I sat through a sales pitch of an AI-native software startup last week. An all-in-one outbound system: contact data, signal detection, drafting, sequencing, call recording, pipeline, and a set of agents to run the parts a junior person used to run. A forward deployed account executive came bundled in, to configure the motion. The commercial terms were $10,000 up front as prepaid usage credits, no seats, no annual contract, expected to last three or four months. Before signing, they offered to size my addressable market, sub-segments, employee bands, titles, account counts so we could work out together how quickly I would consume the credits.
The easy reading is markup. The data and the tokens underneath cost a few hundred dollars a month at retail, and $10,000 for a quarter of that looks like a traditional markup. But it assumes the vendor knows what the thing is worth and has chosen to charge some multiple above it. I don't think they know, they can't articulate it beyond "outbound sales". I don't think anyone knows yet. The number in that email isn't anchored to anything, and the useful question is not whether it is too high, it certainly is. But that what the number is an indicator of is not even clear. Is it $10k worth of tokens? Is it $10k worth of human hours saved? Is it $10k worth of pipeline revenue? The startup offered to work with us to size our addressable market, sub-segments, employee bands, titles, account counts so we could work out together how quickly we would consume the credits.
Every established software category has a unit that both sides accept before the negotiation begins. CRM is per seat, because value moves with the number of people selling. Storage is per gigabyte. Payment processing is basis points on volume. Observability is per host, or per gigabyte ingested. The unit is not a billing convenience. It is the thing both parties have agreed the value is proportional to, and once a category has one, buying gets fast, because the argument narrows to a rate. Every one of those units took years to settle and most were wrong at least once along the way. Per-seat pricing made very little sense for CRM while CRM was still something a company installed on a server in its own building.
AI automation has no settled unit, and every candidate on the table has at least one unsolved problem.
Seats have become irrelevant even before they start. The product's central claim is that you will need fewer of them, and a vendor pricing per seat is asking to be paid less as the thing works better.
Consumption is the replacement the industry talked itself into, and it carries two problems. Tokens measure effort rather than outcomes: the hard case that burns forty times the compute and produces a rejection is frequently worth less than the easy one that lands. And the price of inference falls every few months, so a vendor indexed to it watches margins shrink while nothing about the customer's outcomes change. That is an unfinanceable setup. Very few clients would support a software multiple on a line indexed to a commodity in a price war, but it may be that that's where we're headed.
Outcomes are the honest unit and close to unbuildable in the first years of a category. Attribution is contested — the meeting happened, and whether the agent produced it or the brand did is not a question two parties resolve amicably at invoice time. Half the inputs belong to the buyer: their data, their process, their people's willingness to use the thing. And the cycle required to prove an outcome in enterprise sales runs longer than the runway of the company proposing it.
What is left is activity. Runs, records processed, accounts touched. It is the meter the software can actually see, and it tracks the vendor's cost considerably more honestly than it tracks your value.
So the category reached for prepaid credits, which are not really a price at all. They are a deposit against an unknown, sized by feel. $10,000 is the number that clears a founder's discretion without a security review or a six-week legal cycle, and is large enough that the buyer will actually implement the thing rather than let it sit in a tab. Neither of those facts has anything to do with what the software is worth. The figure floats free of cost, free of value, and free of any unit that would connect the two.
It is worth being fair about the vendor's position here, because most of the commentary from the buying side is not.
Conventional software has a marginal cost per customer that rounds to zero, and that is what allows a vendor to price on value and never lose money on an account. Inference does not round to zero. Cost to serve varies by an order of magnitude between two customers on the same plan, depending on how messy their documents are, how often their agents retry, and how much of the work turns out to need the expensive model. A vendor setting a price is modelling a cost base with no history, against a usage shape they will only learn after signing.
The value spread is just as wide in the other direction. A contract review agent is worth something close to ten times more to a firm processing four hundred agreements a month than to one processing forty, and without a unit that tracks that, the vendor either underprices the large customer or prices out the small one. Every flat fee in this category is quietly doing one or the other.
Then there is the part that is rarely discussed. In most of these deals the vendor is not yet selling a product. They are selling a hypothesis: that a system which has worked somewhere will work inside your particular arrangement of data, process and people. That is a reasonable thing to sell and a reasonable thing to buy. It is not a reasonable thing to price as though it were finished, and most of the bad structures, on both sides, come from pretending otherwise.
Which leaves the buyer with more room than most buyers use.
The first move is to ask what the unit is before anyone says a number. The question that works is this: what is the thing that, if there were twice as much of it, would be worth twice as much to us? A vendor who has thought about their own business has an answer — contracts parsed, tickets closed without escalation, invoices cleared, analyst hours returned. A vendor who answers in API calls or accounts enriched has not yet worked out what they sell, which is diagnostic rather than disqualifying. It tells you the price will change, and that you should not sign a long term. The failure to watch for is a unit that moves against you. If your entire addressable market is eight hundred accounts, a price indexed to outbound volume pays the vendor to burn through it, and you are the one who has to live in that market afterwards.
The second is to price the pilot as the joint experiment it actually is. A prepaid credit balance puts the whole of the discovery risk on your side of the table, which is odd, given that you are also supplying most of what the vendor takes out of it: a reference in a vertical they probably don't have, a workflow they can generalise into product, and the evaluation data that makes their next customer cheaper to serve. Say those out loud during the negotiation. It is the cheapest leverage in the room and it is almost never picked up. Whatever the structure, insist that unspent money converts rather than expires — credited against a first-year licence, not swept off the balance sheet in month four. That costs the vendor no cash, only optionality, and they will usually agree.
The third is to stop demanding pure outcome pricing and offer instead the shape a vendor can sign. They decline to be paid on results not out of cowardice but because they cannot run payroll against your adoption curve, and because half the variables belong to you. A floor plus a share solves for both: a fixed fee small enough that they survive a quiet quarter, plus a payment triggered on one measure agreed in advance that both parties can see in the same system. Then write down who owns which failure. If the pilot produces nothing because your CRM holds four years of duplicate records, that one is yours, and saying so in advance is what makes a vendor willing to carry the part that is theirs.
The fourth is to assume you will be repriced, and to write the mechanism now. This category will settle on a unit within a few years, as every category has, and whoever locked three years at today's guess will turn out to have been wrong in one direction or the other. Keep terms to a year. Ask what triggers a change and get the answer into the contract rather than into the renewal conversation. And if you are among the first customers in a vertical, being first has a value the vendor is collecting from you: ask for it back as rate protection when the model changes.
There is a version of this where the buyer is the unreasonable party, and it is common enough to name. Wanting a fixed fee, uncapped usage, a guaranteed outcome and no commitment is asking a company with fourteen months of runway to absorb every risk in the transaction. A fair amount of the defensiveness in vendor pricing is a rational response to having met that buyer several times already.
None of this requires a view on whether the technology works. It requires noticing that the category has not yet decided what it sells, and declining to negotiate as though it has. The vendor is guessing. The only question is whether they guess alone, with your money, on terms written before either side knew anything — or whether the two of you agree in advance what the experiment is, what it will have proved, and what happens to the money if it proves the opposite. That second conversation is available in nearly every one of these deals. It mostly doesn't happen, because the buyer opens by asking the price.
That's the fourteenth pattern.