Cognition is putting a financial promise behind its enterprise AI coding agent Devin, saying customers can receive credits of up to $10 million if the system does not produce enough engineering value to justify what they pay.
The company said the new program, called the AI Productivity Guarantee, is meant to address a gap in the enterprise AI market. While vendors often report activity metrics such as tokens used or code generated, Cognition argues those numbers do not show whether a business is actually getting meaningful work done.
To support the guarantee, Cognition said it built an estimator that reviews completed Devin sessions and attempts to measure productive engineering output. The system looks at whether a session created useful results and, if so, estimates how long a human engineer would have needed to complete the same task. The company says it validated that approach against time estimates provided by users at enterprise customers.
Cognition said its estimator has access to several inputs, including the user’s prompt, the actions Devin took during a session, any pull request that was created, and additional codebase context drawn from DeepWiki. If a session leads to an unmerged pull request or is otherwise classified as unproductive, it is counted as not useful.
The company said it measures output in hours of productive engineering work rather than lines of code, arguing that the latter can be misleading. A small fix can take many hours of debugging, while a larger block of code may not represent much actual effort.
Cognition described the estimator as a baseline for productivity, not a full measure of return on investment. The company said ROI requires more context about the business impact of each task, and that its customer teams work directly with enterprises to understand that broader value.
According to Cognition, the guarantee applies to enterprise deployments of Devin. Near the end of an annual contract, engineering hours are converted into dollar value using a standard global rate and compared with the customer’s actual consumption. If the estimated value falls short of what the customer paid, Cognition said it will issue credits, with the maximum exposure capped at $10 million.
The company said it believes the arrangement is feasible because Devin is model-independent, allowing Cognition to use different models for different tasks in order to improve price performance. It also said Devin includes granular controls intended to manage spending and steer users toward more productive prompts.
Cognition added that its teams often work inside customer accounts, helping identify high-value projects, pair programming on backlogs, running workshops on how to manage fleets of agents, and tracking outcomes.
The company framed the guarantee as part of a broader effort to make AI vendors more accountable for business results. It said vendors should be able to tell customers what they are getting for their money and suggested that the rest of the industry move toward similar commitments.
Cognition is inviting prospective customers to contact the company for more information about the program, while existing customers can speak with their account teams.