70% of committed code AI-generated, annual AI budget exhausted in four months, zero per-team attribution. This is not a hypothetical. It is what happens when you set "AI adoption" as a target metric without building the financial infrastructure to govern it.
In 2023, Uber published an internal engineering memo describing exactly this dynamic. Engineers were measured on AI adoption — how much they used AI coding tools, how much they integrated AI into their workflows. Rational engineers, responding to rational incentives, maximized the metric. They used AI tools constantly, aggressively, on every task regardless of marginal value. Monthly AI spend per engineer reached $500–$2,000 depending on role and workflow. The annual budget projection, calibrated in Q4 of the prior year, was burned through by April.
This is Goodhart's Law in its most expensive form: when a measure becomes a target, it ceases to be a good measure.
The problem isn't that engineers used AI tools. The problem is that there was no feedback loop connecting usage to cost to team. Finance saw a budget line item: "AI Tools — $X/month." Engineering saw a usage metric: "AI adoption — Y%." Neither team had a view that connected them.
The CFO could not answer: which team is driving 40% of our AI spend? Which use case costs $18 per request when the median is $0.40? Is the backend team's usage returning proportionally more value than the data team's? Without per-team attribution, these questions are unanswerable — and the budget conversation reduces to a single lever: cut or don't cut.
That's a terrible lever. "Cut AI spend" means "reduce adoption" means you're now optimizing against the original KPI. You've created a governance paradox: the metric you want to grow is the metric you need to constrain, because you can't see which part of it to constrain.
"I'm back to the drawing board, because the budget I thought I would need is blown away already."
— Uber CTO Praveen Neppalli Naga, April 2026
Most engineering organizations today have the following visibility into AI spend:
This is roughly equivalent to receiving a cloud infrastructure bill that shows total compute costs but has no resource tags — no breakdown by service, environment, or team. Every serious engineering organization fixed that cloud problem five years ago with tagging strategies and cost allocation tools. The AI cost problem is structurally identical, and most organizations haven't fixed it yet.
The reason: AI spend flows through a single API endpoint (or a handful of them) rather than through dozens of discrete infrastructure resources. There's no natural tagging boundary. Every call to api.openai.com/v1/chat/completions looks identical to the billing system, regardless of which team triggered it, which product feature it powered, or which workflow it served.
| KPI | Without Attribution | With Attribution |
|---|---|---|
| AI adoption rate | A single number — no team breakdown | Rate by team, with cost-per-point context |
| Monthly AI spend | Total provider bill — opaque | By team, use case, and model |
| Budget forecast accuracy | ±40–60% miss typical | Within ±10% with per-team caps |
Attribution requires a layer that sits between your application and the LLM provider — a layer that can:
This is not a complex engineering problem. It's a routing and logging problem. The proxy receives a request tagged with team: backend and use_case: code-review, passes it through to the provider, and records the token cost against those dimensions. At the end of the month, finance gets a breakdown that looks like a cloud cost report — by team, by use case, by model, by period.
The engineering change required is minimal: add a header or metadata field to your existing LLM calls. No restructuring of application code, no changes to model selection logic, no new infrastructure. One proxy URL swap and a tagging convention.
Goodhart's Law doesn't mean you can't use adoption as a metric. It means you can't use it in isolation. The fix is a companion metric that makes the adoption rate legible: cost per adoption unit, broken down by team.
With per-team attribution in place, the dashboard that matters looks like this:
Now you can see that the data team's requests cost 6× more on average than the backend team's — and ask why. Maybe it's justified (complex document processing). Maybe it isn't (using GPT-4 for tasks where Gemini Flash would suffice). The adoption metric is the same; the cost-per-unit metric tells you where to look.
Budget governance also becomes precise. Instead of "cut AI spend by 15%," you can say "data team: requests over $1.00/call require an engineering review." You're not slowing adoption — you're applying routing intelligence to the expensive outlier cases.
There's a second dimension to the Uber failure that attribution alone doesn't solve: the budget cycle mismatch.
Enterprise budget cycles run annually. AI model pricing changes monthly. The Gemini 2.5 Flash price point at announcement was different from the price at GA. GPT-4o's price has changed multiple times since launch. Claude Sonnet 3.5 launched at a price significantly lower than its predecessor. If your annual AI budget was set assuming last year's pricing, and a model you use heavily dropped 40% in price, you may be running significant budget headroom without knowing it — or you may be running over budget because you migrated to a model that's more expensive than the one in your baseline.
Attribution solves the "where is the money going" question. Real-time pricing awareness solves the "is this the right model for this cost" question. Both are required for a CFO to feel confident signing off on AI spend as a budget line.
Trimio's Finance Alpha is built around this problem. The proxy layer adds per-team, per-use-case attribution to every LLM call. The cost dashboard provides the breakdown finance needs for period close. Budget overlays let engineering leads set spend caps at the team level — the data team's cap triggers a routing change (shift to a cheaper model) rather than an outage.
The routing layer also addresses the budget cycle problem: when Gemini Flash drops to $2/M output tokens and your routing rules say "use cheapest capable model," the price drop automatically routes more traffic to Flash. Your budget doesn't need to be re-negotiated — the system captures the savings without a manual intervention.
Goodhart's Law is a warning, not a prohibition. The metric is fine. The problem is running it without the infrastructure to make it legible. Per-team attribution is that infrastructure.
Trimio is the LLM gateway built for teams where AI adoption is the goal — and AI cost governance is the prerequisite. Per-team attribution, budget caps, and routing intelligence in one layer. See how it works.