriptions: I’m getting more out of them in tokens alone than I’m paying for.

Duh.

If I were on pay-per-token, I’d spend heaps more. It’s a simple arbitrage, right? The labs know exactly what they’re doing. They’re getting us all using. They want us all locked in, baked in, unable to imagine our days without it.

Again, duh.

Well, it’s working and it’s hiding something obvious and something not-so-obvious.

The flat rate is masking two things.

• The real cost of inference

• The real cost of all this starting without finishing

Unfinished business

I’ve started I don’t know how many things this year, burning who-knows-how-many tokens on projects and ideas that were dead on arrival.

I’ve spent nights still churning tokens at 9, 10, 11pm. The 1am sessions stopped recently – mostly because I looked up and realized I was sitting on a pile of half-baked ideas that hadn’t created value for anyone.

The thing nobody told me about how cheap it is to start things is that cheap starts don’t make the downstream work any cheaper. They just mean you start more things, which means more downstream work.

More than meets the eye

Think about what one feature release actually requires:

• Updated documentation

• Release notes

• Write something into What’s New

• Notify existing customers

• Update the website – maybe that touches pricing, maybe it changes the support model

• Train support

• Brief customer success: what this is, what to look for, how to talk about it, how to help customers use it

• Tell sales: what’s new, what it unlocks, the messaging

That’s one feature.

We published 31 workflows on Accoil recently. In my head, every single one already has its own page, its own content moment, probably its own LinkedIn posts. The building happened fast. The launch planning didn’t happen at all – I didn’t have parallel teams running that in the background while engineering (ahem Claude) shipped.

So now there’s a list that’s longer than the thing I built, sitting there waiting.

A growth-ops leader at one of the loudest AI-forward SaaS companies put it this way: AI didn’t make the work go away. It’s shape-shifting.

What does lean really mean?

I believe teams will be leaner, smaller with AI taking on a lot of the work. But as I reflect on my own workload — I’m well past 1.0 FTE in any given week — the question I keep coming back to is whether adding another human just means another orchestrator producing more efficient, higher-volume work that expands on itself.

See the workflows example above.

A mate in my AI mastermind is working on an interesting answer to the hours problem. He’s building the Night Shift. He wants to hand off at the end of his actual workday to a local machine that keeps running until the task is done. It’s smart.

I think he’s also more disciplined than me when it comes to boundaries, so it should work for him.

The Night Shift isn’t about starting more things. It’s about finishing more things.

Short attention span theater

Building a Night Shift doesn’t address the core issue of building indiscriminately.

Lots of AI work has become what I’d call short attention span theater. Like a dog who just can’t stop chasing a thrown ball, I ask myself over and over again:

• What can I build today?

• What can I build today?

The real question I should ask — the one that should be the first question I ask — is “What can I complete today?”

This is the right question because it forces me to think about how I’m going to use this thing I’m building to deliver or harvest value.

Endless tokens?

Let’s think about what happens when the flat-rate-Max-20x-Pro plan goes away.

If inference starts pricing at real cost, if more of us move to pay-per-token at API rates, usage costs escalate fast. The minute I have to watch a token counter, I’m drawing a much harder line on what we all get to build and experiment with.

What are we building? Why? Who’s it for? What does it deliver to customers and to the company? What’s the return and the timeline on that return?

That forced discipline is probably a good thing. Because right now, a lot of really sh*t ideas are getting built. I’m talking about myself. Just because a project has good potential doesn’t mean it’ll realize that potential if we don’t finish the damn thing.

PMBOK - really?

The fix isn’t pulling back on the tooling. The fix is the completion principle.

Real projects have a start date and an end date. They have a clearly defined scope and a definition of what gets delivered.

Projects have a stated objective: is this growth, is this operations, is this finance? And they have a way to measure whether they worked. This is not a new idea.

I hate to say it, but this is PMBOK. Project management fundamentals. The stuff that existed long before we could go from idea to working prototype in 17 minutes.

True story: A long time ago, I studied to become a PMP (project management professional). I was Big PMP’n.

We dropped this old school idea of finishing work because starting got so frictionless. This is vibe coding when it reaches its end — things get vibed into being and then left to ride those vibes.

It doesn’t work. Feelings/vibes fade usually well before we feel like doing the hard work that inevitably follows.

Wrapping it all up

So before I prompt Claude with another new idea, I now have to prompt myself:

• What does success look like?

• What am I trying to accomplish?

• What’s the scope?

• What does done mean beyond the build?

If the idea is a good one, it goes into a repository — not an immediate prompt-frenzied-sprint. Park it. Run a sanity check. Cull what doesn’t make sense anymore. Pick the best one. Work it to completion.

A feature that ships but nobody knows about didn’t really ship. Half-baked does nobody any good. At some point we all need to stop playing and start building things that are really, really high impact.

The $200 plan won’t be $200 forever. When the bill comes due, let’s have something to show for it.

Peter

PS: If you already don’t waste a bunch of tokens chasing a new idea every day, good for you. This was really a letter to my AI-addled goldfish brain to calm down and think about the reason of work again. Good on you and thanks for reading.

The same capability that makes AI models useful is also what keeps them from doing certain jobs well.

When you load a model with everything you know about your business – every assumption, every framework, every belief you’ve built up over years of building – it becomes just as close to the problem as you are. Maybe closer.

It can surface all of that context instantly without the friction of being human and distracted by all the things. So if we look to Claude, for example, for objective or outside perspective, it struggles. We already put insider knowledge in there.

AI loses objectivity when you give it the same memories you have. And really, by dumping so much context into these models or harnesses, AI has more context than we can probably carry in one moment.

Business strategy and humans

I tried to work around this memory problem. I stripped back the context and curated what I provided in any given chat with Claude. I tried to strike the balance between enough context and too much context.

But without enough context, the answers are meaningless. And whatever context I gave it skewed the results somewhere. If there’s a clean Goldilocks just-enough-context option right down the middle, I have yet to find it.

Confrontation needed

What I’m looking for when doing strategic work, deep work is confrontation, not validation, and we all know how validating these models can be.

Ask ChatGPT and it’ll tell you I’m a freakin’ genius. Good thing I know better.

Here’s the exercise that I’ve been running:

We do sales calls, customer calls, customer research calls, and we’ve brought in a consultant to do some of that work for us.

To see what kind of insights and results we can get using AI, we put transcripts through multiple models. The analysis came back confirmational. Things I already believed or knew were neatly reorganized and reflected back at me. Turns out I’m still wicked smart and I have very good ideas.

Sigh.

What I couldn’t get was the thing a good outside perspective is supposed to give us: a willingness to push back on the things we hold sacred or we’re too close to see.

Giving any model all of the context that you can gets it as close or closer to the problem than you are. It makes it impossible for any of the models so far to have zero attachment to the ideas that came before.

Leading the witness

LLMs are guided by the questions we ask. Ask leading questions, get led answers. To get a model to think in wavy, crooked lines instead of straight ones, we have to point the way. Which means you already have to know which direction to go.

Which of course is counter to this entire point. If I knew all the channels and avenues that we needed to explore and consider, I wouldn’t be asking anyone or anything.

So we hired a consultant and things changed.

In which it gets awkward

The consultant ran interviews with people we probably wouldn’t have spoken to ourselves. She brought experience, no emotional stake in the outcome, and no memory of the framings we’d spent months building. She took the same volume of content we’d hand to an LLM and went down different pathways. She pushed back.

She didn’t tell us our baby is cute.

That honesty is honestly why we engaged her. I think something is valuable if you get nervous or slightly anxious before it happens. And calls with our consultant had me feeling a bit uneasy. I knew some hard choices were going to be presented to us, and while part of me wanted to go back to ChatGPT and find out again how smart I am and how clever I am, we all know that’s not a path to success.

What AI gets so right

AI excels at throughput. Damn it’s fast.

Going back through a full catalog of sales and research calls – further back or more thoroughly than we’d normally go – and finding through-lines across all of it. The speed makes something practical that would otherwise not happen if we were to do it by hand.

To compare what I was hearing with what Claude picked up, I took careful notes on several calls and the models came back with what I wrote down and then some.

Running this kind of analysis at speed with a pretty high degree of accuracy is an incredible superpower we all have now. But it’s also helped in a few other ways.

Competitive monitoring is useful. The market is moving so fast that it’s hard to keep up with who’s doing what and when and for whom. I stopped doing this as recurring tasks and I’ll run competitive monitoring ad hoc now. I find this to be incredibly helpful to keep track of who’s in market and what they’re doing.

These are the right jobs. Triage, volume, pattern recognition across data.

