Design a three-tier pricing structure for an AI product and justify each boundary.

Perchwave drafts captions and schedules posts, in a brand's own voice, for social teams. Nettlebright Media, a marketing agency, joined the beta running fourteen different client brands under one $69 Team account. Sabelline Thalberg owns pricing at Perchwave, and had built that tier around a five-person team running one brand, the way it was designed a year earlier. On a routine Tuesday margin review, she found the tier's price had nothing to do with what it actually cost to run.

How to answer this, stage by stage

Nobody is grading whether you can draw three boxes and put a price on each one. They're grading whether you know which of the three lines you could quietly redraw tonight, and which one locks the moment a customer sees it.

Let's learn

Every week, before Perchwave existed, a social manager spent about thirty-eight minutes writing a caption and guessing at a good time to post, for every account she ran. Three accounts meant close to two hours, every week, on top of everything else on her plate.

Five steps. The fourth one, a confidence check on the brand-voice match, is the step the whole pricing question actually turns on.

Perchwave cuts that thirty-eight minutes to about six: type a rough idea, get three captions back in the brand's own voice, pick one, and a scheduler has already worked out a good time to post it, per platform. Before Perchwave, a week of posting meant something much slower.

Two hours a week, spent mostly on two things a model is actually good at: matching a voice, and guessing a good time.

Perchwave sells three tiers. Solo, nineteen dollars a month, one brand voice, sixty AI captions a month, a plain caption model with no brand tuning. Team, sixty-nine dollars a month, up to five seats, up to three brand voices, four hundred pooled captions a month, and a brand-tuned caption model plus a batch scheduler that predicts a good posting time per platform. Studio, two hundred forty-nine dollars a month, unlimited seats, eight brand voices included, about three thousand pooled captions a month, extra voices at sixteen dollars each.

Nettlebright Media, a marketing agency, signed onto the Team tier during the beta and set up fourteen separate brand voices under that one login, one per client. Nothing stopped them. Beta accounts weren't yet held to the pooled caption cap, and the original Team tier had never charged for a brand voice at all, it was priced by seats, and voices were unlimited and free.

Here's the turn. Nettlebright using Perchwave harder was never the problem. That's exactly what a growing agency account is supposed to do. The real question was whether sixty-nine dollars still covered what that account actually cost to run, once its real shape, fourteen voices under five seats, replaced the one-brand shape the tier was priced against.

By month six, Nettlebright's fourteen voices were costing Perchwave $153.88 a month in real compute: $130.20 to keep all fourteen voices current overnight, whether or not they generated a single caption, plus $23.68 for the six hundred forty captions the agency actually drafted that month, well past the four-hundred pool that beta accounts weren't yet held to. Against $69 in revenue, that's more than double what the account paid, for an account that, by every usage chart Perchwave had, looked like a happy, growing customer.

The fix: count every brand voice, cap Team at three included, charge sixteen dollars for each extra one, and require a move to Studio past six. Run Nettlebright's real numbers through that new boundary and the account flips from a loss to a real one.

What I would leave alone: the Solo tier's price and shape. It's fine as it is, because it structurally can't hold more than one brand voice, so its real cost is bounded no matter what a solo user does with it.

The lesson: a boundary is a promise about how someone will actually use the product. Set it before you know the real cost driver, and the promise breaks quietly, one account at a time, until the wrong one finally shows up on a margin dashboard.

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 agency doing exactly what Perchwave was built for almost cost the company money every month it stayed a customer.

Sabelline Thalberg has owned pricing at Perchwave since the company had one plan and forty customers, and she still builds her own cost model by hand every quarter instead of trusting the automated one, out of habit more than doubt.

For Perchwave's first year, the seat-based Team tier worked exactly the way it was built to. A five-person marketing team signed up, ran one brand's voice, and the sixty-nine dollars covered what it cost to keep that one voice current, with plenty of room to spare. Sabelline signed off on that pricing and moved on to the next project.

It thinned in three beats, and none of them looked like a mistake at the time. Beat one: the beta opened to agencies, without anyone updating the cost model sitting behind Team's price. Beat two: creating a new brand voice, which used to need a quick internal approval, became fully self-serve, so anyone on a Team account could spin one up in about ninety seconds, with no limit on how many. Beat three: the monthly margin dashboard, once reviewed account by account by the pricing team, got automated into one blended number across every Team account, and the per-account detail stopped getting a manual look.

No single bad week. A slow slide sitting inside a blended number, until a routine Tuesday review finally split it apart.

The trigger wasn't a complaint or an outage. It was smaller than that: on a routine Tuesday, reviewing the quarter's margin numbers before a planning meeting, Sabelline noticed Team's blended gross margin had drifted from a healthy seventy percent down to thirty-nine, over five months, with no single incident anyone had flagged. She pulled the account-level detail apart to see which accounts were dragging it down.

Nettlebright wasn't the only one sliding, but it was the clearest. Fourteen brand voices, five seats, sixty-nine dollars a month. Real cost: $153.88. Nine other agency-shaped accounts were on the same slow slide underneath the blended average, just not as far along yet.

The decision that opened the door traced back to a short meeting, the week the Team tier's pricing first got set. Someone asked whether extra brand voices should cost anything beyond the flat seat price. The answer was no, because at the time the biggest Team account ran one brand voice, with two backups nobody used. It was the sensible call, for the account that existed then.

