AI agents are moving to the meter, so contract for it before the vendor's engineers arrive

Microsoft's results for the quarter to 30 September were strong on the usual measures: revenue of $92.3 billion, up 19%, with Azure up 48%. What was missing was a number. Asked directly on the earnings call, the chief financial officer said Microsoft "did not disclose" an update on the Copilot figure this quarter. Microsoft 365 commercial seats grew 5% year on year, slower than the 7% of a year earlier. Seat growth had run at 6% to 7% in each of the four previous quarters.

Instead, the chief executive led with usage. More than 50 million agents now sit in Microsoft's agent registry, and the number of Copilot customers consuming credits through usage-based billing grew by more than 75% in the quarter, as stated by the vendor. Neither figure came with a baseline.

The same pattern appeared across the market. Revenue at AWS, Amazon's cloud unit, rose 39% to $46.0 billion, and Amazon said its agent service, AgentCore, deployed twice as many agents in the third quarter as in the previous three quarters combined, as stated by the vendor. ServiceNow raised its 2026 target for annual contract value from AI to $1.7 billion and said consumption of agentic AI in production was up more than 13 times year to date, as stated by the vendor.

FDE AI reads this as a change of business model, not of vocabulary. A seat licence caps what a customer pays regardless of how hard the software works. Adding consumption billing alongside the licence, as Microsoft is doing for new Copilot Business licences, removes that fixed cap: what remains is a limit the customer has to set, and the vendor's revenue grows with every task an agent takes on.

The change is not only in investor presentations. Microsoft's partner announcements state that from 1 December 2026, usage-based billing will be enabled by default for new Microsoft 365 Copilot Business licences bought through Cloud Solution Provider partners, with a default limit of 4,000 Copilot Credits per user per month that administrators can adjust. The date was moved from 2 November. On Microsoft's own worked example, a firm with 100 such users could consume up to 400,000 credits a month under the default.

The rollout will initially exclude 11 markets, including Australia, India, France and Germany. For buyers there, the delay is time to set policy before the default reaches them. For firms buying new Copilot Business licences through partners in the US, the UK, Canada, Singapore, the Gulf and other markets not on the list, the window is one month.

None of this makes consumption pricing wrong. Paying for work done can be fairer than paying for licences nobody opens. The point is that the unit of control has moved. Under seats, a buyer controlled cost by deciding who got a licence. Under consumption, cost is controlled by how agents are designed: how many steps a task takes, how often an agent retries, and which model it calls. Those are engineering decisions, usually taken on site, often by someone the buyer does not employ.

A default is a decision someone else has made for you. FDE AI reads the default spending limit as a ceiling the vendor considers reasonable, not one the customer has tested against a budget. Finance teams used to forecasting AI spend as seats multiplied by price will need a second line: expected tasks multiplied by credits per task, with a hard stop.

Vendors and their partners are also investing in the people who wire agents into a customer's systems. Accenture and Workday launched a joint forward deployed engineering programme on 13 October, describing engineers who work "within clients' actual processes, data, policies and controls" and feed what they learn back into the product. Anthropic has committed $100 million to train 10,000 engineers, which it calls Frontier Deployed Engineers, by the end of 2027, as stated by the company. (FDE AI is a paying customer of Anthropic and OpenAI, holds no shares in either company, and receives no free credits or commissions from either.)

The FDE model can work well. It also has a structural incentive built in, whatever the intentions of any particular programme. Where the engineer's employer is paid by consumption, the engineer who decides how many agents to deploy, how often they run and which model tier they call is shaping the customer's bill.

Analysts have raised a related concern. According to Techstrong.ai, Gartner expects up to 70% of enterprises to abandon agentic AI systems built through vendor-assisted models by 2028, as maintenance costs rise and internal teams struggle to change the technology on their own. The Register reported a Gartner analyst warning of lock-in and intellectual property risk in these engagements.

The services side may be next. Infosys reported that AI was 10.5% of its revenue and grew 29% on the previous quarter, while guiding to revenue growth of 1% to 2% for the year in constant currency. FDE AI reads that combination as one sign, among others, that buyers are moving spend towards AI-led work; if that work is priced by outcome or usage, clients in the UK, US and Australia who buy from Indian IT firms are likely to see the same meter appear in services contracts.