Pricing your first pilot

The instinct with a first pilot is to make it free. You want the relationship, the product is rough, and charging feels premature.

It is the wrong instinct, and the reason is not revenue. A free pilot removes the only reliable instrument you have for measuring whether the problem is worth solving.

A price is an instrument

Willingness to pay is not an opinion someone can report accurately. Ask anyone whether they would pay for something and you will get a courteous answer shaped by how much they like you.

A price converts the question into a decision with consequences, and decisions with consequences are the only honest data in early B2B. Four things become visible at the moment money is involved that were invisible before.

Whether there is a budget line, and whose. A free pilot never has to find one. The first paid pilot discovers whether the money exists and who controls it, which is most of what you need to know about how the eventual sale works.

Whether the owner will spend political capital. Getting a small amount approved for an unproven vendor costs someone internal credibility. Whether they will spend it is a much sharper signal than whether they are enthusiastic.

How they describe the value internally. To get it approved, they have to explain it to someone else. That explanation is your positioning, written by the person you are selling to.

Whether it survives the second month. Free things persist through inattention. Paid things get looked at.

Three structures

Fixed-fee pilot. A single amount for a defined window and a defined scope. Simplest to approve, easiest to explain, and the one to default to. Its weakness is that it decouples your price from their usage, which matters once your costs are usage-driven.

Outcome-linked. A base fee plus something tied to a result. Attractive in theory and usually a mistake at pilot stage, because attribution is unresolvable this early and you will spend the relationship arguing about measurement rather than building.

Usage-based with a floor. A minimum commitment plus consumption above it. This fits AI products best because it tracks your actual cost shape, and the floor is what makes the revenue predictable enough to plan against. Its cost is explanation: usage pricing is harder to approve because the buyer cannot bound their exposure without help.

If you choose usage-based, the burden you take on is making cost legible. Every credit-consuming action should show its cost before it runs, and running out should never lock a customer out of their own data or their non-AI workflows. Those two properties are what make consumption pricing tolerable rather than alarming, and they are how our own pricing works — the current shape is on pricing.

Pricing when your costs move

The specific problem for AI B2B SaaS: your marginal cost per customer is real, variable, and not fully knowable in advance.

Three moves keep this from becoming a trap.

Price the pilot to learn, not to margin. A pilot that loses money on inference and produces a clear picture of usage shape is a good trade. A pilot priced for margin on guessed usage is a number pulled from nowhere.

Instrument before you price the next one. Measure cost per customer, per feature, over the pilot. Joining model-provider usage and revenue into per-customer, per-feature margin — with lineage on every cent — is the thing that turns your second price from a guess into a decision.

Quote a range, and say what narrows it. Pilots are exactly the situation where ranges beat certainty. "Between this and this, depending on volume, and here is what we will both know by week four" is more credible than a confident single figure, and it protects the relationship when usage lands somewhere unexpected.

The three numbers to instrument

Whatever structure you pick, the pilot's real output is data about your own economics. Three numbers, measured from the start rather than reconstructed at the end.

Cost to serve, per customer, per week. Inference, storage, and any third-party calls, attributed to the customer that caused them. Attribution is the hard part and it is the part that matters — an aggregate cloud bill tells you nothing about which customer shape is affordable.

Cost by feature. Within a customer, which parts of the product consume the budget. This is what tells you whether your eventual pricing should be per seat, per action, or per outcome, and it is almost never what you would have guessed.

Time to first value. How long from access to the moment they did something they would have missed otherwise. This number predicts conversion better than usage volume, and it is the one to optimise between the first pilot and the second.

Instrument all three in week one of the pilot. Adding measurement later means the early weeks — the most informative ones, because that is when the shape is strangest — are gone.

Writing the end into the agreement

The pilot agreement should be short and should spend most of its words on how it ends.

State the window. State what happens by default when it closes — that the pilot ends, rather than that it silently converts, because silent conversion poisons the relationship you were building. State what they keep: their data, exportable, without a negotiation. And state the success definition in their words, agreed at the start, so that the conversion conversation is a comparison against a written criterion rather than a debate about impressions.

That last one is the piece founders skip. A pilot without a written definition of success ends in a conversation where both sides describe the same three months differently, and nobody has anything to point at.

For where this sits in a broader company shape, solutions for startups covers the operating side. But the pricing lesson is portable and does not require any product: charge something, make the approval happen, and write down what winning looks like before you start.