I Asked AI Who Would Quit Over Our Price Increase. Wrong List.

The night before the price increase went out, Tani couldn’t sleep.

(He runs sales for a mid-size SaaS company — invoicing and scheduling software for clinics. I’ve been helping his team with AI adoption for about eight months, mostly unglamorous stuff. This was different.)

About 140 accounts.A flat 10% increase.

And no real idea how many would walk.He told me the swing was something like ¥4 million a year depending on whether five clients left or fifteen.That’s the kind of number that keeps you staring at the ceiling.

Here’s what most people do at that point.They rewrite the email for the fourth time.

They run the wording past their boss.They add a line about “continued investment in the product.”

Tani did something else. He took the client list, pasted in the last few months of email threads and the usage export, and asked a general-purpose AI — nothing custom, nothing fancy — one question: who’s going to leave, why, and roughly when?

The name at the top of the list stopped him cold.

Six months earlier, that client had said — out loud, on a call — “honestly, we couldn’t run the clinic without you guys.” A reference customer. The kind you put in a case study.

The AI had them as the single highest churn risk in the book.

I’ll get to why. But first I want to pick a fight with three things almost everyone in sales believes about price increases.

The three things we get wrong

One: that churn risk is mostly about the size of the increase. It isn’t.From what I’ve seen, a 10% bump barely registers for a client who’s wired your product into their daily workflow, and a 3% bump enrages a client who was already halfway out the door.

The number on the invoice is not the variable that matters.How they use the thing is.

Two: that the dangerous clients are the ones who complain. Wrong, and it’s the expensive kind of wrong.

Between you and me, a client who complains is a client who’s still talking to you.The ones that scare me are the quiet ones — the account where the order volume slipped 20% last quarter and nobody said a word, the seats that just stopped logging in.Silent contraction.It shows up in the data months before it shows up in a cancellation email, and by then you’re not retaining anyone, you’re doing exit interviews.

Three: that you need a “real” tool — a CRM module, a data team — to see any of this. You don’t.Tani used a spreadsheet and a chat window.

I’m not saying that beats a proper churn model.I’m saying the bar to get a useful first answer is embarrassingly low, and the people who’d benefit most are the ones who assume it’s out of reach.

Side note: there’s a version of “retention” that actively makes things worse. Blast every customer with an apologetic, over-explained “about our upcoming pricing changes” email and you’ll wake up the half of your base that hadn’t even noticed. I watched a team at a different company do exactly that. Their support queue tripled with “wait, what price change?” tickets. Sometimes the most retentive thing you can do is not send the all-hands email.

What the AI actually saw

Full disclosure: back to the reference customer.

The reason they topped the list: their words and their behavior were pointing in opposite directions.On calls, total enthusiasm.In the logs, the rollout stalled about a week after onboarding, only two of the eight licensed users ever logged in regularly, and over the previous three months even those two had been tapering off.

The AI didn’t care what anyone said on a call.It read the deceleration.

That’s the part I keep coming back to. We treat customer sentiment as the signal — are they happy, did they say something nice, did the review meeting go well. The AI treated momentum as the signal. And momentum was telling a much truer story.

It also did something I didn’t expect. Asked to sort clients by churn risk, it more or less handed back three groups:

• Group A — won’t flinch. Clients so embedded that 10% is noise.The AI flagged several of these as underpriced — usage way above what they were paying for.A churn-prediction exercise quietly turned into a “we’ve been leaving money here” exercise.- Group B — silently shrinking. Not complaining.Not threatening to leave.Just. .

using less, every month.This is the group that actually costs you revenue, and it’s the one a wording-focused approach completely misses.- Group C — fine with the price, furious about the surprise. Their issue isn’t the money.It’s being told after the fact, like they don’t matter enough to warn.

Three groups. Three completely different problems. One email was never going to do the job.

The move that worked wasn’t a discount

Here’s the thing I’d want you to take away, if you take away one thing.

What stopped the most cancellations wasn’t a discount.There was no discount budget — zero.It wasn’t a more eloquent email either.

