Cost Per Run Is The Metric UK Businesses Need Before Scaling AI Agents

Uber burned through its entire 2026 AI coding budget by April. Atlassian's AI bill went from 5 million dollars a month to over 15 million in nine months. Neither company was reckless - both were measuring the wrong number.

The number that's hiding the problem

Ask most UK finance directors what their AI systems cost, and they'll give you a monthly total. That number is almost useless on its own, and the data backs this up starkly. A February 2026 survey of 500 finance leaders across US and UK organisations with over 1,000 employees, run by Sapio Research for cloud cost firm DoiT, found that 79% had experienced AI cost overruns in the past twelve months. Only 26% of organisations report full, real-time visibility into what their AI systems actually cost to operate, according to KPMG's AI Pulse survey of 204 organisations in Q2 2026.

Here's the counter-intuitive part: overruns get worse, not better, as FinOps discipline matures. DoiT's data shows organisations with 'very mature' FinOps practices report AI cost overruns in 89% of cases, against 69% for early-stage teams - and the mean overrun for mature teams runs to 30.9%, nearly double the 16.1% seen in early-stage organisations. Mature teams aren't managing costs worse. They're the only ones with visibility good enough to actually see the overrun happening. Everyone else is running the same problem with the lights off.

Total monthly spend moves with volume by definition. It tells you nothing about whether each unit of work is getting cheaper or more expensive, and it gives you no early warning before the bill arrives. Cost per completed run - the total cost to finish one invoice, resolve one ticket, process one document - is the number that survives growth and actually supports a decision.

89% of organisations with 'very mature' FinOps practices still report AI cost overruns - because maturity means visibility, not immunity.

Why agent costs don't behave like software costs

A traditional software licence costs the same whether an employee uses it once or fifty times a day. An AI agent does not work that way, and understanding why is the difference between a controlled rollout and an open-ended liability.

First, an agent decides its own workload. A conventional API call performs one unit of work for one unit of money. An agent receives a goal and works out what's required to reach it - which means two requests that look identical to a user can differ in cost by a factor of thirty, depending on how much work the model decides the task needs.

Second, context compounds. Most agent frameworks resend the accumulated conversation, tool outputs and working notes on every single step. A ten-step run doesn't cost ten times a single step - it costs closer to the sum of a growing series, because step ten is carrying everything from steps one through nine along with it. This is the single most underestimated driver of runaway agent bills, and it's invisible on any architecture diagram that draws each step as an identical box.

Third, retries multiply rather than add. Retry logic commonly sits at three layers - the HTTP client, the agent framework, and the orchestrator. Three retries at three layers isn't nine attempts of equal cost, because each retried run carries its own accumulated context back into the loop. A transient outage on a Tuesday afternoon can generate a full day's spend in twenty minutes, with nothing in the logs flagged as an error.

Fourth, and most dangerous: nothing in the system reports failure. A runaway virtual machine trips an alarm. A runaway agent produces correct-looking answers. Every individual call is well-formed and successful, so no monitoring rule fires, and the cost accumulates through a system that reports itself as entirely healthy.

A runaway server trips an alarm. A runaway agent just produces correct-looking answers and keeps billing.

The receipts: what this actually looks like at scale

These aren't hypothetical risks. Uber's engineering organisation went from 32% to 84% adoption of AI coding agents in roughly a month during 2026, with 95% of engineers using AI tools monthly. Average spend ran 150 to 250 US dollars per engineer per month, with heavy users near 2,000 dollars monthly, and the company exhausted its entire 2026 AI coding budget by April - four months in. Uber's CTO publicly acknowledged being 'back to the drawing board.'

Atlassian's monthly AI expenditure climbed from roughly 5 million US dollars in August 2025 to more than 15 million dollars by May 2026, driven by what industry commentators are now calling 'tokenmaxxing' - unconstrained agent and coding-assistant usage with no cost ceiling attached. Atlassian has since introduced capped 'AI wallets' limiting research and development spend to between 500 and 2,000 US dollars per employee per month. Microsoft has similarly told staff to curb 'tokenmaxxing' and set division-level AI budgets, and reportedly ended some coding-agent licences six months into a pilot once the token bill became clear. Amazon, Adobe and Citi have all introduced some form of AI usage throttling or spend visibility in recent months, according to reporting on the Microsoft situation.