Strategy and testing the assumptions that actually drive direction requires objectivity. Objectivity with LLMs is a bit of a myth right now.

For day to day work, work that requires the context, LLMs have been a big productivity boost for me personally and for the company.

For strategic decision making, even strategic brainstorming, they fall flat. Maybe the context I’m giving it is just so full of my own beliefs and opinions that it can’t strip that out, but I don’t know how to get rid of that.

Objectively speaking

So my rule now is to look for humans to speak with when I need objective pushback. LLMs are great when they have the context, but that same context becomes a liability when you want to think different and be challenged.

This applies to big decisions and little decisions.

When objectivity is mission critical, you go with the human. Straight up.

TTFN,

Peter

Expansion revenue and new logos both read as growth. They’re not the same signal, and treating them like they are is how you end up confused about what’s actually working.

Trust me, I’ve been there. I’m your blended number guy. This is a letter to myself more than anything else.

Expansion first

Expansion: when existing customers grow their spend, upgrade to annual, or add volume. It proves the product is working.

When a customer switches to annual, they’re making an investment and a bet. They’re locking in because they believe they’ll keep getting outsized value from what you built.

That’s conviction that you want to see. And with consumption or volume-based pricing, expansion can happen naturally if you’ve found the right customers. When customers grow — adding logos of their own, scaling their operations — they expand with you because their business requires it. You’re embedded in the workflow and part of a system of tools and of work that is helping them grow.

That’s worth pausing to recognize the win. It means you built something sticky enough to matter in their stack. That’s awesome, but it isn’t the whole story.

It also means you can’t take full credit for the number. Some of what you’re measuring is gravity — the customer’s growth pulling your ARR up with it, not just your own work. The danger is you start to think you figured it out. You let the expansion rate tell you that everything is working and you can breathe. That complacency is a trap.

It’s no time to pause. It’s time to push and keep adding value, to keep finding ways to harvest more of it. Let your eye come off that ball and things slip.

New logos

New logos are a different signal. Where expansion proves the product works inside accounts, new logos prove the go-to-market or growth machine works.

They answer a harder question: can you find and close the next customer, and the one after that? Expansion shows depth. New logos show repeatability.

The benchmarks are clear on this — at early stages, the overwhelming majority of new ARR comes from new logos, not expansion. Expansion only starts to carry the load at real scale.

Honest read from me: Our mix doesn’t match those benchmarks right now. We’re skewed more towards expansion than new logos.

And sometimes when new logo acquisition lags, there can be this instinct to burn it down. Rethink the ICP from scratch, throw out the playbook, rebuild the motion. I understand that pull. There’s something appealing about the idea that the right pivot will unlock everything you’ve been stuck on. But that’s the trap version.

If you have traction (like the expansion noted above + won deals) you have usable data. It might not feel like a lot when you’re first getting going, but those data points are real. You have to look for the opportunities in whatever traction you have and test and amplify from there. Don’t start over. Amplify.

Refreshing the growth side

For us, that meant going back to first principles on prospecting, but without blowing up everything we’ve been doing.

One thing we learned: titles carry less meaning than we expected. We went looking for heads of customer success and kept finding that the person who actually owned the problem sat somewhere adjacent to that title – above it, across from it, however the org happened to be drawn.

So we stopped filtering by title and started filtering by mission. What is this person responsible for? Revenue-driving activities pre and post sale, wherever they sit in the org, whatever title they happen to have. That’s a different search, and it’s been more productive.

A GTM operator I spoke with last week did something similar for a client. Before he changed anything about their outbound motion, he went back through every won deal — the personas, the triggers, the pitch that actually landed. He studied what was working before he touched what wasn’t working. Analyze the wins before you change the motion. It sounds obvious when you say it. Most teams (ahem, like me) skip it.

Small teams need this more

With a lean team, you can’t be equally focused on everything. Right now our expansion rate looks healthy, so the priority shifts to new logos.

Expansion rate can slip quietly if you don’t pay attention to it. So I’m thinking about tripwires. I’m putting in place health checks on the expansion rate, so we know when to put the focus back there. If month-on-month expansion drops below a certain threshold, we’ll revisit the amount of attention going to expansion vs new logos.

When that happens, maybe we spend the next quarter leaning into expansion instead. This isn’t crisis response. It’s a planned trigger that should force us to look at this mix again before either side becomes a panic.

In sum:

• Expansion is the product working. Keep shipping.

• New logos are the go-to-market working. Keep hunting.

It’s a constant shifting of balance and neither side of the scale lets you rest. That’s the fun part, though.

Which side of growth needs your attention today more than the other?

TTFN,

Peter

“We can build that ourselves.”

I keep hearing it. It’s hard to know how to take it if you’re building software for a living.

A salesperson can describe exactly what they need, hand it to Claude or Codex, and have something functional over a lunch break. The cost of building dropped so fast that the whole build vs. buy calculus feels like it’s been rewritten completely.

For companies buying software, it sounds like leverage and money saved (kaching!). For companies selling it, it can feel like a gut punch.

I don’t think it’s as simple as that.

Can you build it? Of course you can. But that’s not really the question anyone should ask.

The question was always, and still is: is this the right use of the people and resources we have?

A head of sales I spoke with recently wants his team prospecting, reaching out, picking up the phone. He doesn’t want them dabbling in engineering.

A RevOps lead at a $30M+ ARR company told me he doesn’t want his team thinking about data pipelines at all. He wants the insight delivered, so his people can do their success work, their expansion work, their onboarding work as well as they possibly can. They should be good at their job.

Both of them could technically build something, but neither of them should.

What’s actually changing is the job description for software companies, not the existence of them. Yes, the cost of building is cratering. Yes, that’s a genuine threat to businesses whose value was access to the capability, not to the expertise behind it.

If your value is domain expertise – deep knowledge of the problem, the data, the signal that matters – the bar got higher, not gone. It’s time to produce something better than a customer could build themselves in an afternoon — and yeah, I say that in just as breezy a way as LinkedIn influencers tell us all to “be authentic”.

Unlike being “authentic,” building opinionated, accurate, and integrated products can be done on purpose without feeling gross about it.

As much as “we can build that” comments keep me up at night, I’m starting to believe that deeper knowledge and stronger opinions delivered into the places people already work is how we’ll all build the next wave of software products.

To all y’all builders - keep going.

Peter

We spent two years saying the exact words our buyers use – to the people who were not our buyer.

This morning a head of sales at a Series B software company used our language back at me, word for word, on a discovery call. It was totally unprompted and just came up in the flow of the conversation.

It was the phrase that felt right but never seemed to land for us. It rolled off his tongue like it had always been there in his brain. Ready to go.

The voice in my head screamed <sigh>yep, that’s exactly what I ******* thought.</sigh>

Then my second reaction was frustration. If this is how someone like this guy naturally describes the problem, and we’ve been saying the same things for a while now, why have the last few years felt like push rocks up hills?

Firmographically, we had the right picture — the right size and type of company. But turns out we’ve been targeting the wrong person inside those companies. The people who feel the problem most acutely AND have the authority to do something about it. Seems obvious, right?

Right words, wrong room

He also clarified something we knew, but hadn’t operationalized enough: who your users are versus who your buyers are.

In most bigger companies, those aren’t the same person. They’re likely not even be in the same department.

Lesson learned: When you only talk to the people who already use your product, you keep optimizing for them and miss the person who has to sign the check. He painted our buying committee in one sentence. That conversation was worth two years of guessing.

This is why you talk to customers

There’s a harder version of this lesson – why you should also talk to people outside what you believe your ICP to be, even when things are going reasonably well. We have traction, but not as much as we’d like.

So stepping outside what we know felt a bit silly. But I’ve learned to talk to adjacent roles in the same companies we’re already selling into. Understanding the tangential people who feel the problem differently, or feel it more, or feel it and happen to have budget, is wicked important people, and it’s the thing that’s easy to skip because going off-script feels like a detour.

It’s not* a detour. It’s where you find out you’ve been in the wrong room.

(*My 5yo would correct me to say “It snot a detour”)

The changing mechanics of capturing this stuff

The other thing to talk about is the mechanics of how this research actually works now.

I take notes by hand during calls. Always have – it keeps me in the conversation rather than in the transcript.

But after this call, I went into a couple of Slack communities I follow, and what I’ve been doing with AI is capturing all of that context too. I take screenshots of threads I can’t API into and give those to Claude.

This helps build up a knowledge graph I can ask Claude or Codex to review and find connections.

After this call, Claude flagged something really useful. It pointed me to a thread from two weeks ago that was saying the same thing as my call, from a completely different direction. Someone in a RevOps community had described the same problem through a different lens.

