By: Audrey Denise B. Cachuela
Reports say Uber exhausted its entire 2026 AI coding tools budget in the first four months of the year. (Source: Fortune, 2026) Employees followed the company’s encouragement and embraced the tools, and consumption outran the budget.
An internal leaderboard ranked teams by AI tool usage, and usage climbed. Then Andrew Macdonald, Uber’s president and COO, questioned in public whether the extra consumption produced more useful features for customers. He could see the activity and could not tie it to outcomes he could measure. (Source: Fortune, 2026)
“Uber didn’t have a spending problem. It had a visibility problem,” says Rajeev Jaswal, co-founder and CEO of Argo Intelligence. “Usage was the only number anyone could see, so usage became the goal.”
If you plan to deploy agents across engineering or customer service, you are heading toward the same spot. A successful rollout can send spending up much faster than business results, so executives who approve another round of agents need three things first: a cost figure for each job, a live view of consumption during the work, and the freedom to swap models when the economics stop working.
Why Agents Break the Software Budget
Software licenses are friendly to finance teams. Multiply the seats by the monthly price, add planned headcount growth, compare the total with the expected benefit, and the annual commitment is easy to defend. Agentic AI breaks that arithmetic because you are buying work performed, and you cannot know the amount of work in advance.
A conventional chatbot processes a prompt and returns a response, whereas an AI agent may interpret a task, pull information from several systems, draft a plan, carry out multiple actions, inspect the results, and repeat steps whenever something fails. Each trip to the underlying language model consumes tokens, the units many AI providers use to calculate charges, so AI token costs grow with each extra step the agent takes.
“As a CIO, I could tell you to the dollar what every software license cost,” Jaswal says. “With agents, the bill depends on how hard the work turned out to be, and nobody knows that until it’s done.”
Three factors set the token cost of a run: the quality of the model selected, the latency required, and how often a task runs or loops, which ranges from a one-step lookup to an open-ended job that repeats thousands of times. (Source: McKinsey & Company, 2026) A finance team can fix a seat price, but it cannot fix how often an agent loops.
Unit prices keep falling while enterprise AI spending climbs, which catches anyone who budgets from a price sheet off guard. (Source: McKinsey & Company, 2026) Forecasts say inference on highly sophisticated AI models will cost AI firms 90% less by 2030 than in 2025, but agentic models need far more tokens per task, and providers will pass only part of the savings on to customers. (Source: Fortune, 2026)
As of early July 2026, a frontier model such as GPT-5.5 costs about $5 per million input tokens and $30 per million output tokens, against $0.20 and $1.25 for GPT-5.4 Nano, an earlier lightweight model. (Source: McKinsey & Company, 2026) That spread turns AI model pricing into a finance decision as well as a technical one: running the priciest model on every task wastes money when a cheaper option handles routine work at acceptable quality.
Enterprise AI Spending Is Outrunning Plans
Uber has plenty of companies. In a survey of 1,636 global enterprise technology decision-makers, 46.9% reported AI spending above their planned budgets, with 35.6% moderately over and 11.3% substantially over. (Source: The Futurum Group, 2026)
Companies that overspend keep spending. Among organizations running over plan, 47.6% asked for more budget, 43.3% absorbed the overrun and planned to settle it later, and 17.2% paused or reduced the AI initiative. (Source: The Futurum Group, 2026)
Many companies cannot see where their AI money goes. About 11% of surveyed organizations did not know what their business functions spent on AI in 2025, and 22% had scaled AI across multiple business units or adopted an AI-first approach. (Source: Gartner, 2026) One in ten organizations in the Futurum survey has no formal AI budget to measure against. (Source: The Futurum Group, 2026)
Returns look uneven: 82% of surveyed C-suite leaders are increasing AI investment, but 23% report widespread, sustained business value from the technology. (Source: Accenture, 2026) That gap becomes a budget risk when finance cannot show what an agent costs per finished job.
Three Controls to Put in the Contract
Overruns and weak reporting both have contractual fixes, and learning how to control agentic AI costs means putting those fixes in place before agents run at scale. The controls below belong in vendor contracts, internal deployment policies, or both, and the software has to enforce them.
1. Set AI Spending Limits Per Job
An annual AI budget states what a company plans to spend. It cannot stop a single agent that starts burning through resources. “A budget for the year doesn’t stop one agent from looping at 2 a.m.,” Jaswal says. “A limit on each job does.”
Suppose an enterprise uses an agent to analyze customer support tickets and summarize recurring issues. Finance and IT might agree that each analysis should cost no more than $2 in model charges, with the system tracking consumption as the agent works. When the job nears its limit, the system should halt further billable operations and report what remains unfinished, leaving an authorized employee to decide whether the business value justifies more spending.
Procurement teams should ask vendors how their systems enforce per-job limits: whether a maximum budget attaches to one job, whether execution stops automatically, and who can approve an increase. The contract should spell out whether the ceiling covers model inference, retrieval, tool usage, and other metered services, because a cap on a single billing category can leave charges flowing elsewhere.
