Cognition's $10M productivity guarantee bets on Devin's measurable value

Cognition AI, the company behind the Devin coding agent, is offering enterprise customers a "Productivity Guarantee" that refunds up to $10 million in credits if the bot doesn't deliver measurable value. The program comes as companies including Uber report burning through annual AI budgets in a single quarter on tools like Claude Code and GitHub Copilot.

Cognition trained a calibration model on a sample dataset to predict two things: whether Devin completed a valuable task, and how many developer hours that task would have taken a human. The company multiplies those hours by an average developer rate, aggregates the results over a long evaluation period, and compares the total to what the customer paid. If the estimated value falls short, Cognition refunds the difference in future credits — up to $10 million per customer.

What stands out

• Calibrated value estimation: A trained model scores each Devin task for completion quality and estimated developer-hour savings, multiplied by a standard hourly rate.
• Long-horizon aggregation: The model's errors are unbiased, so over a full evaluation period (typically a quarter or longer) the aggregate estimate converges to a reliable figure.
• Credit refund, not cash: If the aggregated value undershoots what the customer spent, Cognition issues the shortfall as credits for future Devin usage, capped at $10 million.
• Enterprise-only launch: The guarantee is available to enterprise customers only — no self-serve tier yet.
• : The move highlights a broader enterprise complaint: companies want outcome guarantees, not just per-token billing, especially as Opus-class models drive costs higher than initial budgets anticipated. Whether OpenAI, Anthropic, or other agent vendors follow suit remains to be seen.