The old design wasn't careless. It was right for the account Perchwave had when the price was set, and wrong for the one it was already trying to sell to.

Run it again, with one change: every brand voice gets counted, and the Team tier caps at three included, extra ones at sixteen dollars each, up to a hard limit of six before the account has to move to Studio. Two days after that rule shipped, every existing Team account got recut against it automatically. Eleven agency-shaped accounts moved to Studio, at honest, higher prices. Nettlebright's bill went from $69 to $345 a month. Its real cost went from $153.88 to $170.90, mostly the same fourteen voices, plus a bit more caption volume now that nothing was artificially capped. For the first time, that account made Perchwave money, instead of quietly costing it $84.88 every month it stayed on Team.

One design let a growing account slide past a price that was never built for it. The other counted the thing that actually cost money, and billed for it on its own line.

What Sabelline would tell herself, back in that short meeting about extra brand voices: leaving them free wasn't wrong for the account Perchwave had then. It was wrong for the account it was already trying to sell to, one running content for more clients than it had people on the team.

ORDER, or which of three lines you actually draw first

Not a story wearing a framework's clothes. This is a live ranking problem across three lines competing to protect three different things, and ORDER is what keeps "price it low, medium, high and call it done" from standing in for a real, defensible boundary.

The order a pricing spreadsheet alone won't give you: get the real cost before the number that depends on it ever reaches a customer.

Two things a pricing meeting could get wrong. Only one of them stays fixable the next morning.

Three things worth stating directly, since the real judgment sits here. The alternative Perchwave's team considered, and rejected, was charging strictly per caption generated, a pure usage meter, with no separate charge for holding a brand voice at all. It lost because the nightly refresh that keeps a voice current runs whether or not that voice drafts a single caption that month, so a per-caption-only meter would have kept missing exactly the cost that sank Nettlebright's account, even at zero generation volume. The AI-specific failure worth naming by name is a caption posting unreviewed with a fabricated claim in it, a price, a discount code, a promise about stock, because nobody is watching a batch of forty scheduled posts go out overnight. The guardrail is the confidence check already built into the pipeline: any caption whose brand-voice match falls under a stated cut-off never reaches the schedule on its own, it drops into a human review queue first, no matter how good the model's day has otherwise been. And the trade-off being accepted plainly: routing a shaky caption to a person costs the account its ideal posting minute once in a while, and running a brand-tuned model instead of the cheap generic one costs about nine times more per caption, both accepted on purpose, in exchange for never shipping an off-brand or invented claim to a client's feed with nobody watching it happen.

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

Same five letters, a crop-scouting drone tool instead of a caption tool, and this time what's hardest to undo isn't a customer's trust in a support ticket. It's a season's worth of grower dues, already locked by a co-op's own board.

Furrowscout is Hartveld's drone tool for farms: it flies a set flight path, scans crop stress field by field, and drafts a scouting report flagging disease, pests, or water stress before a person walks the rows. Milosz Vondracek owns pricing on it. Grower, the entry tier, covers one farm. Co-op, the middle tier, aggregates member growers and batch-plans flights across many fields the same morning. Distributor, the top tier, adds multi-crop differential diagnosis, a heavier model that cross-references disease patterns across many different crop types, for an ag retailer reselling reports to farms it services.

Hartveld's Co-op tier billed forty dollars per member a month, built around image-analysis compute costing about five cents per acre per month at a typical eight hundred acres per member. Elderfield's real average ran twenty-five hundred acres a member, over three times the assumption. Real compute cost: about $1,500 a month, against $480 in revenue, a loss more than three times what the co-op paid, on an account that, by every dashboard Hartveld had, looked like a normal, mid-sized member co-op.

The boundary that needed the guard wasn't the biggest jump in dollars. It was the one priced off a count that didn't actually track the cost.

Same rank, different lever: validating the real cost per acre still goes first here too, but the reversibility test points somewhere new. For Nettlebright, the thing that couldn't be undone was a customer's own client contracts. For Elderfield, it's a co-op board's already-set season of member dues: once growers have paid into this year's plan based on a quoted price, that number is locked for the whole growing season, not something you can quietly correct mid-quarter the way a monthly SaaS bill allows.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to the order: publish the bounded floor first, pull real per-unit cost, commit the hardest-to-undo line second, tune the middle line last.
Cost: there's no budget this quarter to audit every co-op account's acreage against its member count. Audit the outliers first, the smallest-membership, largest-acreage accounts, because that's exactly the shape that broke the assumption once already.
The model got better, for real: say Furrowscout's disease-detection model gets meaningfully cheaper to run per acre. The rank barely moves. A cheaper model changes what the Distributor tier is worth, not whether Co-op's real acreage needs checking before a number ships.

Where people run it wrong.
They price a tier by a headcount that used to correlate with cost, and never check whether it still does.
They let a hard-to-undo boundary ship before the cost data behind it is real, because the "obvious" number felt safe.
They treat every account inside a tier as the average account, instead of recutting for the one shaped like the outlier that will eventually show up.

How to use it live. Say the real question out loud before naming an order: "which of these lines can I still move quietly tonight, and which one locks the moment someone else sees it." That buys a beat to rank by what can't be undone, instead of guessing at three tidy price points.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0