AI Token Budgets Decide Which Terms Govern Your Company’s Code

AI token budgets decide more than the size of the bill. When a limit sits below what engineers need, some of the work moves to personal AI accounts, and the company’s code is then governed by consumer terms the company never signed. James Stanier makes the observation that starts this argument in a recent article, and this piece extends it.

Stanier reports that the subsidized seat is ending and limits push engineers elsewhere

Stanier, CTO at Nordhealth, argues that per-seat AI plans are a subsidy that may not last. He reports that Anthropic moved its Enterprise plan to a seat fee plus metered usage in November 2025 and that OpenAI moved Codex on ChatGPT Enterprise to token pricing in April 2026, and he estimates that metered usage costs roughly ten times as much as a per-seat plan at today’s prices. His advice is to make consumption visible to engineers and managers now, never to rank people by tokens, and to match the model and reasoning effort to the task.

Early in the piece he describes a pattern from his network. As companies add guardrails and limits, many engineers get frustrated, sometimes secretly, and go back to their personal plans, which carry flat, generous limits. He notes that this sends company code and data through accounts the company does not control, and that “the real demand never shows up where anyone can plan for it.”

A personal plan runs company code under terms the company never agreed to

Stanier names the data risk in a sentence and moves on to cost. The data risk deserves its own look, because the relevant contract is public and easy to read. On 28 August 2025 Anthropic published Updates to Consumer Terms and Privacy Policy. The update applies to Claude Free, Pro and Max plans, including when people use Claude Code from accounts on those plans. Users choose whether their chats and coding sessions are used to train models, and those who allow it get a five-year data retention period instead of the existing 30 days. The same page says the changes do not apply to services under Anthropic’s Commercial Terms, which include the Team and Enterprise plans and the API.

Put that next to Stanier’s observation. An engineer who hits the company’s limit and opens Claude Code on a personal Max plan has moved the same repository from a commercial agreement into a consumer one. Whether that code can be used for training now depends on a setting in the engineer’s personal privacy preferences. The company cannot see that setting and has no say over it. A token budget therefore draws the line between work done under terms the company negotiated and work done under terms an individual clicked through.

Shadow AI was already common before token pricing arrived

This behaviour does not need metered pricing to start. Microsoft and LinkedIn’s 2024 Work Trend Index, published on 8 May 2024, surveyed 31,000 knowledge workers across 31 markets between February and March 2024. It found that 78% of AI users bring their own AI tools to work, rising to 80% at small and medium-sized companies. The same report found that 52% of people who use AI at work are reluctant to admit using it for their most important tasks.

Those figures describe a baseline in which most AI users already have personal tools and many keep quiet about them. A spending limit that blocks work in the sanctioned tool adds a fresh reason to switch, and the reluctance figure suggests the switch will rarely be announced. That fits Stanier’s description of engineers who get frustrated in secret. My inference is that the risk grows exactly when per-token pricing arrives, because that is when finance teams start tightening limits.

Every token cap needs a sanctioned fallback before it is switched on

The governance fix follows from Stanier’s own advice on model choice. He shows that model size and reasoning effort change cost a great deal, so a limit does not have to mean a stop. When an engineer reaches a budget threshold on the most expensive configuration, the sanctioned tool can drop to a cheaper model or a lower effort level and keep working under the company’s commercial terms. The engineer keeps moving, and the work stays inside the governed perimeter.

Leaders can also measure displacement directly. Two signals are worth tracking per team: how often engineers hit a limit, and whether sanctioned usage drops sharply after a limit hit while delivery continues. The second pattern proves nothing on its own. It is a reason for a manager to ask where the work went. Pair this with a written policy that names which plans may touch company code, so engineers know that a personal subscription is outside the line, even when the policy cannot be enforced technically.

There is a boundary condition. Hard caps still matter. Stanier points to Simon Willison’s argument for hard spend caps that cut usage off by default, and a company that cannot bound its AI bill has a real financial risk. A hard cap on the company’s total spend can coexist with a soft, step-down limit for each engineer. The failure case is a per-person hard stop with no fallback, because it saves the least money and pushes the most work out of view.

Why This Matters for Engineering Leadership

Token budgets are usually set by finance and enforced by an admin dashboard. They should also be reviewed by whoever owns data governance, because the limit decides which contract governs a given line of company code. Stanier’s point that hidden demand cannot be planned for applies to risk as much as to cost: work that moves to personal accounts disappears from both the budget and the data protection record. Leaders who design limits with a step-down path and track what happens after a limit hit keep both visible.

FAQ