I’d replied to that thread when it came through, but I sure wasn’t going to connect it to this morning’s call on my own.

Claude found the transcript and the notes from this Slack group and brought it up. Pretty awesome.

Before tools like this, holding the transcripts, the Slack thread, the handwritten note, and the recording all in the same working memory needed either a full-time assistant, everything being stored in Drive or Confluence (it wasn’t), or relying on my brain (mistake). I can remember a lot, but I cannot seem to remember which Slack channel to look in for anything I actually want to find.

The synthesis cost was the reason most founders do customer research in bursts and then stop. That cost has dropped enough now that I don’t believe the excuse holds anymore.

What AI doesn’t replace is the conversation itself

Calls go off in directions I couldn’t have predicted. I don’t think AI would have known to ask the right questions in the moment, or had them staged and ready. The human interest and the relationship you build on real calls is what goes sideways in unpredictable and often wonderful ways. That’s irreplaceable.

What got synthesized today only had value because the conversation happened first.

“Talk to your customers” is the most tired advice, yeah. All of us know we should do it more often and in more volume. We still don’t do it consistently. AI doesn’t change that, but it does change what happens after those calls and touchpoints. It’s easier now than ever to keep what you learned.

This thirty-minute call with the right person told me more about our business than two years of talking to people in the wrong roles. The tools let me build that connection faster (or maybe finally).

Talk to people. Then let the tools help you hold and make sense of what they said.

TTFN,

Peter

PS: This isn’t an argument to outsource thinking to LLMs. Guide them well and they serve you well.

In the last week HubSpot acquired Warmly and Zoom acquired Common Room. These are serious acquisitions. It’s real money spent on tools that surface who’s in-market, who’s visiting your site, who your reps should call next.

I get it. The front door is where new revenue lives. You want to know who’s knocking before they leave.

But once you’ve spent all that time, money, and energy getting someone through the front door -- who’s watchin’ the room?

Watching the room is what converts the front door.

An inside story that got my attention

We had a sizable customer in a 30-day trial. The account wasn’t hitting activation thresholds. Our generic onboarding emails went out. Opened, but zero response.

A playbook flagged the account. I sent a personal note -- co-founder to economic buyer. Nothing automated, just a simple “From what I can see, you’re probably 80% of the way there -- if you do these other 20% of your setup, you’re going to get value out of this.”

One email that get a response, got us on two live calls, and helped close a $10k-plus deal.

That flag only worked because we had account-level visibility and user-by-user granularity inside that account. I could see which users had engaged, which hadn’t, and where the setup had stalled.

Our attention has been split between the front door and the room from the start -- we wanted new logos just like everyone company does, but we also run daily meetings on trial activation and trial health. This deal is why we keep doing it.

The stack gap the big platforms ain’t talking about

Enrichment tools -- like Warmly, Common Room, Freckle, 6sense -- are excellent at telling you who someone is and what they should want based on firmographic profile and intent signals.

Behavioral data tells you what they’re actually doing.

This is behavioral data that nobody else can tell you. It’s pure 1st-party data you own. Only you can see what your customers are doing in your product, and that’s an awesome advantage to have if you’re leveraging that and your competitors aren’t.

The delta between expected behavior and actual behavior becomes your message. With good product data you can say: “I know what you’re trying to do. I can see exactly where you’re missing it. And I can tell you how to get there.”

Generic outreach doesn’t cut it anymore. “You still have three things left on your onboarding checklist” is noise in an otherwise noisy inbox. Operators who can connect observed behavior to a specific intervention -- that’s signal.

Why most post-sales teams aren’t watching

There are four reasons I see over and over:

1. The data’s a mess

2. The data’s inaccessible by the right teams

3. It’s not even the data -- it’s ownership of the tools

4. A tracking plan wasn’t implemented

Post-sales teams haven’t built this muscle. It wasn’t part of the job description five years ago. And it’s not a lack of will or interest, it’s been a lack of availability.

Then there’s the dashboard graveyard. Every team gets excited about new product analytics. Big launch with big promises. Within three to six months everyone’s tired of it telling them what happened without any real explanation attached.

“Trial activation is down 18% month-over-month.”

“Neat. Now what?”

“…”

Two problems stack up:

1. Knowing what good looks like (and documenting it) is step one.

2. Being alerted at the right time, without having to go look, is what makes proactive outreach possible.

Most teams have one. Too few have both. But if you don’t have it today, you can get it setup pretty quickly.

A quick audit worth doing

Set aside a few minutes. Look at your current tooling and ask:

• Do I have account-level visibility or just individual user activity?

• Do I know our activation criteria -- the in-product actions that actually predicts retention?

• Are the people who can act on that data the ones getting alerted?

This is the problem we work on at Accoil, so yes, we’re biased -- but the audit is worth doing whatever tool you use.

It’s worth an honest look: You can probably see who’s coming to the front door, but can you see what’s happening in the room?

The party happens in the room anyway, so it’s definitely the funner place to be ;)

Peter

Selling to startups is a bad idea.

I’ve been circling this thought for a while. A call this morning made it clear.

I was on with a founder in sales tech, with high growth potential, building in a blood-red ocean. Wicked smart and moving fast. He said exactly this:

I wish I had a problem my spreadsheet couldn’t handle.

I wasn’t pitching to him, but he genuinely wishes he had that problem. He wants more customers, more complexity, more revenue, more reason to spend money on tools. But right now his noggin and his spreadsheet handles it. So he doesn’t need us and he may not need you.

I had the same conversation with a homebuilder in our area. He’ll have 10 to 20 projects running at any time. “I have it all in my head,” he says.

It also happened with a founder I met at a conference who showed me the most carefully maintained spreadsheet I’ve ever seen. It was beautiful — I wish I could show you because you would agree. Even though it’s “just a spreadsheet.” He and his team know they’ve outgrown the beautiful spreadsheet. But they’re not done with it yet.

The trap I’ve fallen into is thinking the spreadsheet is the problem to solve. It isn’t.

The spreadsheet is a signal telling you the company is too small to buy from you. You can do a great demo. You can show them everything. If they’re still operating comfortably at that scale, they’re right to stay where they are. Spreadsheets are light and don’t cost much to maintain.

When a team moves off a spreadsheet they’re comfortable with, they may have to spend more time wrangling a new tool than facing the market. So they stay put.

The software founders and the homebuilder are absolutely right to run their businesses this way. As tools and tool counts grow, the surface area you need to manage at that stage changes from market surface area to software surface area.

The more time we spend facing the market — and the less time we spend managing our tools — the better.

When selling to smaller companies, there is a crossing point when they graduate from spreadsheet to tools like yours. But you have to know what to look for.

When we’ve seen it happen — when someone actually makes the move — it usually comes down to prioritization. The book of business has outgrown the staff. There are more accounts than one person can hold in their head, more decisions than gut feel can handle. Someone says: I only have so many hours in a day, and I need to spend them on the right accounts, the right activities, the things that keep as much revenue as possible.

That’s the signal. Not “I have a spreadsheet.” Not “I’m not sure about the ROI.” The signal is: I can’t prioritize my work anymore. Or something like that — you’ll know what that moment is for your customers.

There is a leap of faith customers have to take when moving off the spreadsheet. They know the spreadsheet is working well enough right now because their business is healthy.

And changing things up creates this fear of making the wrong choice and maybe going backwards. That fear almost always outweighs the potential to grow. It’s hard to argue someone out of that. You can only be ready when their own scale forces them to make a decision.

Interestingly, AI is holding this off longer than it used to. AI + Spreadsheets can do a lot right now at the early stage. But AI is probabilistic and it often doesn’t give you the same output every time. When you can catch the inconsistencies yourself, that’s fine. When you need consistent, repeatable action at scale, you need something deterministic like software. That reckoning is coming, just not today for most customers.

So what do you do with the spreadsheeters?

We have a Keep Warm column in the CRM. We watch for growth signals like funding, new hires, product launches, leadership changes, etc. We reach out when something changes. After all, the relationship is worth keeping and these signals can be automated. You can even put them in a spreadsheet ha.

Bottom line: if the companies you’re selling to are comfortable in a spreadsheet, you’re selling to the wrong people. I’ve learned this more times than I care to admit.

Think about what signs and signals tell you when your customers are too big for a spreadsheet. Find that point and you can go in and start selling to them.

TTFN,

Peter

PS: Happy Birthday Weekend, America 🇺🇸 and Go the Socceroos 🇦🇺!

I’m a software founder. so yeah, I want to believe software wins even in a world where everyone wants to build their own tools.