Side note: it was changing the order and the timing of who got told.

Group B — the quiet shrinkers — got a call from their account manager before the notice went out. Not a sales call. A “how’s it actually going, what’s getting in the way” call. The price came up at the end, in context.

Between you and me, group A got the notice on the normal schedule, no special handling — and a couple of them got a separate, totally unrelated conversation a few weeks later about moving up a tier, because the usage data said they’d outgrown their plan.

Group C got a heads-up from someone senior — a short note with an actual name on it — a few days ahead, so the official notice wasn’t the first they heard of it.

Same price. Same product. Mostly the same words. Different sequence. That’s it.

The numbers, roughly: Tani had braced the team for somewhere around a dozen cancellations and that ~¥4M swing. What they got was three. The loss came in around a fifth of the worst case. And the list-building itself — the part that used to be two days of “I have a gut feeling about these accounts” with no confidence behind it — was a couple of hours, with a written reason next to every name.

The bonus nobody planned for: the tier-upgrade conversations with Group A landed three accounts at a higher price than before the increase. A defensive exercise threw off offense.

I want to be careful here, though, because this is the part where it’d be easy to oversell.

Where I wouldn’t trust it

I’ve now seen this done a few times, and the AI is wrong in predictable ways.

It’s wrong about thin-data clients.New accounts, light usage history — the model has nothing to chew on, so it tends to call them “safe.

“ They are not safe.They’re unknown.The newer the client, the more a human needs to look.

It’s wrong about anything that isn’t in the history. A competitor launching a cheaper product next month, the client’s own business hitting a rough patch — none of that is in the logs. The prediction is always an “if nothing else changes” prediction, and something else always changes.

And the list itself is radioactive.A document literally titled “clients likely to leave” leaking — to the client, internally to the wrong people — is its own disaster.

Names get coded.Distribution stays tight.You decide on purpose what you’re willing to paste into a general AI tool and what you’re not.(For what it’s worth, I’d keep actual company names out of the prompt entirely and map them back yourself.)

The subtler trap: letting the prediction turn into more work instead of less. If the answer to “who’s at risk” is “everyone gets the white-glove treatment,” you’ve broken your own capacity and learned nothing. The job of the AI here is to rank. It’s a triage tool, not a to-do-list generator. The minute it’s making your team busier, it’s being used wrong.

One more, and this one’s real: if you take “Group A can absorb more” too literally and start stacking increases on them every few months, you’ll burn the trust of the most loyal clients you have. “They can take it” is not the same as “you should keep taking.”

I haven’t figured out a clean rule for how often is too often. I just know I’ve watched it go bad.

— -

I’ll be real — Everything below is the part I’d actually hand to a sales lead — the prompts, the rollout templates, the failure modes with fixes. If you only wanted the argument, you’ve got it: a price increase isn’t a writing problem, it’s a sequencing problem, and the client most likely to leave is rarely the one making noise.

— -

The kit

Step by step

Step 1 — Build one sheet. One row per client.Columns: client (coded — C-01, C-02…), monthly revenue, usage or order volume this month vs.six months ago (a % change is enough), last meaningful login or order date, and a one-line read on the tone of recent emails (warm / neutral / tense).

You’re not building a data warehouse.You’re building one screen’s worth of truth.

Step 2 — Run the ranking prompt (below). Paste the sheet in as text. Don’t attach a file on the first pass — you want to see it reason.

Step 3 — Pull the top 5–8 and check them by hand. The AI’s job is to point. Yours is to confirm. Half the value is in the cases where you go “no, that one’s fine, here’s why” — that’s you calibrating the model out loud.

Step 4 — Design the rollout in three groups, using the timing template below. Then send.

Prompt 1 — Rank and explain

You’re helping a B2B sales team plan a price increase. Below is a list of
clients with revenue, recent usage trend, last activity date, and the tone
of recent email exchanges.

