Microsoft's engineers loved Claude Code. That turned out to be the problem.
The company handed the AI coding tool to around 5,000 engineers in December. By spring, 84 to 95% of them were using it. A dream adoption rate. But the pilot ran on flat licences, so the real cost stayed hidden, until the billing switched to usage-based and the true number landed: $500 to $2,000 per engineer, per month.
And it wasn't alone. Uber's CTO said the same thing out loud: the company burned its entire planned 2026 AI coding budget in four months.
The division's annual AI budget was gone in a matter of months.
Why the bill changed, and why yours will too
You might file this under big-tech problems, the kind of enterprise deal that never touches your $19 Copilot seat. It was, until June.
Here's what actually drove it. Flat-rate AI was quietly bleeding the people who sell it. A heavy user running agents on a flat subscription was consuming an estimated 15 to 30 times the compute they paid for, and no vendor subsidises that forever. So the price moved to match the real cost. Anthropic put heavy Claude Code usage onto metered rates, which is the bill Microsoft suddenly had to look at.
And it didn't stop at big contracts. On 1 June, GitHub Copilot itself moved to usage-based billing. Your seat still costs the same $19 or $39. But it now comes with a monthly credit budget, and once your team works through it on the stronger models and the agentic features, you pay per token from there, exactly like Microsoft did.
A couple of weeks back I made the case that an AI agent is really a hire, not an IT cost, something you onboard and manage like a colleague. If that's right, then this is payday.
The catch
So the 2026 question isn't about capability. It is quieter and it's about money. Not "can it?" but "does it pay?"
And here's the trap both Microsoft and Uber fell into: adoption is not the same as return. A tool 90% of your team uses every day feels like a win. But a genuinely engaged team runs up a bill far past what the seat-count maths predicted. Popularity is a promising signal. It is not proof the money came back.
What BCG suggests you do about it
The useful part is that there's now a clear framework for this, and it's simple enough to use on Monday. BCG's advice starts by killing one bad habit: treating "AI spend" as a single line in the IT budget. Every pound you spend on AI is really one of three things, and each is judged differently.
Investment. Money spent building something reusable, like an agent or a redesigned workflow. Treat it like a capital project. You don't expect a return this week, you expect what it builds to earn for months.
Running cost. Money spent on internal work, the drafting and summarising and day to day. Give each workflow a budget and an owner, the way you would a marketing line. Then judge the owner on output, not usage.
Cost of goods. Money spent inside a product or a customer interaction. This one hits gross margin directly, and it's the one most P&Ls miss. Bolt an AI feature onto an $80 seat and the inference behind it can add $15 of real cost, and the margin drops from 80% to 65% overnight. Watch this in COGS, not IT.
Sort your spend into those three and the ROI question stops being vague. BCG gives the ratio a name:
© BCG
In plain terms, cost per realised outcome. Take a workflow, add up everything it costs, and divide by what it genuinely delivered, hours truly saved, work that shipped, tickets that didn't come back.
To sum up
This is not an argument against AI. Microsoft, Uber and BCG all agree the tools are worth having. It's an answer to one question: how to read the meter.
Call to action? Pick your most-used AI workflows and sort them into separate buckets. Work out their cost per outcome. Then you'll know which ones to fund, which to cap, and which to retire or adjust to a different model.