In July 2026, an executive who wants AI in production picks from three kinds of vendor. All three get paid whether it works or not. None of the three carries your outcome, and that is the only thing to know before you sign.
The first bills hours
A consultancy selling man-days makes money when the project runs long. It makes more when the project goes off the rails. No steering committee will correct that, because it is not a flaw in execution. It is the product.
Look at what that produces when public money pays and the Cour des comptes opens the file. The ONP, the state payroll system: seven years, 346 million euros, not one civil servant paid. The auditors’ verdict speaks of a patent failure of particular gravity and of resources spent to no purpose at all. SIRHEN, at the Ministry of Education: cost multiplied by five, close to 400 million. On AI, the November 2025 report is drier and no more flattering: pilots that stayed local, carried by isolated teams, never taken to scale.
The private sector produces the same ruins, with no public report to record them. I have watched a law firm pay for an unusable solution, delivered by people who had no solution, with enough vocabulary to leave the executive committee feeling clever on the way out of the meeting. They kept paying for months. It is a scam, and the word is exact: billing for a result you know you are not producing.
The client is not innocent. Whoever demands four hundred lines of specification and a monthly committee is buying, with his own hands, the eighteen-month project he says he dreads.
The second bills tokens
The lab does not sell time. It rents intelligence and collects on consumption. Gartner measures between five and thirty times more tokens for an agentic task than for a chatbot query. EY puts a customer service interaction at 0.04 dollars in 2023 and 1.20 dollars in 2026. Uber burned its annual AI budget in four months.
The number that settles it: across 2.4 billion API calls analysed in early 2026, routing every task to a frontier model comes to 18.40 dollars per million tokens, against 2.31 with a tiered architecture. A factor of eight, from orchestration alone. Your inference vendor has no reason to build it for you.
That is not an oversight, it is a position. In May, OpenAI launched DeployCo, four billion dollars, to put its own engineers inside your walls and convert your data and your processes into systems built exclusively on its models. The investors hold a guaranteed minimum return of 17.5% and bring their two thousand portfolio companies as a captive customer base. McKinsey, Bain and Capgemini are on the cap table: your reputedly neutral advisers own a share of the product they recommend to you.
What a lab learns inside your walls does not go back into the model weights. Enterprise contracts forbid it, and on that point they are honoured. It goes back into the heads of its engineers, into its playbooks and into its product roadmap. Ask Windsurf, whose capacity Anthropic cut almost entirely with less than five days’ notice, at the exact moment Anthropic was launching its own coding agent into the same market.
And on 12 June at 17:21, a Department of Commerce decision cut off worldwide access to two frontier models, simultaneously on AWS, Google Cloud, Microsoft Foundry and the direct API. No notice. Restored eighteen days later. Renting intelligence is not a vendor risk, it is a sovereign risk, and it belongs to another state.
The third bills seats
SaaS has sold the seat for twenty years and nobody objected. The same per-user rate for a bank of twenty thousand people and for a factory of four hundred. The same product too, down to the pixel, whatever value you get out of it. You pay the market average and you receive the market average.
The scam runs twice. An identical price for value created on scales that have nothing in common, and no customisation to close the gap. A seat that today does ten times the work of a 2019 seat costs the same as a seat nobody opens.
The market is starting to correct it. Pure per-user pricing fell from 21% to 15% of SaaS in twelve months, and outcome-based billing has just reached parity with per-seat subscription in signed contracts. Late, and slowly.
What they share
None of the three bills on value. The consultancy bills its presence, the lab bills your consumption, the software vendor bills your headcount. In all three cases you pay at the door. Your outcome appears in none of their income statements.
These models do not reform, they die out. McKinsey reports around 25% of fees indexed to outcomes, so three quarters still on time spent, and it took the US federal government wrenching more than twenty billion dollars of concessions out of its largest firms to impose performance on them. The movement comes from the destruction of the billable hour, not from a conversion.
What we sell
At Galahad, we do not sell an ontology-driven orchestration and security platform. That is what we build. It is not what we sell.
What we sell is a public service that answers first time. A production line that surfaces, in minutes, every unit touched by a defective part. An insurer that handles a claim without reaching for the wrong clause. An intelligence service that correlates twenty compartmented sources without ever pushing classified material down where it has no business being.
The rest is an engineering problem, and it is ours. You do not have to fund it.
We do not deliver a solution, we deliver a result. We own the creation of value inside the organisation, and we accept being paid on that alone.
The scale
A good AI deployment walks in front of an organisation that has a problem worth a million. It does not bill for studying the problem, it bills a share of solving it.
What we hold ourselves to: a measurable result in week zero, not a pilot, a result. At six months, a value-to-cost ratio a finance director defends on his own in front of his board. At twelve months, a venture-scale return, in a company as much as in a public administration.
We have just signed on that basis: 30% of the value created, payable on production. If the value does not arrive, we are not paid. No consultancy, no lab, no software vendor will offer you that, and all three have excellent reasons not to.
The objection
It is serious and it always comes. Value pricing only holds where the outcome is binary. Intercom bills 0.99 dollars per resolution because resolved can be proven. On an industrial ontology deployment, value is multivariate, attribution is arguable, and the client will happily say half the gain came from the reorganisation he started before us.
Our answer takes one line: the price is not contractual. We do not sell an attribution clause, we set an honest price at the start and we hold to it. Whoever demands a notarised formula before anything has even begun is not trying to win, he is trying to cover himself. He has already chosen his hours vendor, he just does not know it yet.
Hours, tokens, seats: three ways to bill at the door. The fourth invoice arrives after the result. It is the only one that commits whoever sends it.