Most organisations still treat AI procurement as a future question: at some point we will decide whether to take Copilot. In practice, with most vendors that decision has already been made for you. AI is shifting from optional add-on to part of the package — and the price shifts with it.
Four vendors, four ways of doing the same thing
Microsoft raised prices across virtually the whole Microsoft 365 portfolio on 1 July 2026, bundling extra functionality into the base including Copilot enhancements. Office 365 E3 went from 23 to 26 dollars per user per month, over 13%; Microsoft 365 E3 from 36 to 39 dollars, 8%. The sharpest rises hit the smaller packages: Frontline F1 rose by a third, from 2.25 to 3 dollars, and Business Standard by 12%. Copilot also remains available as a separate add-on at 18 dollars per user per month on annual commitment — a promotional price running to the end of September 2026. The effect is dual monetisation: a base increase everyone pays regardless of usage, plus an add-on for those who want more.
Google chose precisely the opposite presentation for the same move. Gemini was folded into the base plans and the separate add-on disappeared. A customer on Business Standard at 12 dollars who added Gemini paid 32 dollars; now it is 14. For that group it is a cut of more than half. For everyone who never bought the add-on it is a two-dollar increase. Same strategy, opposite framing.
Salesforce prices Agentforce in Flex Credits: 500 dollars per hundred thousand credits, half a cent per credit. A standard action costs twenty credits, so ten cents; a voice action thirty credits, fifteen cents. There are also user licences and editions with an annual credit bundle. The point is not the rate but that your invoice now depends on how often something happens rather than how many people you employ.
ServiceNow went furthest: since 9 April 2026 its licensing has been rebuilt into three AI-native tiers with AI included in every tier. There is no variant without it. Consumption is metered in «assists», where a lightweight summary costs less than a multi-step agentic workflow, and the pool is tracked at tenant level. Exact assist values per action are not published. Legacy SKUs went end-of-sale on 1 July 2026.
From seats to consumption — and what that does to budgets
The common pattern is the move from a predictable price per user to a variable price per action. Gartner estimated in July 2026 that 234 billion dollars of enterprise application spend is at risk in the coming years from agentic AI — around 20% of SaaS spend by 2030 — because the link between user count and vendor revenue breaks. For you as a buyer the flip side matters more: your costs no longer track your headcount.
The same firm predicted in June 2026 that the cost of AI-assisted coding will surpass the average developer salary by 2028, with an explicit warning: many vendors lack transparency about how token consumption is calculated and billed, limiting enterprises' ability to forecast and control cost. It also names the absence of built-in vendor cost controls as a governance gap. That is not theoretical: 78% of IT leaders reported unexpected charges from consumption or AI pricing in 2026, and 61% had to cut projects as a result.
Are you paying for what you do not use?
The uncomfortable part: the return is far from proven. Gartner predicted in 2025 that over 40% of agentic AI projects will be cancelled before the end of 2027 due to escalating costs, unclear business value or inadequate risk controls, and noted that of the thousands of vendors claiming to be agentic only about 130 genuinely qualify. Forrester expected in October 2025 that organisations would defer a quarter of planned AI spend to 2027, and found fewer than one in three decision-makers can tie AI value to financial growth. ITAM research shows only 29% of organisations actually measure the value of AI software.
Rising, less predictable cost combined with barely measured return is exactly the situation in which contractual protection pays.
The clauses you now need
Start with the unit of measure. Have it recorded what a credit, assist, token or action precisely is, with a rate card fixed for the term, a consumption cap with alerts, and no automatic billing above that cap. What is undefined can become more expensive without the price changing.
Then settle the data question: an explicit prohibition on using your data and inputs to train the vendor's or a third party's models, clarity on where processing takes place, and disclosure of subprocessors. Record who owns the output and secure an indemnity for infringement claims on AI-generated output — and check that the indemnity does not lapse the moment you use a feature your own way.
New with AI: the right to advance notice when the underlying model, version or model provider changes, with a remedy if quality measurably degrades. And, given the moves above, price protection: a cap on renewal increases and protection against AI features migrating from add-on into a repriced base tier. Finally, the AI Act role split — who is provider, who is deployer, and what documentation and logging the vendor supplies so you can meet your own obligations.
You do not have to invent this. The European Commission publishes model contractual clauses for AI procurement, updated in March 2025 and aligned to the AI Act, in a full version for high-risk AI and a light version for the rest. They are written for public buyers, but they are free, authoritative and perfectly usable as a starting point for a commercial contract. We use them as the basis in our contract renewal work.