Do three things:
1.Rank the clients from highest to lowest risk of churning or significantly
reducing spend after a ~10% price increase.2.For each, give a one-sentence reason grounded in the data I gave you — 
specific (“usage down 30% over 6 months, no login in 5 weeks”), not vibes.

3.Sort every client into one of three buckets: A) will absorb the increase
easily — and flag any that look underpriced for their usage; B) quietly
contracting — using less, not complaining; C) will accept the price but
react badly to being surprised.

For what it’s worth, if a client has too little data to judge, say so — don’t guess.

[paste sheet]

Prompt 2 — Find the silent shrinkers (run this even if you do nothing else)

From the same list, ignore churn risk for a second. Show me only the clients
whose usage or order volume has dropped meaningfully over the last 6 months
but who have NOT complained or raised concerns in email. For each, tell me
what the drop looks like and what you’d want to ask them on a call. Order by
revenue at stake, biggest first.

Prompt 3 — Pressure-test your own list

Here’s my churn-risk ranking and my reasons. Push back. Which calls look weak
or under-supported by the data? Where am I likely letting a client’s friendly
tone override what their usage is actually doing? Name a client I’ve rated
“safe” that you’d look at again, and say why.

[paste your ranked list with reasons]

Template — the three-group rollout

| Group | Who | What they get | When | | — -| — -| — -| — -| | A — solid | Embedded, price-insensitive (some underpriced) | Standard notice, no special handling.Separate tier-upgrade conversation 3–4 weeks later for the underpriced ones.

Honestly, | On schedule | | B — shrinking | Using less, saying nothing | A “how’s it really going” call from their AM first.Price mentioned at the end, in context.| 1–2 weeks before the notice | | C — proud | Fine with money, hate surprises | Short personal heads-up from someone senior, signed with a real name.| 3–5 days before the notice |

Template — the Group C heads-up note

Subject: A heads-up before you get the formal note

Hi [name],

Quick personal note before the official one lands, so it’s not a surprise:
we’re adjusting pricing across the board from [date], about [X]%. You’ll get
the formal details from [team/system] in a few days.

I wanted you to hear it from me first. Happy to walk through the reasoning
on a call if that’s useful — no agenda beyond that.

[Senior name, title]

Failure patterns, and what to do instead

”We blasted everyone an apology email and woke up the people who hadn’t noticed.” Don’t pre-explain to the whole base. Group A and the quiet majority just get the standard notice. Hand-handling is for B and C only.

”The AI called our newest accounts low-risk, so we left them alone — and two churned.” Thin data reads as “safe.” Treat any account under ~6 months old, or with sparse usage, as unrated rather than low-risk, and look at it yourself.

”The risk list got forwarded too widely and it got awkward.” Code the names before anything goes into a prompt or a shared doc. Keep the mapping somewhere narrow. Decide deliberately what you’re willing to paste into a general AI tool — I’d default to no real company names at all.

When to change the prompt

Here’s the thing — — Wholesale or recurring orders instead of SaaS logins: swap “usage trend / last login” for “order volume trend / last order date / average lot size.” The silent-shrinker signal here is the lot size quietly dropping while the order still comes in.

• Professional services (retainers, fees): you may have almost no usage data — feed the AI the tone and cadence of email threads instead, and ask it to flag relationships that read as cooling. Weight Group C heavily; fee increases there are almost always an emotional reaction to feeling taken for granted, not a math problem.
• No clean usage exports yet: run Prompt 1 on revenue + last-contact-date + email tone alone. Cruder, still beats ranking by gut. Then make the sheet better next quarter.

— -

So that’s the playbook. The part I can’t hand you is the judgment call on the loyal clients — how hard to push Group A, how often. I don’t have a rule for it yet.

But here’s the question I’d sit with before you send anything:

Between you and me, your “we couldn’t run without you” client — are they saying it, or are they showing it? Because the AI only checked one of those. And it had the answer in about three minutes.

For what it’s worth, (I’m thinking about writing next about what happens when you point the same kind of prediction at your own team’s workload instead of your clients. Different can of worms.)

— - This article was written with AI assistance and edited by the author.