What Do These Companies Have in Common? They Get Guaranteed AI Impact

A growing number of large enterprises are discovering a painful truth: AI coding tools can burn through budgets faster than anyone expected.

What Do These Companies Have in Common? They Get Guaranteed AI Impact — Or Their Money Back Uber’s CTO recently made headlines when he revealed the company had already exhausted its entire 2026 AI budget just four months into the year after rolling out Claude Code to thousands of engineers. Similar stories are popping up across the industry — teams disabling GitHub Copilot, Claude Code, or other agentic tools because token spend exploded beyond all forecasts.

The core problem? Companies want real productivity, not just token volume. And now, one AI company is putting its money where its mouth is.

Cognition’s Bold Move: The AI Productivity Guarantee

Cognition (the team behind the autonomous AI software engineer Devin) just announced something remarkable for its enterprise customers: a Productivity Guarantee.

Here’s how it works:

Cognition built a specialized AI evaluator that reviews every completed Devin session.
For each session it answers two questions:

1. Did this produce genuinely useful output?

2. If yes — how many hours would a human engineer have needed to achieve the same result?

These hours are multiplied by a standardized developer hourly rate and aggregated over a long period (typically the annual contract).
At the end of the period, if the total estimated value is lower than what the customer paid, Cognition refunds the difference in credits — up to $10 million.

It’s the first time a major AI agent company has offered a true performance-based financial commitment at enterprise scale.

How the Estimator Actually Works

Cognition’s technical blog post explains the system in detail. The estimator has access to the full session trace: user prompt, every action Devin took, PRs created, and deep codebase context. It deliberately tries to be conservative and unbiased.

Individual session estimates can be noisy (as expected with any LLM-based predictor), but because errors are roughly unbiased, they cancel out when you look at thousands of sessions over months. The model was validated against real human engineer self-assessments and shows strong correlation in aggregate.

This isn’t vaporware — it’s already running with enterprise customers.

Why This Matters

What Do These Companies Have in Common? They Get Guaranteed AI Impact — Or Their Money Back For years, AI vendors have sold on hype, benchmarks, and usage metrics. Enterprises are now demanding something much simpler: prove that this is worth the money.

Cognition is betting that Devin delivers enough real engineering value that they’ll rarely (if ever) have to pay out the guarantee. By putting a $10M cap on the line, they’re forcing the entire industry to think harder about outcomes instead of just token consumption.

Other players are watching closely. OpenAI and Anthropic, whose models power much of the current agent wave, face similar pressure from customers whose spend has gone parabolic. Expect more outcome-based pricing experiments in the coming quarters.

The Bottom Line

The age of “just use more AI” is ending. The new era is about guaranteed impact.

Companies that adopt serious AI agents now have a new question to ask vendors:

“What happens if it doesn’t actually save us time and money?”

Cognition just gave one of the strongest answers in the market so far.

If you’re a CTO or engineering leader wrestling with AI ROI — this is the conversation you should be having with every vendor.

The tools are powerful.
Now they’re finally being asked to earn their keep.