It's Not a Budget Story. It's a Measurement Story.

In February, Accenture told senior staff that using its AI tools would count toward promotion. In October, it told many employees it would centrally fund $29 a month of Claude per person. The easy read is that AI got expensive and a company tightened its belt. The real story is that the measurement flipped without anyone announcing it, and the employees who followed the old one are holding the risk.

Budgets are reasonable. AI shouldn’t get an unlimited line item. But a budget is also a measurement, and people behave according to how they’re measured.

From token maxing to token rationing

Token maxing, the instinct to throw more context, more agents and more model calls at a problem because the marginal cost felt close to zero, had a short shelf life. The invoices arrived, finance noticed, and in a lot of organizations the conversation flipped from “use it everywhere” to “who approved this spend?”

Accenture is the clearest example so far. 404 Media reported on October 15 that the firm now centrally funds $29 a month per person for Claude for many employees, effective October 1. Some of those employees had been spending as much as $1,000 a month. Anything above the allowance has to be charged to a specific project or team. Staff on a Premium plan get $250 a week in usage, but many are on the General plan.

To be fair to Accenture, the company told Forbes that describing this as a firm limit on access is incorrect. It’s a funding model, not a shutoff. That distinction matters less to the person on the General plan who now has to ask a project lead for permission to keep working the way they were told to work. This is the same firm that has been urging clients to invest heavily in AI.

The timeline is the problem. In February, the Financial Times reported that Accenture told associate directors and senior managers that use of its key AI tools would be “a visible input to talent discussions,” and Fortune reported the firm had started monitoring the weekly AI logins of some senior staff. CEO Julie Sweet later said on the Rapid Response podcast, “If you want to get promoted, you’ve got to do the things that we do in order to operate Accenture.”

So in February, using AI was a condition of advancement. In October, using it heavily means asking someone else to pay. Nobody announced that the yardstick changed. The people who built their work around the old one found out from a licensing email.

The risk lands on the people who followed instructions

This is asymmetric risk. The person cutting the budget gets an immediate, visible win: a smaller number on this quarter’s spreadsheet. The costs show up later and somewhere else, on the people whose daily work now runs through these tools.

Think about who reorganized their work around AI over the last two years:

• The engineer who now reviews and directs code more than they type it
• The analyst who rebuilt a reporting workflow around a model
• The writer or PM whose throughput doubled and whose role was redefined to match

They adapted because they were told to, and many were measured on it. Pull the tool back abruptly and their output drops, their skills look dated, and their performance review doesn’t mention that the floor moved under them. The decision-maker never sees that cost, because it never lands on their scorecard.

I’m on that list, with a twist. At work, budget has never been my constraint. Choosing is. I have Claude Code, Microsoft Copilot, GitHub Copilot, Cloudflare’s newer AI products (including a customizable “company OS” that works like a Claude built for one organization), and a shelf of internal tools. I’ve been settling on GitHub Copilot alongside Microsoft Copilot, which I expect to converge, with Cloudflare integrations and our own models behind them. Now I’m hearing that one of those subscriptions may not be renewed. If it goes, I either rebuild around what’s left or ask for an exception, and I’m the one holding that risk.

It’s the driver waving you through an intersection. They can’t see the car in the next lane, and they won’t be the one who gets hit.

A cut can add the limitation back

I’ve been listening to Eliyahu Goldratt’s Beyond the Goal, and one idea keeps pulling me back: technology delivers value only when it diminishes a limitation. Adopt it but keep the rules you invented to cope with the old limitation, and you’ve added cost and complexity and gotten nothing back.

AI went through that trap first. Many organizations bought the tools and kept the old rules: the same headcount math, the same approval chains, the same way of measuring output. Usage climbed, value was hard to see, and the line item became an easy target.

Now comes the second-order version. Teams did change their rules. They built new workflows on the assumption the tool would be there. Cutting it doesn’t return them to the old state. It strands them between two systems: the old rules are gone, and the new capability is rationed. That’s technology adding a limitation, the opposite of the point.

Goldratt also argued that most constraints are policies, not physics. A token budget is a policy constraint. It can be designed well or badly, and right now a lot of them are being designed in a hurry.

His best-known line, from The Haystack Syndrome, is the title of this post in other words: “Tell me how you measure me, and I will tell you how I will behave.” Measure senior staff on weekly logins and you get weekly logins. Bill the heavy users back to their projects and you get people rationing the tool you made a condition of promotion. Neither behavior is irrational. Both are the measurement working as designed.

Dependency is its own fragility

I’d be guilty of my own complaint if I only argued one side. If someone can’t do their job without a specific model from a specific vendor at a specific price, that’s fragility, not productivity. Vendors change pricing. Models get deprecated. Outages happen.

Some token maxing deserved to be cut, too. Not every agent loop was producing value, and “I used a lot of AI” is not an outcome. A budget forces the question we should have been asking all along: what limitation did this remove, and what is it worth?

So the problem isn’t budgets. It’s cuts made on cost alone, by people who don’t carry the consequences, without asking which uses changed the work.

Change the yardstick before the tool

For leaders setting the budget:

• Cut by value, not by volume. Ask which workflows AI changed and protect those. Trim the experiments that never stuck.
• Put the risk owner in the room. The people whose work depends on the tool should help design the constraint, not learn about it in an email.
• Change the measurement before you change the tool. If people were rewarded for AI-driven throughput, a cut without a new yardstick punishes them for following instructions.
• Give a runway. Workflows took months to build. They shouldn’t be unwound in a billing cycle.

For practitioners:

• Know what limitation your use removes. If you can’t say it in a sentence, a budget owner can’t either.
• Keep the underlying skill warm. Use the tool to go faster, not as a replacement for understanding.
• Avoid single-vendor lock-in where you can. Portability is cheap insurance.
• Write down the tradeoff when you’re overruled. Name the risk and who owns it. That turns a silent decision into an accountable one.

The budget is the next yardstick

Budgets will keep tightening, and they should. Organizations that change the cost model without changing what they reward will save money this year and pay for it in stranded workflows, dated skills and employees who stop believing the next initiative. Organizations that line up the budget, the incentive and the value will spend less and keep the gains.

The question isn’t whether to budget for AI. It’s whether the people writing the budget can see what the people using the tool can see, and whether they’re willing to look.

Resources

• Accenture Tells Employees They Can Now Spend Only $29 on Claude Per Month, 404 Media
• Accenture Clarifies Claude Spending Limits As AI Cost Questions Grow, Forbes
• Accenture caps staff Claude spending at $29 a month, The Next Web
• Last year, Accenture trained 550,000 workers in AI—now it’s warning senior staff to use it or don’t get promoted, Fortune via Yahoo Finance
• If you want a promotion at Accenture, CEO says you’ve got to use AI, Fortune
• Eliyahu M. Goldratt, Beyond the Goal, audio lectures, 2005
• Eliyahu M. Goldratt, The Haystack Syndrome, 1990