None of these are stories about badly run technology organisations. They're stories about a pricing model that shifted under the floor. A flat seat licence made token spend invisible because the price didn't move with usage. Consumption pricing moves with usage by definition - and very little in most organisations' cost stacks was built to watch it in real time.

Atlassian's AI bill tripled from 5 million to over 15 million US dollars a month in nine months - then it introduced capped AI wallets.

Building a cost-per-run framework before you scale

The cheapest possible prevention is arithmetic done before a single line of production code exists. Teams that avoided runaway agent costs did one thing differently at the start: they costed a single run in a spreadsheet, including the bad cases, before deployment.

Start by counting steps, not requests. Estimate the median number of steps a run will take, then the ninetieth percentile, then the worst case your step cap allows. Multiply each by average tokens per step - remembering that later steps carry more accumulated context than earlier ones. A model that assumes constant cost per step will understate the true total by a factor of two or more on any run longer than about six steps.

Then build a range, not a single point estimate. Present three numbers to whoever signs off on the budget: expected cost, plausible-bad cost, and capped-worst cost. That third figure is what prevents a cost overrun turning into a governance incident, because it's a promise the architecture can actually keep. If your team cannot state a capped-worst case for an agentic workflow, the design has no cap - full stop.

Finally, set a walk-away threshold before anyone is emotionally invested in the project. Agree the cost per completed run at which the automation stops being worth it, in writing, before go-live. If a human handles the task for four pounds and the agent costs three, a 40% cost overrun erases the entire business case. Writing that number down in advance turns a future argument into a pre-agreed trigger instead of a political fight after the bill lands.

If a human does the task for four pounds and the agent costs three, a 40% overrun erases the entire business case.

What UK finance and ops teams should do this quarter

KPMG's Q2 2026 data shows the real gap isn't awareness - it's the ability to act. 66% of organisations now maintain AI cost monitoring dashboards, and 61% have cost reviews built into their approval process. But only 36% have direct token or usage controls in place. Two-thirds of organisations are watching AI spend. A third can actually intervene when it goes wrong.

Attribution is a specific, fixable failure inside that gap. DoiT's survey found 55% of organisations place AI spend accountability with technology teams and 53% place it with finance - the numbers exceed 100% because both groups assume it belongs to the other, which in practice means nobody owns it. The first practical step for a UK business scaling agents is simply naming a single owner for AI unit economics, with authority to set hard caps, not just visibility into a dashboard.

Second, move budget conversations from monthly totals to cost-per-completed-task figures - cost per invoice processed, per ticket resolved, per document reviewed - so a rising number is visible as it happens rather than at month-end reconciliation. Third, implement actual usage controls at the API or gateway layer: hard step caps, retry limits shared across all three retry layers rather than compounding independently, and spend alerts tied to the walk-away threshold agreed at design time, not to the total monthly bill.

Only 36% of organisations have direct token or usage controls in place - the rest are watching a dashboard they cannot act on.

"We're too small for this to matter yet"

The most common objection from UK SMEs is that this level of cost discipline is over-engineering for a business running a handful of pilots on consumer-tier subscriptions. That confidence has a shelf life. Industry commentary on the 2026 pricing landscape notes that smaller organisations are currently able to stay under the radar on flat-fee personal accounts precisely because usage is still low enough not to attract attention - but that gap is expected to close as providers face shareholder pressure to convert flat-fee usage into metered revenue.

There's also a harder truth in the unit economics themselves. Goldman Sachs projects businesses will consume 24 times more tokens by 2030 than in 2026, reaching 120 quadrillion tokens a month, even as the price per token continues to fall roughly 60 to 70% a year for inference. Falling unit price and rising total bill are not contradictory - volume is winning by a wide margin, and volume is exactly the variable that a flat monthly subscription hides from a growing UK business until the day the provider notices and reprices the account.

The businesses least prepared for that moment are the ones that never built a cost-per-run habit while the stakes were low. Building the discipline into a five-person pilot costs almost nothing. Retrofitting it onto a live customer-facing workflow after a surprise invoice costs a great deal more, in both money and trust.

Goldman Sachs expects 24 times more token consumption by 2030 - falling unit price and rising total bill are not contradictory.