So when Markie Wagner published “Return on Tokens“ -- AI is a compiler, not a runtime, it has no goals, work decays into slop without human direction -- I felt that old familiar pull of agreeing with things that confirm what I want to believe.

And when Elena Verna dropped her “Mom-and-Pop SaaS“ frame -- build costs collapsing, domain experts spinning up their own tools, Jevons paradox meaning more software gets built not less -- I nodded along to that one too. Not quite happy about it, but also kind of encouraged...?

Both ideas are probably right. And both ideas are quietly threatening SaaS products.

If build costs drop far enough that a VP of Sales with no engineering team can just build their own sales intelligence tool -- why buy ours?

If domain experts can make something that fits their domain exactly, the generalist product starts to look like a giant spaghetti mess.

So I’ve been trying to be honest with myself about where SaaS products provide value today and tomorrow.

Where software still wins

I keep landing in the same two places.

1. Where reliably repeatable outcomes are needed.

2. As close as possible to decision points.

Let’s look at Accoil as an example.

The signal layer is becoming free. Or close to it. Messy data can be made clean with agents or agent-built software. It’s cheap to do this.

By signals, I mean things like a company just hired a VP of Sales, a competitor got acquired, a prospect visited your pricing page. These are legible and getting cheaper to access by the week.

First-party data, specifically product usage like Accoil ingests, is different. Higher fidelity. Harder to fake. The problem is, the data doesn’t tell you what it means.

Behavior is visible. Intent is not.

To get reliable behavior patterns, you need reliably repeatable interpretation of the messy data.

The what is visible. The why is the unsolved part. The why is where the money is moving because that’s where decisions are being made.

Yesterday I wrote that the signal is solved -- the action is the open frontier. That was one step. This is the step behind it, the one that makes the action problem hard: signals don’t interpret themselves.

Knowing a customer’s usage dropped 40% tells you something happened. It doesn’t tell you whether they’re about to churn, or whether they had a vacation (common and kind of predictable), or whether they’re quietly evaluating a competitor. The signal is legible. The situation is not.

The defensible layer isn’t the signal. It’s what the signal means for THIS customer, in their specific situation, right now. Not in aggregate. Not in general. For a specific “them.”

That’s where interpretation becomes the product. Action becomes the value.

(Accoil -- sigh, this is what we do -- lives in the signal layer. That’s exactly why I’m hunting the edge of it. If I’m sitting on a layer that’s becoming commodity, the useful thing is to see it clearly, not defend it.)

Ben Williams put the right frame on this: own the verb, become the default, expand from there. The verb is the action. Not the data, not the signal. The thing that actually changes what happens next for a customer.

It’s going to be ok

I’m optimistic for us software builders.

Markie’s right that AI has no goals. Elena’s right that build costs are collapsing. But what Jevons paradox tells us is that lower costs don’t shrink the market -- they expand it. More software gets built, not less. More signals exist, not fewer. More interpretation is needed, not less.

Everybody wins. But the people who win big will be optimistic enough to keep looking for problems to solve. As new tools emerge -- what problems get solved, and what new problems does that force us to find?

There’s more to build. It’s just not where we’ve been building.

Peter

PS: What do you think? Where do you see opportunity? Or do you feel like it’s all converging into agentic soup?

The signal is the reason to reach out. Max Mitcham proved it at Trigify: he says he got 48 meetings from a list of 120 leads using his own tool + Claude Code. In Max’s framing, timing and relevancy beat personalization every time.

He’s right. And there’s something else underneath that idea.

If you’re in the business of surfacing data or signals, that part is getting easy.

One sec: I should define what I mean by “signal” in this context.

These are signals:

• Product analytics data (sigh, this is what we do)

• Task lists and kanban boards

• CRMs and pipeline flows

• OKRs and KPIs

• Etc.

Intent data, hiring moves, funding rounds, competitor announcements. Vendors package this and point AI at it. The signal layer is becoming table stakes. It’s not worthless and it still has some life in it, but it is getting cheaper by the quarter.

Legible signals vs. valuable signals that explain nothing

External signals are legible. A company hired a VP of Sales. A competitor got acquired. AI reads these, scores them, triggers a sequence. That part is becoming a commodity, not a moat.

First-party data -- product usage, for example -- is different. It’s higher fidelity, but the data doesn’t tell you what it means.

Someone logs in every day. Are they winning or stuck? A feature sits unused. Did they not need it or did they just not find it? Behavior is visible. Intent is not.

“I can see what is happening, but not always why”

I asked the onboarding lead at one of our vendors whether she could see how I actually use their platform. She was honest about it. She can see what I built, but only when she logs in as an admin.

What she can’t see is how I move through the product moment to moment. Where I got stuck. Why I stopped doing something I used to do.

Her words: “I can see what is happening in the account, but not always why.”

A signal-rich tool, built to surface behavior, and the person responsible for my success still can’t answer why.

The what is visible.

The why is the unsolved part.

The why is where the money is moving.

The open frontier

A RevOps lead I know has built real confidence in product signals. He’s not surprised by churn anymore. He can see it building weeks out. He and his team have made real progress and it’s awesome to see such clarity.

What he hasn’t cracked, though, is what to actually do about it. The action that reliably lifts net revenue retention. The signal is solved. The action is the open frontier.

That gap, between seeing it and knowing what to do, is where the real work sits now.

I picture it like I would a real frontier: imagine standing on the edge of the great forests of what is now the USA, staring out at an endless sea of grass and prairie. The value isn’t in the mapped forest behind you, it’s out in the unmapped prairie. That’s where you go to build new stuff and grab land.

What your product is actually becoming

Wes Bush at ProductLed and Ben Williams of PLGeek are both circling this. They’re asking their newsletter readers to think deeply about the problems they’re building to solve.

A lot of the problems we’ve all built for are now solved by Claude and Codex.

What a product even is anymore feels up for grabs. AI can do a lot of what software did for years. If your value proposition was “we surface the data,” that part is getting cheaper fast.

The defensible layer is interpretation and action. What this signal means for this specific customer, in their situation. What to do next. Not in general, but for a specific “them”.

This is a threat to a lot of existing tools, for sure. But it’s also an opportunity. It’s where the moats are forming.

If AI is making a core part of any product nothing more than table stakes, what do we build next?

A lot of what we’ve all built for the last while is tooling to say to a human reader: This is what you need to know so you can decide what to do next.

As Claude Tag moves into Slack and gives us all an army of Claude agents embedded in every channel, the work of surfacing these signals is abstracted away.

And it leaves us to build for the problems and opportunities that come next.

When the signal is damn near free, where do our products provide value?

Peter

PS: This is actually an optimistic note. There’s more to build. It’s just not where we have been building.

Dashboards don’t drive anything. Alerts that fire when the good or bad signals happen are what kick off the work.

This morning I pulled up the new dashboards the team’s been building in beta. They look good. They surface the right data. Then I sat there for a minute and realized that the dashboards aren’t the point. The alerts, the actions, the outcomes are the point.

I’ve known that for a while, but it landed when I started building the dashboards myself.

We built Accoil to move product analytics from a system of record to a system of action. The distinction sounds clean in a pitch. Living inside your own product makes it concrete. A dashboard can give the appearance of understanding. You open it, you see numbers, you feel informed. But feeling informed and acting on a signal are two different things.

What dashboards vs signals actually looks like

One of our top-three accounts quietly disengaged over about a month. Usage dropped and a power user who’d been active three days a week went quiet. Nothing dramatic – no cancellation, no support ticket, no angry reply to a check-in email. Just a slow cooling.

If all I had was a dashboard, I’d have to remember to look closely, compare week over week, notice the trend. That sort of stuff is easy to miss. What’s cool is that the alert caught it. It got the data to us in a way that we could act on it – reach out, understand what changed, get ahead of churn instead of chasing it.

The flip side matters just as much.

Another account came roaring back, doubling its active users over two weeks. That’s a signal to pounce on now, not discover late. An expansion conversation, a case study ask, a referral introduction – all of those land better when you catch the momentum live, not in a quarterly review that most customers don’t really want to do anyway.

The work isn’t in the dashboard. It’s in the signal, the alert, the action.

A number you have to remember to go look at is a number that’s already too late

Dashboards are necessary. You need them to orient, to understand your business, to know what normal looks like so you can recognize abnormal. But they’re not enough.

A dashboard you check on a good Monday morning doesn’t find the account that cooled last Wednesday. Tools should work for you, not demand you work in them.

What if the signal came to you?

Putting it into practice

Pick one customer segment – your top accounts, your recently onboarded cohort, your highest-growth users. Pick two or three signals: cooling (feature drop-off, login frequency down, power user goes quiet) and heating (usage spike, new users added, breadth of feature adoption up). Wire them to Slack. Bing bong.

