The Bill That Ended the Pilot

Uber's CTO announced in April 2026 they had burned through the company's entire annual AI coding budget in four months. Not a typo. Not a controlled overrun flagged in Q2 planning. The full year gone by spring, and somebody had to tell the board why the largest line item under Technology Innovation was going back for a second helping.

Nobody at Uber is incompetent. They instrument everything, they know how to model usage, and they run one of the tightest FinOps operations in the valley. What happened there is happening everywhere the POC went to production without the guard rails production actually requires. The debt is coming due, and it is landing on the desk of whoever owns the PagerDuty rotation when an agent loops itself into a $22,000 weekend or a prompt change drops eval scores by 18 points and customer support finds out on Monday morning.

Nearly three-quarters of enterprises watched their AI costs blow past budget last year, according to the 2026 State of FinOps Report. FinOps for AI is now the top forward-looking priority, AI cost management is the number-one skillset that teams need to develop, and 98% now manage AI spend, up from 31% two years ago. That jump is not adoption enthusiasm. It is triage. The bill arrived, and finance wanted to know what they were paying for and who approved it.

The technical shape of the debt is predictable once you have run this loop a few times. Prompt versions proliferate across repositories, environment variables, dashboards, and notebooks, and when an AI feature breaks in production, engineers often spend hours trying to answer a simple question: which prompt version is actually running right now? First Central and PayPal both named the same end state: you have every other person in the organization building their own bespoke agents, many with system prompts which most of them are not too great, and context sprawl where you have 30 different agents and a bunch of different sources of truth.

Gartner estimates that by 2028, the average global Fortune 500 enterprise will have more than 150,000 AI agents in use, yet only 13% of organizations believe they have the right governance in place to manage those agents. You can project that forward and see the shape of the incident review: which team deployed it, what was it allowed to touch, who evaluated the prompt, and what budget envelope was it running under. In the enterprises I work with, those questions take two days to answer because the system that would answer them was never built. The agent shipped because someone had API keys and a Slack thread, not because there was a release gate.

The cost failure and the governance failure are the same failure. When an agent misreads a response or receives a malformed payload, it frequently retries the same action, hits the same wall, and retries again, and without a hard ceiling on iteration count, that retry loop runs until token budgets are exhausted or the orchestration layer times out. One documented case in June showed a writer crashing 51 times in 75 minutes, and a third ticket recorded 97 failures over roughly eleven hours during the same spending-limit incident, prompting the observation that "it self-healed" is a flattering description of an uncontrolled loop.

A 2026 Concordia University study clocked the waste: a 2-to-1 input-to-output ratio, a "communication tax," with code review alone eating 59% of every token spent, and agents burn 5 to 30 times more tokens per task than a chatbot. EY put it in dollars: one customer-service interaction went from $0.04 to $1.20 in three years, 30 times higher, while token prices fell. That is the shape nobody priced during the pilot, because the pilot had ten users and the agent called the API twice per session. Production has six thousand users, the agent loops an average of eight times, and one in twenty sessions triggers the pathological case where it calls the retrieval API 43 times before giving up. The budget assumed two calls. You are paying for 344.

The evaluation gap compounds it. A super-majority of YC agent builders said evals often under-deliver because keeping them up to date becomes an impossible task, and about 38% explicitly raised evaluation challenges: building suites, running A/B tests, improving behavior over time. Prompt sprawl shows up as an incident review with no good answers, when something breaks in production and the incident review asks three questions the tool cannot answer: who changed the production prompt, what did it score before it shipped, and did the customer data in the test set ever leave the network.

Industry analysts have projected that a large share of agentic AI projects will be cancelled by 2027, largely because costs, risks, and value were never measured rigorously, and the overwhelming majority of large organisations are piloting AI agents, but only a small minority have promoted those agents to production scale. The gap is not model capability. The gap is the operating discipline that would make the thing governable once it touches money or customer data. AI technical debt compounds faster than traditional software debt because it does not live in code alone but in data, models, prompts, APIs, workflows, vendors, security controls, compliance gaps, and the undocumented decisions made when speed mattered more than structure.

