An internal notice says Accenture funds $29 a month per person centrally; heavier use must be funded by a project or team. That moves an overhead tool into the cost of delivery and leaves projects to absorb the difference.
Accenture has told many employees it will centrally fund Claude only up to $29 a month each, according to an internal notice reported by 404 Media. The notice took effect on 1 October. Anything above that must be funded by a specific project or team. Accenture sells clients on AI at scale, and now it is deciding who pays for its own.
The $29 is a cap for some staff. 404 Media reports an employee whose Claude access was revoked altogether. For most people it works as a transfer price. Use above $29 moves from a central overhead budget onto the engagement that benefits. Each project then has to absorb the cost.
The notice says that costs are charged on consumption. By 404 Media's account, staff who go over the central amount must find a specific project or team to fund the rest. The publication calls that a Work Breakdown Structure, or WBS, at Accenture. A WBS is a project or team code against which an engagement books its time and costs.
The funding rule controls access as well as accounting. The notice says that "providing a WBS and obtaining a license does not guarantee indefinite or ongoing access". It adds that licenses "may be reallocated" to where they provide the greatest value. Funding a project code is necessary, but it does not secure access.
The notice sets up two tiers. 404 Media reports that Premium plans come with $250 of token usage a week, but many employees are not on that plan. It concludes that Premium users "might be okay". The exposed group is staff on the General plan whose usage is already high.
Internal messages quoted in the report describe monthly bills of $500 to $1,000 from work use. One of those employees builds an "AI accelerator" for clients. These are employee anecdotes rather than company data, but they show the scale a heavy user can reach. Anthropic's billing page describes Enterprise as a seat fee plus usage charged at standard API rates. So a heavy user's bill grows with the work done and not with headcount.
Take a heavy user at the top of that range, $1,000 a month. Over a year that is $12,000. Accenture's fiscal 2026 results release gives revenue of $74.2 billion, approximately 814,000 people and a gross margin of 32.0%.
Divide revenue by people and each employee brings in about $91,000. At a 32.0% margin, that leaves about $29,000 of gross profit per head. So one heavy user would consume roughly 41% of the gross profit an average employee generates.
The central $29 is small enough to sit in overhead unnoticed. The usage above it is large enough to change the economics of a single project team, which is why it has been pushed to the project code. Overhead is spread thinly across every client. A project line lands on one client's engagement and has to be defended there. Usually the person defending it is also trying to hit a delivery date and a margin target.
The policy follows the biggest rollout the company has run. In December 2025 Anthropic announced that about 30,000 Accenture professionals would be trained on Claude. It also said Claude Code would be available to tens of thousands of Accenture developers, which Dario Amodei called "our largest ever deployment". A tool handed to a very large workforce now produces a bill large enough to need an owner.
Who pays depends on how the work is sold. On time and materials engagements, the client pays for staff time at agreed rates. AI consumption then has to be argued into a rate or passed through as an expense. On fixed price work, the overrun falls on the engagement's own margin.
Cloud computing teams know this structure well. Chargeback to project codes is standard practice there: a central pool, then consumption billed back to the cost centre that used it. The difference is that Accenture's cost centres are also its clients' projects.
The impact should show up in the margin line. Accenture's gross margin for fiscal 2026 was 32.0%, against 31.9% a year earlier, so the starting point is steady rather than stressed. At the central rate, Claude costs a fraction of a percent of revenue and is easy to absorb. At heavy-user levels across many engagements, it becomes a recurring input cost that somebody has to price.
Itemising AI use as a pass through makes the price transparent. It also exposes the vendor's productivity gains, because the client can see what AI work cost and ask for a share. Pricing it inside the rate card keeps the bill simple but leaves the vendor carrying the usage.
Some employees are already trying to get upgraded to a Premium account so that their projects are not charged for additional AI use. One wrote that the change "obviously has a significant impact on project costs". The spokesperson's general statement that "we make AI available to all our people" does not answer who pays for the heavy users.
The first quarter of fiscal 2027 is the earliest period when the effect could appear. It ends on 30 November 2026, with results expected in late December. Revenue is guided at $18.95 billion to $19.60 billion.
Reported gross margin is unlikely to show it. Suppose 10,000 heavy users each cost $12,000 a year. That is $120 million, about 0.16% of $74.2 billion in revenue, or roughly 16 basis points of margin.
The first quarter of fiscal 2026 already moved 20 basis points, from 32.9% to 33.1%, for unrelated reasons. A heavy user population would have to be far larger than that to stand out. The better test is how Accenture prices AI work in its statements of work, and whether management says so when it reports in December.