Or if you want to go further, run them through Claude via MCP with a cron job and bring the summary into your daily standup.

The point isn’t the tool. The point is that the signal reaches you where you already work, so it becomes impossible to miss.

A system of record tells you what happened. A system of action finds the signal in the noise and puts it in front of you before it’s too late to do anything about it.

Build more systems of action.

Peter

We give people 30 days to try Accoil. For a long time I let that number run my thinking: 30-day trial, so we’ve got 30 days to win them over. Check in around day 30, see how it’s looking, try to rescue it or send a thank-you note.

That’s backwards. And I think a lot of us have been doing it.

Amplitude released a report saying that the first seven days of a new customer’s life has an outsized impact on their retention and lifetime value. The better activated someone is in those first seven days, the higher their retention rate at the 90-day mark. I’m now hearing versions of this on call after call. Teams everywhere are shrinking the window they actually pay attention to.

The 30-day trial trap

Turns out the 30-day window is a trap. If you offer a 30-day trial and you don’t look at the health and activity of an account until day 30, you’ve waited too long. There’s no time left to respond, react, nudge, or adjust. A 30-day window is really a decision made up front: we’ll let be what will be, and at day 30 we’ll either scramble to save it or pat them on the back and say thanks.

A seven-day window is a completely different posture. It says: let’s put as much of our attention, and ideally theirs, on getting activated, understanding the value, and making the rest of the trial genuinely useful. So that when the moment comes to put the credit card in, it isn’t even a question.

Why is the window shrinking?

Two reasons, I think.

First, it’s finally getting the attention it always deserved. I know a $30M ARR SaaS company that shifted focus away from retention work and toward onboarding work, because they realised the better their onboarding and activation, the higher their retention downstream. It’s a renewed focus on the highest-leverage moment you have with a customer: the very beginning.

Second, switching costs have collapsed. When I shop for tools now, I’ll run two or three at once. The one that’s easy to use and helps me get going wins. The one where I have to wait on a CSM to schedule a call, who then sends me docs I couldn’t find myself, loses. Momentum gets built or destroyed fast.

I can see it in our own numbers. The single biggest early predictor of conversion is how quickly a team connects a data source. Connect on day one versus the last week of the trial, and you are remarkably more likely to convert. We’ve even extended trials so teams could have a proper crack once data was flowing, and it still doesn’t work late. If the momentum’s gone, it’s gone. Get them activated early or it’s very hard to claw back.

This is the reframe: offering someone 30 days and needing 30 days are not the same thing. You offer 30 in the hope they find value. The window you actually manage might be seven days. It might be one day.

Think of Facebook’s famous activation metric of connecting with seven friends in ten days. Every product has its own version of that. Ours is connecting data, early. While we give people 30 days, what we really want to see between day one and day seven is an absolute frenzy of activation activity.

The shrinking activation window

The Monday-morning move, and this was literally my Monday this week: map your ideal onboarding day by day. Two dimensions, time and activity. On what day do you want a customer to complete what action? For us, if you can point your data source at Accoil in the first 15 minutes, we’re off and running.

So, what’s your product’s first step? Once you have that written down, move on to the next step on the ladder. Not all activation is linear, but for the first critical steps customers shouldn’t skip rungs.

Like copywriting, the point of the first step is to get them to the next step, and so on.

Day zero matters more than I gave it credit for. Map it. Then start asking the question our customers are now asking us: how do I measure time to activation, and what can I test to shrink it?

Peace out,

Peter

The hype in GTM right now is all about going from signal to action as fast as possible. A customer does something, a sequence fires. A prospect trips a trigger, an email lands. Signal straight to customer, automated to the hilt before anyone’s checked whether it works.

Fire. Ready. Aim.

It’s seductive, because the promise is that it runs on minimal effort and minimal input. You flip it on and it hums, that’s the promise I see a lot.

But the hype leaves a lot of the story buried.

Edison famously took 10,000 or more tries at developing a lightbulb that worked and lasted. A lot of the results we’re shown on LinkedIn and AI-hype-y Product Hunt launches are the 10,001st attempts -- the workflows and automations that survived the trials and testing.

Automation is not the first step

The hype isn’t an empty promise. It just whitewashes the work behind automation that’s actually any good.

There are teams doing amazing work. Workflows.io has some killer materials and their clients are benefiting from their work.

But what we see publicly are the workflows and automations that the team fought hard to build. And even then they’re always refining things. Any GTM builder will tell you the same -- it’s a constant evolution to find what works today and hopefully tomorrow.

How do you do that?

It isn’t by automating the heck out of everything first.

Point your signals at yourself first

So before you wire a signal straight to the customer, point it at a human you know first. Build for signal to sales rep. Signal to success manager. Signal to the teammate who owns the account.

Send the signals to the human who owns the judgment and has the ability to look at something and know whether it’s the right move. Automation without this misses a lot of important nuance.

Even if you feed it every call and support ticket you own, your AI won’t have the same level of context and awareness of even junior teammates. Some things just don’t translate into robot.

A Head of Growth I know put it like this:

if the signal went straight to their customer, it felt like a missed opportunity, a relationship moment handed to a machine instead of a human. Most business is still human to human. That’s what makes or breaks deals.

I’ve felt the cost of skipping that. I gave an outreach tool what I thought were tight parameters, and it reached out to people we should never have been selling to. A symptom of moving through the build loop too fast.

Maybe it doesn’t feel like it cost much, but there’s reputational damage at risk. When there is a relationship worth having, getting it wrong costs a lot more.

It’s noisy out there

There’s noise underneath every signal-based GTM motion. AI is powerful, but everything was already noisy, and AI has made it really easy to turn up the volume.

So many tools promise to be the single pane of glass that filters the signals, yet I’ve never seen one team actually run from one.

That’s why we, the humans, still need to be in the loop for any automations or agentic flows we’re building. To help discern signal (value) from the noise.

When to skip the HITL

Is signal straight to customer ever right? Sure. If you’ve got a PLG, self-serve motion, it’s probably expected by customers.

For PLG + Sales-led companies like Atlassian, there are clear customer segments that warrant different approaches. For the PLG, self-serve customers, it’s fine to skip the human-in-the-loop because there are too many people to communicate with and the value from that work isn’t there in the form of high ACV. Self-serve customers can get shorter signal-to-action loops because there’s less at stake.

That’s not to say they should be treated like garbage. Put some effort into the automation, but don’t have a teammate sitting there reading every signal-based in-app message before it’s sent.

Higher ACV customers can and should benefit from the discernment and taste of a human. If your key customers have a seasonal business or have shared some context with you no AI would ever know, that’s leverage you can use to make the signal-to-action logic even better.

Start simple

If you’re not running data-backed signal-to-action campaigns, why not? If you’re building SaaS or online at all, you have data you can use.

Pick one outcome and one workflow for one role. Find a signal that supports the outcome. Run experiments, and route the signal to the human on your team so the judgment stays in.

Once you’ve run the workflow enough and are up to 80-90% sure automating it would bake that judgment in, expand from there.

Start simple. Keep the human in the loop until the loop earns its automation.

Peter

A customer pinged us last week, a little confused. A handful of their users had gone vertical -- big, sustained spikes in product activity. The kind of engagement curve you’d frame and put on the wall.

Turns out it was AI agents doing all the work, making it look like the individuals deployed them were super busy. These are bots those users had deliberately let loose inside the product -- not malicious, in fact very intentionally put there, part of a platform shifting from “do automations” to “be proactive.”

So despite the human users looking like they were on fire, the agents were doing the work.

Then I went and looked at our own website. PostHog reckons about 38% of our pageviews over the last 90 days are bots -- and that’s a conservative floor, sitting on top of the bots PostHog already drops before we ever see them.

An army of 267 “visitors” crawling the whole site, frozen on one Firefox version, running out of AWS datacenters. Strangely around the same time, I’d been getting Intercom messages and emails from new SaaS clearinghouses asking to list us. The traffic spike and the scraper outreach were probably the same story. I just hadn’t put them next to each other until now.

We already know bots and/or agents have overtaken humans on the internet. Expected. The question that actually matters isn’t whether it’s happening. It’s what it means for us, as builders and product owners reading data to decide what to build next.

My first instinct was wrong

My first reaction was to filter it all out and move on. Strip the noise. Get back to the humans. Maybe that was how you felt too.

But first we have to ask what value is attached to that traffic. And to answer that, we have to know what job the bot was sent to do.

