Insights

What an AI Agent Actually Costs

Uber burned through its entire 2026 AI budget in four months, an extreme case, but an instructive one. Why AI agent costs behave so differently from classic software, and what that means for margins.

5 September 2026 · 5 min read AI AgentsFinance Auf Deutsch lesen Share on LinkedInShare on X

In February, 32% of Uber's engineers were using agentic coding tools. By March, it was 84%. By spring, 95% were using AI tools monthly, and roughly 70% of committed code came from them. That's impressive adoption by any measure. It's also why Uber had burned through its entire 2026 AI budget by April: four months in, not twelve.

How exactly a year's budget disappears in four months

Average monthly costs, per consistent reporting, landed at $150 to $250 per engineer, manageable. Heavy users, though, generated bills between $500 and $2,000 a month, and one executive reportedly managed to spend $1,200 in a single two-hour coding session. Making things worse: an internal leaderboard ranking teams by AI usage, an incentive system that produced exactly what it was built to produce, more consumption, not more value. Uber COO Andrew Macdonald names the actual gap directly: "That link is not there yet" — the connection between AI spend and real consumer benefit simply isn't established yet. Uber has since introduced spending caps of $1,500 per tool per month.

The important part isn't the anecdote itself. Uber is an extreme case with thousands of engineers. What matters more is the mechanism: agent costs are variable, not fixed, and that's a fundamental break from the software-licensing logic finance teams have relied on for decades. A seat costs what it costs. An agent costs what it does, and "what it does" is hard to predict without a cap.

Not everyone sees the same picture

Worth a brief detour here, in fairness: not every analysis paints the bleak picture the first article in this series drew from MIT's 95% figure. a16z explicitly pushes back on MIT: "Based on our internal data and conversations with corporate executives, we find that statistic hard to believe," Kimberly Tan wrote in April 2026, pointing instead to roughly 29% of the Fortune 500 and 19% of the Global 2000 already being paying customers of leading AI vendors.

Two things worth noting here, in the spirit of the critical sourcing from the earlier articles. First: a16z itself is invested in many of the AI startups this discussion is about. That doesn't make the number wrong, but it's worth knowing. Second, and more important: being a "paying customer" is adoption, not value realization, exactly the distinction MIT and BCG independently drew in article one. By a16z's yardstick, Uber is an adoption success (95% usage). By its own COO's cost-control standard, it isn't, as of now, a proven case of value. Both statements are true at the same time.

Why cost often isn't even the real problem

A team led by Arvind Narayanan (Princeton, known for, among other things, "AI Snake Oil") showed something more fundamental with the paper "AI Agents That Matter": most agent benchmarks optimize purely for accuracy and ignore cost entirely. The authors find that state-of-the-art agents are needlessly complex and costly, and that much of the community draws the wrong conclusions about where accuracy gains actually come from. Their central finding: jointly optimizing cost and accuracy, instead of accuracy alone, can drastically cut effort without losing performance.

In practice, that means a large share of what an agent "costs" isn't the price of the task. It's unnecessary complexity nobody ever questioned, because nobody ever asked it to. Not surprising, given that most teams (see article two) aren't even tracking the metrics that would answer that question in the first place.

What that means for margins if you're building, not just buying

For founders who build AI products rather than just use them, this has a second, more direct consequence. Bessemer Venture Partners, in its AI Pricing and Monetization Playbook (February 2026), puts it plainly: AI companies typically see 50 to 60% gross margins, versus 80 to 90% for classic SaaS. The reason is the same as Uber's, just viewed from the seller's side: every request incurs real, variable compute cost. The COGS line that software licensing let you ignore for decades is suddenly central again for AI products. Bessemer's core line captures it well: if the unit economics don't work at 10 customers, they definitely won't at 1,000.

Three questions before you scale

From the Uber lesson, the Princeton finding, and what I see with founders, a small, concrete checklist falls out, deliberately meant to complement the three-question framework from article two, not replace it:

  • Is there a spending cap before there's a usage incentive? Uber's leaderboard went live before any spending cap did. Incentive without a ceiling is an invitation for cost to explode, regardless of how good the tool is.
  • Is complexity being questioned, or just accuracy? Per Narayanan's team, the question "do we actually need this extra tool call, this extra retry loop" pays off more often than most teams assume, and the answer is frequently no.
  • Is the cost line tied to the same single metric already defined in article two? If the spend curve is rising but the pre-defined success metric isn't moving, that's not growing pains. That's the kill signal.

Adoption is easy to measure and feels good, Uber proves that vividly. Whether it translates into value only becomes clear once someone puts both numbers side by side: what it costs, and what it actually delivers. That connection is still missing at most projects, at Uber, and, if the data from article one is to be believed, at 95% of everyone else too.

Sources:

Case study

Counterpoint (see framing in text)

Research

Venture Capital

Direct follow-on

More Insights
Albert Schaper
About the author

AI expert, entrepreneur, and founder with a finance background and a focus on execution. I build companies, invest in ventures, and advise teams on putting AI to work. LinkedIn

← All Insights