// A standing SPP frame

Three eras · One decision

Pricing the AI Harness

A harness buys capacity in a supplier’s metered unit and sells capability in units of its own. What you charge for is a decision software has faced before.

The harness is the third arrival at the same two doors. The API connector met them. The OEM component met them. Every arrival chose: keep the supplier’s metric and pass its cost through, or recast onto a metric of its own.

01 · The claim

The invariant

A layer that accretes capability on top of someone else’s unit eventually arrives at a binary. Door one: pass the underlying cost through. Bill in the unit you buy, calls, tokens, compute hours, marked up enough to carry your margin. Door two: recast the metric. Define the unit on what the layer now does for the customer, and accept that your input costs become yours to manage rather than your customer’s line item to audit.

Neither door is right in the abstract. Pass-through fits while the layer is thin. It strains as the layer thickens, because the unit on the invoice describes what you consume, not what the customer receives. Recasting captures what the layer adds, and it is the harder door: it moves consumption risk onto you and demands a unit that can hold at renewal. Where the line falls for a specific layer is diagnosis, not doctrine.

ASP→ SaaS→ platform→ copilot→ wrapper→ agent→ harness→ the next noun

What makes the binary hard is not its difficulty. It is that the industry keeps renaming the thing that faces it, and every rename resets the memory.

02 · The depiction

Three arrivals. The same two doors.

Read the unit each era carried in. Then read what each door does with it.

The two doors, three eras of arrivalsThe unit changes its noun, the call, the host seat, the token, and the decision at the end of the corridor never changes: keep the supplier’s unit and pass its cost through, or recast on a unit of your own.

03 · Three sightings

The decision has a history.

Era 01

The API connector

The API era opened with a unit that fit: pennies per call, while a call was CRUD. Then business logic moved into the layer, connectors began finishing work no call described, and layers that had become applications met the two doors for the first time at scale. The per-call rate card outlived the transition only where a call remained one uniform act.

The full arc, and the test that fell out of it: the API pricing era.

unit · the call

Era 02

The OEM component

Older than the API era. When one vendor embeds another’s product, the license grant has to answer: whose units is the grant denominated in? The supplier’s units, and the host is passing through, relaying a meter with a royalty schedule attached. The host’s units, and the component has been recast: the supplier’s meter stops at the host’s cost line. Every OEM negotiation that stalled on denomination was arguing pass through or recast.

The contract-side view: channel, OEM, and white-label licensing.

unit · the host seat

Era 03

The agent harness

Vendors shipping agent harnesses buy capacity in a supplier’s metered unit and sell capability in units of their own. The default answer repeats the API era’s opening position: the supplier’s token, passed through with margin, describing the vendor’s consumption while the customer bought a finished task. Credits soften the invoice without changing the decision.

The layer itself, taken apart: what an agent harness is.

unit · the token

03b · The side door

The doors also open from the side: the acquisition.

Every arrival so far came from below, a supplier’s metric entering through the stack. Acquisitions bring the same decision in through the side.

The acquired product arrives carrying a metric somebody else’s architecture chose, and the acquirer faces the doors again: meld it into the existing licensing architecture, or pass it through as an add-on sold on its own basis. The tell reads straight off a pricebook. A catalog where every acquired product still carries its original unit is a catalog where the doors were never chosen, one bolt-on at a time. Melding is the harder door here too: migration, renewal terms, a portfolio’s worth of variance. Passing it through is the quiet default, and it compounds with each deal.

04 · The sequence

Licensing, then packaging, then pricing.

Whatever the layer is called, the sequence does not move.

FirstLicensing

The value metric decision: which unit travels with the grant. Everything downstream operates on the unit it defines, which is why both doors open here.

SecondPackaging

Editions, add-ons, and allowances only mean something once there is a unit to allocate across them.

ThirdPricing

A rate on a unit nobody chose deliberately is a rate defending nothing.

The industry runs the sequence backwards at every rename: pick the pricing shape the new category converged on, inherit the unit embedded inside it, discover the mismatch at renewal. A recast only holds when the new unit does what a value metric must do: the buyer understands it, can forecast it before signing, and watches it track the value received. The whole argument, sighting by sighting, runs through pass through or recast.

05 · The test

Name the unit. Ask two questions.

The vocabulary will churn again. Whatever noun replaces harness, the test runs the day the noun arrives.

Question one

Did the unit exist under the previous name?

If it did, the pricing question is not new, whatever the category branding claims. Some earlier layer already faced it, and its history is available the moment you stop calling your situation unprecedented.

Question two

Does the unit describe what the customer receives, or what you consume?

A consumed unit passes a cost through, however many abstractions sit between the meter and the invoice. A received unit is a recast metric, with everything recasting demands of it.

Neither answer is a verdict on its own. But a company that cannot say which door it stands before has usually walked through one by default, and the default is pass-through, because the meter was already running.

06 · The practitioner

About the expert

Chris Mele

CEO, Software Pricing Partners

Ranked #1 on OpenView’s list of B2B SaaS pricing experts. Chris leads every AI-pricing engagement, surrounded by a senior team that has held CFO, CPO, and CIO seats inside software companies. You get the senior in the room, not their junior.

LevelSetter runs the pricing infrastructure end-to-end so the experts focus on the calls only humans can make. It scales practitioner judgment; it does not replace it.

Frequently asked questions

How do you price a product built on top of model APIs?+

Start with the unit, not the number. A product built on model APIs buys capacity in the supplier’s metered unit and sells capability in units of its own, so the first decision is which unit travels on your invoice: pass the supplier’s unit through with margin, or recast the metric on what the product delivers. Licensing carries that decision; packaging and pricing operate on whatever unit it chooses.

When is the right time to move from pass-through to your own metric?+

The tell is drift between the invoice and the product. Pass-through fits while the layer is thin, and it strains as capability accretes, because the supplier’s unit describes what you consume rather than what the customer receives. When customers buy finished work and the invoice still reads in the supplier’s meter, the decision is already on your desk. Where the line falls for a specific product is diagnosis, not doctrine.

What happens to your margin when the model supplier reprices?+

The exposure depends on which door you chose. A pass-through vendor relays the supplier’s move onto the customer’s invoice, so the customer absorbs the change and audits the line item. A recast vendor absorbs the move on its own cost line, and margin becomes something it manages rather than something it passes along. Neither door removes the exposure; they decide who sees it and who owns the response.

What makes a good value metric for an AI product?+

The same things that make a good value metric anywhere. The buyer understands the unit, can forecast it before signing, and watches it track the value received. A recast only holds when the new unit clears all three, which is why recasting is the harder door: it demands a unit that can survive a renewal conversation, not just a launch announcement.

Is cost-plus pricing ever acceptable for AI software?+

As an opening position, yes. Pass-through is cost-plus by another name, and it is legible, easy to defend, and often the right first answer while the layer is thin. The trouble is not choosing it; the trouble is never having chosen anything. A vendor that cannot say which door it walked through has usually defaulted into pass-through because the supplier’s meter was already running.

08 · The doorway

Standing in front of a door is not the same as choosing it.

If the unit on your invoices belongs to a supplier, and your product long ago started doing more than that unit describes, you have already faced this decision. You just have not answered it yet. Bring us the layer, in whatever vocabulary this year gave it, and we will show you which era already priced it.

Talk to a pricing expert Or take the Pricing Architecture Assessment →