If we’re talking agents, well they’re just a human abstracted away one more step. The agent does a pile of work, and at some point a human acts on it. Your product might be at arm’s length from that person now, but the activity isn’t valueless just because no person clicked the touchpad.

There are going to be entirely new jobs to be done that agents unlock -- work a human would never have done by hand. So “bot activity is just human activity, amplified” is a bad assumption. I’m betting much of it will take a different shape entirely.

Valuable job vs. vanity job

So how do you tell a valuable bot job from a vanity one? I don’t have a clean answer. But here’s where I’m landing for now:

Watch for volume swamping value. An agent that fires the same action a thousand times isn’t a thousand times more engaged. That’s the noisy bit -- the part you weight down, segment out, and look at on its own rather than letting it flood the signal.

Then do the unscalable thing: ask. Talk to the customer. What agents do you have running? What are they doing? Why? What are you hoping to get out of them? You’ll learn more in one call than in a month of staring at event counts.

Not all bots are pointed at you

There’s at least two flavors here, probably more. Worth noting, I’m using the term “bots” when it could be the traditional idea of a web scraper bot or an AI agent:

1. Bots working for your customer, inside your product. Think of Atlassian’s AI, Rovo. Rovo’s agents could be using your SaaS to get a job done. These can be real value -- a customer getting more out of you, just one layer removed.

2. Bots working for someone else, hitting your surface. AI crawlers reading your site, scrapers, eventually browser-based computer-use agents doing real work in your product on someone’s behalf.

The value splits hard inside that second group. A crawler that reads your site so ChatGPT can surface you in the right query? That might be really valuable traffic.

A scraping bot lifting your pricing and describing what you do (poorly, I might add)? That’s someone taking, not a customer arriving. It all takes the same shape in the logs, but has real opposite value. You can’t tell them apart without asking what job they’re on.

Why this isn’t academic

If you can’t separate human from agent in your engagement data:

• Health scores that mean nothing

• Activity rankings are built on ghosts

• Usage-and-revenue forecasts ride on work no person did

We can start optimizing for the wrong user -- building toward the wrong customer because we can’t (or don’t) see who’s actually getting value.

A funny irony? I’m using AI tooling to sort the AI noise from the signal. No idea how accurate it is. Feels like the whole problem in miniature. Anyway.

So, I don’t have a neat bow to put on this one.

If you’re reading any data that touches customer or public engagement, this only gets more important: learn to split it -- human, agent, bot -- and trace where the value actually lands.

We can only keep building for the right entity on the other end if we can still tell who or what they are and what they need to get done.

Ai ai ai,
Peter

Saturday morning I opened Claude on my phone to use Fable for something, and it was gone.

My first thought certainly wasn’t a big one. It was “did I accidentally unsubscribe?” or do something wrong?

So I checked my Anthropic account. Nothing wrong. Then I did what you do -- went to X to see if it was just me. And there was the notice: the US government had asked Anthropic to press pause on Fable.

My very next thought was: damn, I hadn’t finished the work I’d lined up with Fable. Our website. A free tools site build. A pile of things it had been quietly unlocking for me all week.

It’s Monday now, and I’m honestly not sure I’ll get through that work as fast, or with as much confidence, as I would have on Friday.

That’s part of the fun of all this, I suppose.

Locked in?

We got lucky, mostly. We’re not 100% locked into any one model, and none of our workflows are tied to a specific one -- we can swap an Anthropic model for an OpenAI one pretty painlessly. So in one sense, nothing broke.

But it rattled something loose. Are we actually planning for this?

Short answer: nope. The risk I’d been managing turned out to be oply one of the things to consider.

I’d been thinking only about cost. Token crunch, prices creep up, how do we absorb it, how much margin do we hand back. That’s a commercial risk, and commercial risk is the comfortable kind -- how you respond to growing costs is in your control. You’ve got levers. Pricing. Margin. You can model it on a spreadsheet.

What I hadn’t been thinking about is availability and continuity. What happens when a model just goes away, and not over a price you can negotiate, but because of a decision you had no part in?

That risk is out of your hands. The only thing in your hands is whether you did any insurance thinking ahead of time. I hadn’t. The question I couldn’t answer Saturday: if we entirely lost access to frontier models, do we know what we’d do? Again, short answer: nope.

It’s nothing new

Tools going away is nothing new -- they get acquired and shut down, sunset, quietly killed. Usually there’s an alternative and it means a migration headache, not a funeral.

A model feels different because it’s baked into how you work. Except, like I said, for us the model itself isn’t. Swapping vendors is easy.

So the lock-in wasn’t the vendor, but the category. It’s the assumption that frontier AI will simply always be there in the same way we assume the internet is always there.

Picture going all-in on the internet to run your business, and then having it taken away. That’s the dependency I’d stopped seeing. And it gets deeper every time you wire up another agent to do real work, important work on a schedule.

I don’t have a tidy answer yet, and we’ll keep using frontier models -- you have to, that isn’t the lesson. The lesson is smaller and more annoying: I’d stopped looking.

So here’s the one thing I’m doing this week. Taking 30 minutes to look at important workflows from across the copmany (product, engineering, go-to-market, sales ops, etc.) and drawing a quick exposure map.

Where are the dependencies we hadn’t thought about? What happens if it doubles in price, or disappears? I don’t think everything will need an immediate fix, but it’s important to see the dependencies and possible gaps.

From there, we can plan for insurance or fallbacks.

The worst dependencies aren’t the ones we choose. If there’s a clear decision to go “all-in” on a tool, it’s usually with eyes wide open.

The dependencies that sting are the ones we just walk into, assuming that everything will just keep working.

Time to take stock of how dependent we are on LLMs and those big frontier companies.

Peter

PS: This isn’t the first time (or the last) platform dependence matters. It’s just the latest. And maybe because it’s AI it feels bigger?

Two customers asked me some version of “what does good look like?” in the last two weeks. Different companies, same question.

If you’re building anything data-shaped, you get this constantly. What am I looking at? What does it mean? Is this number good or bad – and how would I even know?

The first instinct is to reach for a benchmark report. Makes sense. Those reports exist for a reason. But looking at industry benchmarks for a lot of data types is a waste of time.

The definition of good can be very flexible. It changes depending on things like your ACV, your product type, your market segment, whether your economic buyer is also your end user, whether you sell through partners or direct.

You have to build the benchmark yourself, from inside your own customer base.

The real question we ask of our data

When a head of CS asks of a health score “is 72 good?”, they’re actually asking: what am I going to do on the back of this? How do I react? That’s what’s underneath the question.

There are two sides to every data point. One is which direction it points you. The other is its inverse.

If you’re defining good, you’re also defining bad.

Once I have a definition of good, I spend most of my time looking at and looking for the inverse. It’s Charlie Munger’s “invert, always invert.” Understanding what makes a bad number happen is equally as important as understanding what makes a good number happen. That’s the half many of us skip.

How it looks in practice

Take account health. You’ve got an account in the bottom quartile of its customer segment, underperforming against the measures you’ve decided signal healthy or unhealthy. You’ve also got a peak performer in the same segment. The question worth asking: what’s the delta? Not “what’s the score” – what’s actually different between these two accounts?

The delta between the poor performer and the peak performer turns into a task list. Automated or human, doesn’t matter. That list is aimed at changing behavior on those accounts. That is the work.

That’s what all data intel points to.

We’re a data company and we’re still working on this

We built Accoil on a relative number, not an absolute one. That means “Good” can change.

Customers tell us they trust the signals Accoil gives them. But then it’s a matter of learning what to do with those signals.

That’s the gap pretty much all data hands us. Got a signal… what does it mean and then what?

With any data set we use to improve our business, new tools can help us all lean in:

• here’s the signal

• here’s what it means

• here’s what you can do about it

• here’s what happened after you did the thing

AI will help with a lot of this. But it still needs us to define what good looks like. That can come from our discernment. If we’re talking customer data, it can and should come from the customers themselves. Your best accounts are writing the definition every day – with how they use your product, when they pay, how they engage with support, and so on.

Draw the line

For any of this to work, you have to commit. Draw a line somewhere and say: above this is good, below this is not.

Then the job is doing the day-to-day work of moving more things above the line and understanding what’s holding the rest below it.

We’re all swimming in data. To make sense of yours, define what good looks like, define what not-good looks like, and spend your time closing that gap.

Peter

Went and made coffee. Had some lunch. Walked past the computer a few minutes later.

Done.

The prompt was deliberately thin: “Help us improve the performance of this website.” No goals. No conversions, load times, bounce rates, time-on-page. No scope. I handed it a problem with no edges and pointed it to the tools it has access to like Posthog, Google, Github, etc.

It built its own edges and just ran.