I have spent enough time in these incident reviews to recognize the pattern. The agent was working in staging. It passed the eval suite. Somebody promoted it to 5% of production traffic on a Friday afternoon, and by Sunday morning it had burned $18,000 calling a vendor API in a retry loop that nobody thought to put a ceiling on. The on-call got paged. They killed the deployment. Monday's post-mortem asks why there was no budget cap, no iteration limit, no diff view showing what changed between the staging prompt and the production prompt, and no centralized registry showing which other six teams are running variants of the same agent against the same API.

The answer is that none of those systems exist, because the POC did not need them. The POC ran on a $2,000 monthly budget, used one prompt edited by one person, touched no production data, and calling the API 40 times in a loop was fine because the API was a mock. Production is not a scaled-up POC. Production is a different threat model, a different cost model, a different ownership model, and a different compliance surface. The systems that would govern production were supposed to be built during the six months between pilot and launch, and instead that time went into adding three more agent types and connecting them to Salesforce.

Field reports from operators who have audited their stacks suggest that 40% to 60% of token budgets in production LLM applications are pure waste, money paid for capability never used or for inefficiencies nobody priced at design time, and that figure is consistent across teams from venture-backed SaaS startups to enterprise platform groups. This is the number every operator holding the pager should take to their next planning meeting, because it is the gap between what the budget assumed and what production actually costs. Half the bill is paying for loops that should have been capped, prompts that are three versions out of date and 600 tokens longer than they need to be, agents that reload 80 KB of tool definitions on every call, and caching that was turned on in the SDK but never configured on the provider side so it is doing nothing.

The FinOps correction is coming in two forms. The top requested tooling capability in the State of FinOps 2026 Report is granular monitoring of AI spend such as tokens, LLM requests and GPU utilization, with one practitioner noting that teams are having to balance the complexity of managing AI spend over a more diverse range of projects with the directive to avoid limiting AI usage, which might slow down time to market. Companies are increasingly issuing developers personal token budgets as a managed resource, and at GTC 2026, Jensen Huang announced that NVIDIA engineers will receive token budgets worth roughly half their base salary, and that for a $500,000 engineer who consumed less than $250,000 in tokens, he would "be deeply alarmed."

That is not a thought experiment. That is the CFO telling engineering that tokens are now a first-class cost category with the same governance surface as head count, and if you want to ship another agent you will show the unit economics and the budget envelope before you touch production. I am seeing this conversation happen in three-quarters of the enterprises Real AI works with. The conversation starts with finance, escalates to the CIO, lands on the senior engineering leader who owns AI, and ends with somebody getting paged at 3 a.m. to explain why an agent burned 140,000 tokens retrying a tool call that was never going to succeed.

The wedge that breaks this open is the incident nobody can defend. A $22,000 weekend is hard to justify when the agent was supposed to summarize support tickets. A prompt injection that exfiltrates customer email addresses ends the pilot the day it is discovered. A customer-visible incident does not appear to stop the move toward autonomy, and respondents with firsthand proof that testing can miss defects are also moving most aggressively to let those tests authorize production changes, so if their per-deployment failure rate remains constant while deployment volume rises, the total incident count could grow even without the percentage of affected companies increasing.

That is the truly grim read: the correction will not arrive as collective learning but as serial disasters, each one large enough to force one organization to build the governance layer they skipped, and everyone else will keep shipping until their own disaster forces the same retrofit. The cost of that retrofit is higher than the cost of building it correctly the first time, and the cost is carried by whoever owns production reliability when the thing breaks.

If you are that person, these are the questions to ask now, not after the post-mortem. Does every agent have a maximum iteration count enforced at the orchestration layer, not model discretion? Is there a budget cap per agent, per feature, per user, per day? Can you diff the production prompt against the staging prompt and see what changed? Is every prompt version immutable, tagged to an environment, and linked to the evaluation run that cleared it? Can you trace a production failure back to the exact prompt, model, tool configuration, and retry policy that caused it? Does the bill break down to feature and team, or does it arrive as one number that finance cannot allocate?

The enterprises that can answer yes are still in the minority, but they are the ones that will scale AI without the CFO pulling the budget after the second unplanned overrun. The ones that cannot are running on borrowed time, and the debt is not theoretical. It is denominated in dollars, and it comes due the moment production traffic hits the thing nobody thought to put a ceiling on.

Tarry Singh is the founder and CEO of Real AI (realai.eu), an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan (earthscan.io) for Energy AI, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.