Governance
AI budget: how to control spending without stalling adoption
Apple's lower daily AI cap shows that consumption is now a variable cost, and that governing it takes strategy
A company's AI budget now tracks what each team consumes, and Apple made that visible when it lowered the daily amount employees can spend on AI models. When vendors bill by usage, spending grows with the very adoption the company is encouraging. Keeping it under control takes rules on who consumes, how much, and for what, tied to the strategy behind each initiative.
Why did Apple cap employee spending on AI?
Apple wants to contain costs while it expands the use of AI across its operations. According to Bloomberg, in a report carried by Reuters, the company cut the daily AI spending allowance for some engineers from $300 to $200. It is considering a limit of $100 for some engineering teams, and some non-engineering groups that already operate under a $100 daily cap could drop to as little as $25. Employees who need more access would be able to seek exemptions from senior executives. Apple did not comment.
The allowance covers tokens, the units of data that models such as OpenAI's ChatGPT, Anthropic's Claude, and Google's Gemini process and bill for. The backdrop is research and development spending: in its latest annual report, Apple recorded $45.8 billion in R&D, 32% more than the year before, and attributed the increase mainly to infrastructure costs, including investments in AI, and to headcount.
Goldman Sachs has made a similar move. According to the same report, the bank restricted access to some AI coding tools after reviewing their productivity and risk. Both companies encouraged broad use of AI and are now checking what that use costs.
AI spending grows with use
A per-seat license has a fixed price, and the budget plans for it at the start of the year. Consumption of tokens and application programming interfaces (APIs) varies with what each person does that day, much like cloud usage, and an agent running long tasks consumes far more tokens than a short chat. That gives the AI budget a variable component that the annual budgeting process does not follow.
The volume is rising as well. According to Boston Consulting Group's (BCG) AI Radar 2026, a survey of nearly 2,400 executives, 640 of them CEOs, in 16 markets, companies expect to double their AI spending in 2026, from 0.8% to about 1.7% of revenue. Technology companies and financial institutions plan to reach about 2%. More than 90% of respondents intend to keep investing at current or higher levels even if the investment does not pay off in the following year.
Some companies still reward consumption. Researchers from BCG and the University of California, Riverside, reported in Harvard Business Review that firms are measuring and rewarding token consumption as a proxy for performance. Under that rule, the heaviest spenders look like the most productive employees, and costs climb with no one accountable for them.
More tools do not bring more value
Spending can rise without returns following. Researchers at the BCG Institute also measured AI adoption at more than 600 U.S. public companies using external signals: hiring, technology vendors, and what companies say in filings and earnings calls. About three quarters of them have moved past pilots and earn no premium for it. The average company in this group draws on about 13 AI vendors across nine functions, and the group's median firm trailed its industry peers by 1.4 percentage points a year in total shareholder return over the past three years. The leading 6% achieved, on average, 9.3% higher annual returns than their peers.
The authors attribute part of the gap to tool overload. Self-reported productivity rises as workers pick up a second and a third AI tool and falls after that, because switching between tools and checking their output costs more than the tools return. Saved time also goes to waste: in BCG's AI at Work 2026 survey of nearly 12,000 people, 66% of frontline employees who use AI regularly said they receive limited or no guidance on what to do with the time they save.
With 13 vendors, a company manages 13 contracts, each with its own renewal, security review, and integration, and pays more than once for tools that do the same job.
Caps per person, per team, or per initiative
A daily cap per person, like Apple's, is simple to apply and makes spending predictable. It treats the engineer fixing a production failure the same as the one generating drafts nobody will read, so it needs a way out for the use that matters most. When that way out is an exemption granted by senior executives, every request becomes an executive decision, and the most valuable work waits in line.
A team budget puts the decision with the manager, who knows the work and can shift funds between people during the month. A budget per initiative ties spending to a project with a goal and a metric, and it is the easiest to defend before the CFO, because every dollar has a purpose. The three can coexist: the per-person cap contains everyday use, the team budget adds flexibility, and the initiative budget funds work with its own objective.
All three need a clear approval threshold for exceptions. A rule that sends requests up to a set amount to the department manager, and anything above it to senior leadership, keeps control without stopping the work.
Each strategy calls for its own budget rule
The same researchers identify three paths that lead AI leaders to returns, and each one tolerates a different kind of control. The first is doing the same with less: using AI to make work faster and cheaper, with off-the-shelf tools and processes redesigned around them. On this path, a firm cap and a comparison against the savings achieved work well, because the goal is already to cut costs.
The second is making existing products better, and it depends on many AI-fluent people close to customers. According to the authors, 12% of employees at the leaders have AI-related skills, compared with 1.7% at the average firm in the middle group. This path calls for budgets per product team, with room to experiment, and a tight per-person cap cuts precisely the use that generates the ideas.
The third is building a new offering that would not be possible or economical without AI, usually with proprietary tools. The leaders on this path see their margins contract relative to peers, at least for a while. Judged like an efficiency project, against a payback period or a margin target, an effort like this would be shut down before it earned anything. The authors propose treating it like a venture, with stage gates defined in advance that release funding as evidence arrives that the capability exists and that customers would pay for it.
What to measure to govern spending
Governing the AI budget takes a few indicators, reviewed as often as the bills arrive. Cost by department and by vendor shows where the money goes. Cost per outcome, such as cost per ticket resolved or per deliverable completed, shows whether the spending buys anything. The number of tools per person points to overlap, and the rate of exceptions to the caps shows whether a limit is too tight or too loose.
These indicators complement the value metric for each project, which the article on AI ROI covers in detail, and budget decisions need cost and value side by side, with someone accountable for both. When IT manages the tools and a separate team runs the AI projects, spending has no owner, one of the effects of the conflict between IT and AI teams.
Conclusion
Apple's cap shows that AI spending has reached the point where it needs rules. Control starts with knowing how much each department consumes and with which vendors, then separates fixed from variable costs, consolidates overlapping tools, sets caps with a clear threshold for exceptions, and gives each initiative the budget rule of its strategy. The monthly dashboard closes the loop, alongside the value metric for each project. The first step is the review of the last three months of spending, which shows who consumes the most and which contracts overlap.
Takter's AI Governance service starts with an assessment of which AI the organization uses, with what risk, what value, and what governance maturity, including an inventory of all AI in use, vendor AI included. From there, Takter implements an AI Management System under the ISO/IEC 42001 standard, with an owner for each system, approval thresholds, and objectives with indicators tracked on a dashboard.
Sources
- Prathik Jayaprakash, Apple cuts employee AI budgets in cost-control push, Bloomberg News reports, Reuters, November 2026.
- Apple, Form 10-K, fiscal year 2026, November 2026.
- Boston Consulting Group, As AI Investments Surge, CEOs Take the Lead, January 2026.
- Julie Bedard et al., When Using AI Leads to "Brain Fry", Harvard Business Review, March 2026.
- Boston Consulting Group, AI at Work: Strategy Matters More Than Tools, June 2026.
- Adam Job, Ulrich Pidun, Matthew Kropp, Tania Wang, and Matthew Molloy, 3 Paths to Delivering Shareholder Returns with AI, Harvard Business Review, October 2026.