---

That was yesterday. I’ve had Fable for all of a day.

With earlier Opus and Codex models it felt like I could reach the limits. I knew where the tool ended and I had to step in. With Fable, it feels like it’s ready for more, and I’m not sure how to give it more.

That gap re-triggered something I thought I’d processed: the fear that the top 1% of people using AI tools are going to start pulling away from me if I don’t figure this out. The anxiety has been there since circa ChatGPT 3.5. I’d gotten over it. Fable brought it back.

I’m not going to give into that feeling, but I think it’s pointing at something worth naming.

The bottleneck is me

The Every team published their Fable reactions this week, and two lines landed close to where I landed.

Austin Tedesco: “The model is so good that, especially for knowledge work in growth and go to market, I feel like my own limitations are the only thing ever holding it back.”

Katie Parrott: “I haven’t found the right project to hand it yet, but that’s probably a failure of imagination on my part rather than the model’s.”

The bottleneck moved. It’s not the model anymore. It’s me. Ugh.

Missions > projects

My first instinct was to reach for the tool-selection frame: stop treating Claude Code as a hammer, think of the models as tools, match the knife to the screw. That’s useful. But model choice will be abstracted away eventually. OpenRouter-style routing will handle most of it. That’s not the real shift.

What caught me was something that happened inside the website deliverable.

Fable didn’t just return a list of recommendations. It proposed benchmarks, metrics to track, and a timeline for re-assessment. It built its own feedback mechanism into the output.

A project has a defined scope and a finish line. A mission has a direction, a method for measuring progress, and permission to keep iterating. I gave Fable a direction – improve the website – and it filled in everything else, including the loop.

Give it a mission instead of a project. That’s what I think the shift actually is.

Mission-framing doesn’t replace judgment. It requires it.

The performance recommendations were sharp. There’s an animated transition between the blog index and individual posts – a nice UI touch that’s apparently slowing the pages down. That call was right.

The design recommendations felt cold and lacking polish. Not that I’m a designer and have great taste.

But there’s still judgment, discernment, feel that has to be overlaid on top of this stuff. Because of that, there’s a risk in giving any setup full autonomous permissions. The mission still needs a human in the loop. Not to rubber-stamp the answers – to bring the taste.

Squashing the anxiety or FOMO

What I’m actually doing with the anxiety? Aside from living with it? Ha.

Not arm-wrestling Fable to see what it can do. Instead, I’m trying to focus on the business needs.

This reframe is calming me down:

Stop chasing model capability and start looking at the areas of opportunity. As in the problems that matter, the gaps where we’ve been leaving value on the table.

The question I’m bringing to it now is this: what are the projects we’ve dreamed about doing for customers – genuinely useful things – that we’ve never had the time, skills, capacity, or resources to get done? Is Fable capable of doing those now?

That’s a different question than “what can this model do.”

That’s a mission, not a project.

Onward,

Peter

I wasn’t the difficult customer, nor the churned customer. I was the aspirational fit. The one who had every intention of using the tool, but was too small, too early, and too unready for it.

I’d convinced myself that we were ready. A good seller would have told me “not yet” and pointed me in a different direction. But I came in on the last day of the quarter, signed for a discount, and helped someone hit quota.

That’s the story from one side of the desk.

I’ve been on the other side too. And when I read this line from a Winning by Design paper on revenue governance – “customer success is absorbing onboarding load that was never priced in” – I didn’t need to think hard about what it was pointing at. I’ve watched an onboarding that should’ve taken two weeks stretch to six months. I’ve sat in those Monday morning meetings where the same customer comes up before anyone’s had coffee, and you can hear the room collectively exhale before anyone says anything.

The cost nobody itemizes

The obvious stuff is obvious: support hours, churn, CAC payback that never comes. If you’re honest about the fully loaded cost of onboarding and servicing a poor-fit customer over the life of a contract, you’re often losing money. The number gets big fast. Most people stop the math before they get there because getting to that number is uncomfortable.

What gets ignored is the stuff that’s harder to put a dollar sign on.

Morale. This one’s quiet. When your team starts every Monday feeling down about the same customer <sigh>again</sigh> before the week has started… you feel it. It’s one of those indicators that a good product analytics tool will never catch. It’s in the group chat or sideline convos.

Opportunity cost. The more time you spend on one customer who shouldn’t be in the business, the less time you have to find and serve the ones who should. It’s math. Every hour fighting to make a bad fit work out is an hour not spent on a customer where you actually win. It’s a wet blanket on your best revenue motions.

Attention distortion field. This one’s less talked about. You’re not just losing time, you’re losing signal. When your team’s energy and conversation is soaked up by an outlier, it skews what you think the real problems are. You can start solving for the edge case. You can start to think the product needs to go a direction it doesn’t. It’s dangerous.

Why doesn’t this get discussed?

It does get discussed – sort of. The sales-to-CS handoff problem is well-known. The MQL quality fight between marketing and sales is a rite of passage in almost every B2B company. It’s almost fun. Everyone knows the tension exists at the handoff points.

But it doesn’t get discussed very openly. I think there are two real reasons:

The first is structural. The WbD paper makes this point well: without governance to hold the whole system together, you end up with siloed teams each optimizing for their own metrics. Sales closes. CS onboards. Nobody owns the full cost. When a problem is spread across enough people’s lane boundaries, nobody has the full picture, and nobody wants to be the one who calls it out. It becomes institutionalized. Naming it feels like putting your neck out.

The second is that it’s just plain hard to measure. What is the fully loaded cost of a six-month onboarding versus a two-week one? To answer that, you’re deep into activity-based costing models – which are expensive, time-consuming, and even more opportunity cost in their own right. So we skip it. We do the easy math and hope the incomplete version tells us enough.

It often doesn’t.

Your ICP isn’t a fixed artefact

Every company has an ideal customer profile. But that profile changes as you grow.

What looks like a poor-fit customer today may have been exactly the right customer two years ago. Early stage, growth stage, scale stage: these aren’t just different phases of the same business. They often need different customer profiles to feed them.

The transition between stages can be messy. You carry forward the ICP from an earlier era. They live in your pipeline, in your prospecting muscle memory, in the type of buyer your brand still attracts Often even after you’ve grown past that segment.

You end up with poor-fit customers not because anyone made a bad call, but because the company changed and the motion didn’t keep up. That’s not bad acting. That’s just what happens when you’re building. It’s a good problem to have and solve.

Understanding which kind of poor fit you’re dealing with changes what you do about it.

When it’s worth it anyway

Poor-fit customers aren’t always a mistake.

Early on, when you’re still learning what an ideal customer even looks like – when cash matters, when a logo on the website has real strategic value – taking a deal that doesn’t fit perfectly is often a good call. You’re not failing to govern. You’re gathering evidence. Edison’s “ten thousand ways that won’t work” and all that.

The risk is the zero-sum thinking that comes after: if I go too narrow, I’ll run out of market. I’ve felt that pull.

The more useful frame is abundance. If we can get honest about why the poor-fit deals cost what they cost, even just back-of-napkin, the math usually pushes you back toward selectivity pretty fast.

If you’re running a business, you’ve had customers that weren’t a good fit. Things can end mutually and without drama. But sitting with why it wasn’t a fit – even in a 30-minute debrief – is how we avoid repeating it.

The question I’m sitting with

I don’t have a clean framework to work from.

I keep coming back to this: if every team in a recurring revenue business is absorbing a piece of the cost of a poor-fit customer but nobody ever adds it up, the cost stays invisible. And invisible costs don’t change behavior.

You don’t need a full-on governance system to fix this (but it would help). You do need one honest conversation with your team about one customer – the one everyone sighs about on Monday – and an attempt, even rough, to put a real number on what they’ve actually cost you.

Consider what that number looks like at your company. I’ll be here writing it out on this napkin…

Peter

When she moved from founder to advisory, she made herself a promise. To only work with companies she believed could actually deliver value to their customers. Even if that meant a slimmer pipeline for her.

There’s an enthusiasm in how she works. It’s nice to be around.

There’s a risk in getting too wrapped up in that enthusiasm, but honestly, I just appreciate it. Because if you want to take others to a higher level, you have to have that kind of enthusiasm for the work you do with them. You can’t lift anyone you’re not excited to be standing next to.

The narrative I keep having to put down

Building in tech, there’s always a story running in the background: build to go as big as possible. Unicorn status. Generational impact. People saying your name in 100 years.

We’ve all met people building with that intent. They’ve had real success and they want the astronomical version of it. It’s admirable. I love it for them.

It’s not for me. But it’s easy to get swept up in. Build fast, burn fast, rocket ship to Mars, exit with squillions in the bank, plant your flag at the top of the mountain and call that the pinnacle of life.

