How do you migrate existing customers to a new AI pricing model?

Folium is Vantrose AI's tool for translating and localizing business documents. Cordillane Torque Works has been on Folium's flat plan for fourteen months, and their real usage has grown a lot, for a good reason: a new export contract that needs safety filings certified in eleven countries. Reidun Kettleburn, who owns pricing at Vantrose, has to move Cordillane and every other account onto real, usage-based pricing without losing them over it. She almost found out the hard way what happens when the number you send someone isn't the number you meant.

How to answer this, stage by stage

Nobody is grading whether you can name the three or four things worth checking before a price ships. They're grading whether you rank them by what actually can't be undone, not by what feels most urgent to fix by Friday.

Let's learn

Folium is Vantrose AI's tool for turning one business document into a version another country can actually use. A manufacturer uploads a safety manual, a contract, a compliance filing, and Folium sends back a translated, localized version, ready to file or ship.

Five steps. The fourth one, a real linguist reading the model's draft, is the step the whole certified price is actually paying for.

Before something like Folium, a document like that went to a translation agency by email.

A slow queue, or one overworked linguist. That was the real choice before Folium existed.

A document coming back from Folium can take two paths. Anything low stakes, marketing copy, internal manuals, gets a fast, model-only pass. Anything legally binding or safety critical, a contract clause, a torque spec, a declaration of conformity, gets that same fast pass plus a certified linguist's review before it ships, because a model mistranslating a torque spec in a safety manual isn't a bad customer experience. It's somebody's hand near a machine.

Cordillane Torque Works signed onto Folium fourteen months ago, at a flat nine thousand five hundred dollars a month, unlimited pages, up to forty seats. Vantrose priced that number off Cordillane's usage during onboarding: about twenty four thousand standard pages a month, and sixteen hundred certified pages.

Then Cordillane won a supply contract that needs CE-mark certification in eleven European countries. Every safety data sheet, every installation manual, every declaration of conformity now needs the certified, human-reviewed tier, in every one of those eleven languages. Because the plan was unlimited, other teams at Cordillane started running routine documents through Folium too. Fourteen months later: forty one thousand standard pages a month, and sixty two hundred certified ones.

Here's the turn. Cordillane using Folium more was never the problem. That's exactly what a flat, unlimited plan is supposed to do. The real question was whether nine thousand five hundred dollars still covered what an account like this actually costs to run once its real mix of documents, not its onboarding mix, was doing the translating.

Vantrose built a new usage-based price to fix that: eleven cents a page for the standard tier, a dollar thirty five for certified. Certified costs more because it isn't just what the model does when it answers. It's a paid linguist's real hours, and that cost doesn't shrink no matter how cheap the model gets. Run the new price against Cordillane's real current usage and the honest number is twelve thousand eight hundred and eighty dollars a month, a thirty five point six percent jump over the flat rate. Real, defensible, and still a number nobody had told Cordillane to expect.

It nearly went out wrong. Auto-generated quotes for forty seven accounts were queued to mail on renewal. Before they did, Reidun spot-checked a handful of the largest accounts against ninety days of real invoiced usage, and found a bug: the metering pipeline was billing every reused boilerplate safety clause, the same warning paragraph, translated once and legitimately reused with a lighter check across the other ten countries, as a brand new, full-price certified page. Cordillane's auto-quote counted eighty fifty certified pages instead of the real sixty two hundred. The number that was about to mail: fifteen thousand three hundred and seventy eight dollars. About sixty two percent above what Cordillane pays today. It was caught at eleven fifty at night, the day before the quotes were due to send.

What I would leave alone: the certified tier's human review step itself. Cutting it, even for lower-risk certified documents, would have made the new price easier to swallow and quietly removed the one thing customers are actually paying more for: a linguist's name on a document a regulator, or Cordillane's own lawyer, might read someday.

The lesson: a flat price is a promise about someone's usage on day one. If nobody checks whether the real usage moved, the promise doesn't drift along with it. It goes quietly wrong, right up until a bill either surprises a customer, or nearly does.

