For companies that are really leaning into AI, one of the trickiest problems is figuring out how to allocate budgets, and how to calculate the ROI of the spend.
Last week 404 Media reported that Accenture had capped employee Claude spend at $29 a month. The story made the rounds quickly, mostly for the irony: a firm marketing itself as being on the forefront of AI Transformation finding itself scrambling to control its own AI bill and landing on such a blunt solution seemed…. ironic.
But if you look under the hood, Accenture is doing a lot of things right.
Responding to the initial uproar, an Accenture spokesperson clarified that $29 is not a hard cap. It is the amount centrally funded per employee - available without any additional approval. Spend above that threshold is still allowed, but it has to be mapped to a project or a client budget.
(note; when I first researched this topic, I assumed that Accenture was just passing through AI costs to their customers, just like how many AI-powered software companies do. While this might be partly true, 65% of Accenture’s revenue comes from fixed cost billing, meaning that in most cases, Accenture’s P&L needs to absorb any unexpected AI spend).
Heavy users also have their own lane. Premium accounts (allocated by “role and business need”) reportedly come with a $250 weekly token allowance, roughly 37 times the general amount. Licenses are reviewed periodically and can move to where they create the most value.
That design comes close to the model we see working at the companies we advise.
Set a budget, with a carve out for your proven builders. Make every dollar above budget the track to a key initiative, tied to an outcome they can name.
Accenture has a centralized budget. It has a builder lane, and every dollar above $29 now has an owner. Tracking token consumption to outcomes remains the hard part; we recommend that our clients lean on their people managers for this.
We have written before about the strongest-link company: find your builders, give them large budgets, and get out of the way. Accenture’s own data shows why the builder identification process matters.
In June, 404 Media published leaked audio of Accenture’s agentic AI strategy lead saying non-engineers were driving much of its token consumption. Converting PDFs was one of the big “token chewers.” Heavy users and builders are different groups.
How do you differentiate between the two? This is where good old fashioned people management matters. Your managers should be able to smoke test the correlation between token consumption and production.
The companies we work with that are furthest ahead can draw a straight line from where their tokens go to what the business gets for them.
We hold ourselves to the same standard at Legible. Routine, repetitive work runs on cheaper models. Anything public-facing gets the best model available.
We splurge while training a new AI teammate, then step down once its job is routine. Each teammate has a job description, so we judge it the way we would any hire: by what it produces.
Many leadership teams are setting 2027 budgets right now. Before you pick a number for AI, decide what everyone gets, find your builders and fund them properly, and give every dollar above the floor an owner and an outcome.