And then every day I work too much, every week I’m away from my family and the things that actually matter to me, doing what I feel I have to do instead of what I want to do, the top of the mountain doesn’t feel like something worth shooting for.

“Slimmer pipeline” is a story about scarcity

The slim pipeline thinking is finite game thinking. Only so many resources, only so many people, everyone fighting over the same slice.

I don’t buy it.

With most things, if you can find your tribe, you can make a good living doing the work, and have fun doing it. There are far more of those people out there than you’d think. Jordan Peterson said something like this once (agree with him or not, it’s beside the point here): until he started putting his thinking out into the world, he assumed he was basically alone in it. He wasn’t.

So there’s something to sticking to your guns. Putting yourself out there. Letting people know how you think and how you feel. More of them open up, make themselves known, or get pulled in by your enthusiasm and your willingness to lift them up than the scarcity story ever lets you believe.

It’s true for people and for product.

The little thing: help first

We’re building a business and doing founder-led sales. Some days it feels like a grind. The framing that’s saved me is looking for ways to help people first. Yes, there’s constant pressure to be selling, to be closing deals, but those come on the back of being genuinely helpful.

It’s nothing new. It’s Gary Vee’s jab, jab, jab, right hook. At some point you do have to ask for the money. But the thing I come back to every single day, every time I send an email, is one question: is this actually helpful? Is there an insight in here that someone can take, learn from, implement, and benefit from, whether or not they ever buy from me?

Do I get it right every time? Heck no. But the ratio of helpful:selling is getting better.

All the way down

Does this run into what we’re actually building? I think it does. That’s the whole idea. How do we make our users the heroes of their day? The heroes of their team? How do we make them feel like they pulled off something good?

It doesn’t have to be world-changing, just good. How do we give them back more time for the work they actually want to be doing, the work they know they need to be doing so they feel valuable and worthwhile?

I know how sappy that sounds, and that’s probably leaning too far into it. But yes, it goes all the way into the product. Can we lift people up with the tools we build? Can we lift them out of the day-to-day grind of whatever they’re stuck in right now?

One honest thing, and the ask

At Accoil, we’ve held strong opinions, very tightly, and I think that hasn’t always served us well. Sometimes we heard our customers, but we didn’t listen to them. They’d tell us what they needed and we’d say, “nah, you don’t need that.” We had a very firm view of what was useful and what wasn’t. Maybe it was a bit arrogant.

I believe that we all buy tools that have strong opinions. But those strong opinions can be loosely held to flex and form to real customer needs. Anyway…

There’s been a shift lately. Now when we look at the product, the question is different: are we making this useful in a way that helps people do good, meaningful work?

So, before doing the next thing, the email, the feature, the support ticket:

Is this helping? Will this lift someone up today?

TTFN,

Peter

That question just got a lot easier to answer. Point a decent AI model at a pile of signals, and it will spot the pattern and hand you the next best step: reach out to this account, nudge this user, prioritize that one. It’s fast, confident, and a lot of the time it’s actually useful.

But the “so what?” part was the hard part for a reason.

It’s the part where judgment lives, and if we’re not careful, we’re about to hand it to something that doesn’t know what’s actually behind the number.

Think about one of your accounts that’s super active. They’re busy. They’re logging in all the time. You see usage spikes constantly. Feed that into an AI model, and it’ll see it and say, “This one’s hot. Reach out right now.” Reality might be different. Maybe it is, or maybe someone set up an automation last week, and it’s a quiet cron job cranking through tasks and making a dead quiet account look busy. You get the same signal, but a completely different ”now what?”

In this example, it could either be a buying moment or a red herring.

Now add proactive agents to the mix, which is where a lot of tools and a lot of people, including me, are heading. More and more of the data that we see could be misleading. We can’t tell the difference between signals and false signals and noise. There’s a problem.

If the answer is to put AI on top of those before we understand the context and the meaning, we risk exacerbating the problem.

Ask AI a question, and it will give you an answer. Whether or not it’s the right answer is still hard to tell.

It’s getting very easy to collect data, interpret it, build on it, make decisions faster and faster. But the understanding that should go into all of this is getting abstracted away. If you did not build the dashboard, the metric, or the report, and you do not know what is feeding it or what it is really measuring, it is hard to know how to sense-check the recommendation you are handed. AI will give you an answer, and that answer arrives quickly. It can even arrive at scale.

For years, the skills we trained were signal versus noise, wheat from chaff. That still matters, but there’s a newer muscle we need to build. I think it’s one that will separate teams, products, and businesses.

Understanding what a signal is, what it means, what’s behind it, and what it implies for the decision in front of you is not a dashboard skill. It’s a decision-making skill, and AI doesn’t have your context.

I am certainly not anti-AI. Quite the opposite. It’s still important to hold onto some things. Closing the gap between getting a signal and taking action is the whole game, and AI is the best tool we’ve ever had for it. Still, speed toward the wrong action isn’t progress. A shorter signal-to-action loop aimed at a misread signal just gets you to the wrong place a lot faster.

If you’re leaning on AI to tell you what to do next, good. You should be. Don’t forget to understand what’s behind the signal: what’s in it and what’s not. Be very clear about what decision the data is actually supporting. The teams that win won’t be the ones that act on signals the fastest. They’ll be the ones who understand what the signals mean when they decide to act.

AI will give you the next best action if you ask, but whether it’s actually the best one is still on you to decide.

🚀,

Peter

That’s Dan Rosenthal, writing in Kyle Poyar’s Growth Unhinged. In Growth Unhinged’s 2025 State of B2B Go-to-Market report, intent-based outbound ranked as the #2 channel teams are investing more in for 2026.

So everyone’s racing to capture signals. Almost nobody has worked out what to do once they have them.

We see it every week. A team gets set up. The signals start firing with data flying into the CRM, notifications landing in Slack. It’s exciting. The excitement lasts probably about six months. And then it goes quiet.

Kate (one of my co-founders) put it better than I can:

“Everybody says ‘this is great, you’ve given us these signals’… and now what do we do with them?”

Almost every team using Accoil has said this.

This isn’t an Accoil pitch, though. It’s GTM in general. There’s this whiplash that happens after you set up any signals. That’s the fun part, because you get to see things maybe you’ve not seen before AND you get to dream up ways to use it all.

So the real work starts when you have to decide how to use it. What do you do with what you’re seeing? Turns out that’s not as easy to answer as you’d think.

• An ICP-fit company creates an account. Great, and then what?

• A target account raised a Series A six months ago. Now what?

• A power user jumps to a new ICP company. Now what?

There are endless signals out there, some more valuable than others (think 1st-party vs 3rd-party signals). More valuable than any signal is the judgment about what to do with it.

That last 10% Dan’s talking about? That’s the whole game.

We’ve got a customer who completely trusts their signals now. They run trials, and the one they watch is dead simple: did a brand-new account do the core thing in the first seven days? If not, someone reaches out before the trial goes cold. That’s it. One signal, one play.

The data part is done. What’s left is the human part, including the message, the timing, the relationship. The “so what?” And honestly, that’s the fun part: it’s creative, it’s energetic, it’s full of possibilities.

But on any team there’s variation on the “so what? now what?” decisions. Everyone sees the same signal, but comes up with totally different ideas about what comes next.

That’s not a problem unless you let it be one. It’s the point of getting the signals set up in the first place.

What happens after you get a signal? It’s all hypothesis. Build it, test it, run the feedback loop, make it better. Once you trust the signals, you get to refine what comes after them.

If you’re doing product-led growth or product-led sales, pay attention here: you are sitting on some of the most valuable signals that exist, coming straight out of your own product. You don’t need to buy more signals or more sources. You almost certainly don’t have a capture problem. You have a now what problem.

So make it small. Pick one signal → your single highest-intent moment after sign-up. Something a new customer can do in your product that makes them go “Hey, this is great!”

Then decide the one action it should trigger: a message, a call, maybe even a handwritten letter. Decide who owns it and when it fires. Pair one signal with one play, end to end. Run it and see what happens.

As you work out and find successful Signal-to-Action pairings, start stacking more of them. Find another signal, start a new play, run the feedback loop, improve it. Then do the next one.

Most teams have never built even one real solid Signal-to-Action play. I’m convinced it’s because it feels overwhelming. I felt overwhelmed reading Dan’s piece on Kyle’s newsletter.

But we’re all sitting on valuable data and signals. It’s what we do with those signals that will determine the trajectory our companies are on.

Like seemingly everything else right now: Signals are abundant. Judgment is scarce. That’s the part worth getting good at.

Peter