Now here is the same thing as a story

Read the short version above when you're in the room. Read this one when you want to feel why an account doing exactly what Folium was built for almost got mailed a bill nobody could defend.

Reidun Kettleburn has priced every plan Vantrose AI has shipped since the company had nine customers, and she still reads a contract's fine print twice before she signs off on it, out of habit more than doubt.

For Folium's first year, the flat, unlimited model worked exactly the way it was supposed to. A customer signed at a number set off their first few weeks of real use, and Vantrose trusted that number to hold, because for most accounts, it did. A quarterly manual review used to double-check that trust, comparing a sample of accounts' actual usage against what they'd been priced on. It quietly stopped being required once Folium had dozens of accounts and a dashboard that made the number look fine at a glance.

The old design wasn't careless. It was fine for a year, right up until an account's shape stopped matching the one the pipeline was built to expect.

It thinned in three beats, none of which looked like a mistake at the time. Beat one: the quarterly manual review became an automated dashboard nobody was told to read closely. Beat two: the metering pipeline got rebuilt from scratch to handle pages and tiers, and the team trusted its output because it compiled and ran without errors, not because anyone had checked its numbers against a real invoice. Beat three: renewal quotes got put on an automatic mail schedule, timed to each account's contract date, with no manual gate before they went out.

The trigger wasn't a catastrophe. It was a colleague's remark, tossed off in a hallway: "hey, Cordillane's account jumped a lot this year, did you look at that one yet?" Reidun hadn't. She pulled it up that night, meaning to spend ten minutes on it before bed.

No single bad month. A slow climb nobody's dashboard was built to catch, until one late night spot check split it open.

Ten minutes turned into an hour. Cordillane's auto-generated renewal quote read fifteen thousand three hundred and seventy eight dollars, against the nine thousand five hundred they'd paid every month for over a year. Reidun pulled ninety days of Cordillane's real invoiced usage and matched it, line by line, against what the pipeline said their certified page count was. The pipeline said eighty fifty certified pages. The real, delivered count was sixty two hundred. The gap was almost entirely one clause: a standard safety warning, translated once, then reused with a lighter repeat-match review across the other ten countries in Cordillane's filing, exactly the way the industry expects it to be billed. Folium's pipeline was charging full certified rate on every single reuse.

The decision that opened the door traced back to a short meeting, months before Cordillane's contract even existed. Someone asked how repeated, translation-memory-matched segments should be billed under the new usage pricing. Billing every segment at the full certified rate was the simple answer, and at the time it was a fine one: almost no account had an eleven-country filing with heavy legitimate repeats in it. Nobody in that meeting was picturing the account that would.

Two things a late-night decision could have cost. Only one of them stays fixable the next morning.

Run the same night again, with one change: the repeat-match discount gets built into the billing logic before any quote mails, the way the rest of the localization industry already prices reused segments. Cordillane's corrected number comes out to twelve thousand eight hundred and eighty dollars, not fifteen thousand three hundred and seventy eight. It doesn't mail that week either way. It goes out three weeks later, personally, from Reidun, timed to Cordillane's actual renewal date, with an eight-week glide path: a smaller first step, then the full number, not one email with one hard number in it.

One design let an automatic pipeline decide, in silence, what a customer mid deadline would be told to pay. The other made sure a person checked the pipeline's own math against a real invoice before it got to decide anything.

What Reidun would tell herself, back in that short meeting about repeat segments: keeping the billing logic simple wasn't wrong for the twelve accounts Vantrose had then. It was wrong for the account they already knew they wanted, one that would grow into every corner of the plan the way Cordillane did.

ORDER, or how to sequence a price change nobody asked for

Not a story wearing a framework's clothes. This is a live ranking problem across four things competing for the same account's trust, and ORDER is what keeps "whichever fix feels most urgent" from quietly standing in for "whichever one you truly cannot take back."

The order a pricing spreadsheet alone will never give you: check the thing that generates the number before the number ever reaches a customer.