2. Make Consumption Visible at the Workflow Level
A report that says the company spent $100,000 on AI last month satisfies accounting and leaves the CFO guessing whether the money was well spent. A CFO needs to see which business processes consumed the budget, what individual jobs cost, and whether those jobs delivered the intended results.
Imagine agents handling contract review, employee onboarding, and technical incident analysis. Each workflow carries different complexity and expected benefit, and lumping them into one AI expense line hides which ones pay for themselves. Finance should attribute spending to the workflow, the model, and the completed job.
3. Preserve the Right to Change Models
The third control covers a risk that stays hidden until a vendor changes its prices. A company that builds an agent around one model provider can find that provider hard to replace, because a cheaper alternative may demand integration changes, fresh testing, and operational disruption. That is AI vendor lock-in. Jaswal, who spent 25 years in enterprise IT including CIO roles at Rapid7 and Red Hat, puts the buyer’s problem in two lines. “The invoice is the smallest number in the conversation,” he says. “The number that matters is the exit cost.”
Consumption pricing also creates a conflict of interest. A vendor that supplies the model and bills per token earns more when agents use more. “If your vendor makes money on every token, ask what they do to help you use fewer,” Jaswal says. “Per-job limits and a record of what each job consumed should be standard. Hold every vendor to that, us included.”
Reducing AI vendor dependency starts with the ability to test competing models on the same workflow and switch when pricing, quality, or operating needs change. A central team can evaluate models, tools, and providers, negotiate commercial terms, and spot lower-cost alternatives as they emerge. (Source: McKinsey & Company, 2026)
Contracts should say whether customers can use their own model accounts, choose alternative providers, and keep their workflow definitions and execution records when they change vendors.
How to Forecast Spending and Measure Cost per Outcome
Once the controls exist, AI budget forecasting can use real workflow data. Start the forecast with the work agents will perform, and count users or licenses second.
Consider a company planning to automate 40,000 monthly invoice reviews. During a controlled pilot, IT measures the average token consumption, model charges, retries, and other billable operations behind each completed review. If the average model and tool expense comes to $0.40 per completed job, monthly consumption charges reach $16,000.
Finance should also test what happens when volumes climb to 60,000 or 80,000 reviews, or when complicated invoices push per-job costs higher. An agent that costs $0.40 per review may save real money. At $4 per review, a $160,000 monthly bill at 40,000 reviews, the savings may disappear.
A forecast built only on tokens misses most of what it costs to get work done. In many banking customer service tasks, token costs make up 20 to 25 percent of an AI agent’s variable run costs, and human oversight makes up 70 to 75 percent. (Source: McKinsey & Company, 2026) A company can cut the tokens an agent uses, raise the human correction needed afterward, and end up with a smaller invoice and a costlier process.
“The cheapest token is the one that doesn’t create work for a person afterward,” Jaswal says. “Track what it costs to finish the job, not what the model charged.”
For an invoice review agent, that total covers processing charges, failed attempts, employee intervention, and the time spent correcting errors. Redesigning the workflow to cut exception rates and simplify review often saves more than switching models. (Source: McKinsey & Company, 2026)
The same logic applies to AI coding tools, the category that drained Uber’s budget. More generated code does not guarantee better software, so finance should ask for evidence of completed business outcomes before treating higher consumption as grounds for a bigger budget.
Make Cost Controls a Buying Requirement and a Habit
Agentic AI stays affordable only if enterprises keep control of the systems that generate their bills. Finance can build a plan from forecasts and cost-per-outcome figures, but only the contract gives it the power to carry that plan out. Buyers have the most negotiating room before signature, so the original request for proposal is the place to ask for each control above.
Contract controls need a management routine behind them. Organizations built a FinOps capability to manage cloud spend, and they now need the same discipline for agents: a cross-functional team that reviews AI unit economics each quarter and decides which workflows still justify their spend and which to redesign, consolidate, or move to lower-cost models. (Source: McKinsey & Company, 2026)
Argo Intelligence, the enterprise AI platform co-founded by Jaswal, is one example of how a product can approach these controls. It ships with models and inference included, can run on customer-controlled infrastructure, and lets organizations bring their own model or point it at their own compute.
Before you approve your next enterprise AI rollout, put these controls in front of finance, IT, and procurement, and ask each vendor to demonstrate them with live examples before you sign. Hold every vendor you evaluate, Argo Intelligence included, to the same standard, and get the answers in writing before the first large invoice arrives.
Sources
External sources published in 2026.
- Fortune. “Uber Burned Through Its Entire 2026 AI Budget in Four Months. Now Its COO Is Questioning Whether It’s Worth It.” May 26, 2026.
- McKinsey & Company. “Where AI Agents Pay Off: A Practical Guide to the Economics of Agentic Workflows.” August 24, 2026.
- The Futurum Group. “46.9% of Enterprises Report AI Spend Over Budget in 2H 2026.” September 10, 2026.
- Gartner. “Gartner Survey Finds Only 22% of Organizations Have Successfully Scaled AI Across Multiple Business Units.” September 1, 2026.
- Accenture. “Pulse of Change.” July 2026.