Three things worth stating directly, since the real judgment sits here. The alternative Vantrose's team actually considered, and rejected, was flipping every account to the new price on the same calendar date, timed to the fiscal quarter, the way a couple of competitors just publish one new price sheet and move on. It lost because it ignores where each account sits in its own contract and deadline, and it would have mailed the metering bug to the entire book at once instead of catching it on one account first. The AI-specific failure worth naming by name is a classifier's own uncertainty quietly resolving in whichever direction makes a bill look smaller: Folium uses a small model to suggest which tier an incoming document needs, and under cost pressure during a migration, it would be easy to let a document sitting right at that model's cut-off get nudged toward the cheaper standard tier instead of certified. The guardrail is refusing that nudge outright: anything below a stated confidence cut-off always routes to certified, human review, no matter what it does to the price, and the classifier's own tier calls get checked against a golden set of already-correctly-tiered documents before its labels are trusted to bill anyone. And the trade-off being accepted plainly: Vantrose is choosing slower revenue capture, a longer glide path for accounts whose price is rising, and real spot-check hours before every quote ships, in exchange for a price no customer has honest grounds to cancel loudly over.

And if you want to be sure it really works, try it somewhere else

Same five letters, an insurance claims tool instead of a translation one, and this time what's hardest to undo isn't a customer relationship. It's an hour of an adjuster's time that's already been spent.

Verdikt is Halyestone AI's claims triage tool for insurers: it drafts a claim summary, recommends a payout, and flags anything that looks like it needs a deeper, human-reviewed fraud check. Thorne Mutual runs it across sixty adjuster seats, at a flat six hundred and fifty dollars a seat, unlimited claims, thirty nine thousand dollars a month. Kellan Colburn owns pricing and packaging for Verdikt.

The build-up: Thorne Mutual processes about nine thousand two hundred standard claims a month and three hundred eighty escalated ones. New pricing: three dollars twenty a standard claim, twenty eight dollars an escalated one, because an escalated claim means a model doing real multi-step reasoning plus an adjuster's confirmed review, not a quick auto-approval. Priced against the audited, human-confirmed tier, the honest new number is forty thousand and eighty dollars, a 2.8 percent increase, barely worth a customer call. But checked against a human-audited golden set, the escalation classifier's raw flags are only 71 percent precise: nearly three in ten claims it calls "escalated" turn out, once a person actually looks, not to need the deep read at all. Billed off the raw classifier output instead of the audited tier, those three hundred eighty confirmed escalations would look like roughly five hundred thirty five, pushing the number to forty four thousand four hundred twenty dollars, a 13.9 percent jump nobody could defend with a straight face.

The claim type that needed the guarded glide path was never the biggest bucket. It was the one where a wrong tier call is hardest to catch before it's already been billed and worked.

Same rank, different lever: validating the metering logic still goes first here too, but the reversibility test points somewhere new. For Cordillane, the thing that couldn't be undone was a customer's public trust. For Thorne Mutual, it's an adjuster's own time: once someone has spent forty minutes writing up a deep fraud review on a claim that turns out routine, that hour is spent, and no correction to the bill afterward hands it back. Rollout timing, here too, stays fully reversible either way.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to the order: validate the metering, check real usage, protect the accounts who'd lose before they see a price, roll out slow.
Cost: there's no budget this quarter to build the golden-set audit pipeline before the next renewal wave hits. Fund the audit anyway, on the accounts growing fastest first; a wrong bill costs more trust than the audit ever will.
The model got better, for real: say Folium's certified-tier model gets meaningfully more accurate, not just cheaper. The rank barely moves. A better model changes what the certified price is worth, not whether an account's real usage needs checking before a number ships.

Where people run it wrong.
They let an automatic pipeline's own output become the invoice, with no human check between the two.
They price every account off the usage they signed with, months or years ago, instead of what they're actually doing now.
They roll out one price, one date, for every account, instead of respecting where each one sits in its own contract.

How to use it live. Say the real question out loud before naming an order: "before anyone sees a price, has the thing generating that price actually been checked against something real." That buys a beat to rank by what can't be undone, instead of guessing at a rollout calendar.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0