Why Acceptance Criteria Matter
The most common cause of rework in software development is not technical failure but communication failure: the developer built what the story said, not what the product owner meant. Acceptance criteria close this gap by making the intent explicit in testable terms before development begins. They also serve as the basis for quality assurance: a tester can read the acceptance criteria and design test cases without needing to interpret ambiguous requirements. For AI products, they are essential because AI features have failure modes that do not exist in deterministic software. A traditional feature either works or it does not. An AI feature may work in 80% of cases and fail silently in 20%. Without explicit acceptance criteria that define what percentage of cases must work, at what quality level, how edge cases are handled, and what user feedback is provided on failures, there is no objective way to determine whether the feature is complete. Acceptance criteria force these conversations to happen before the feature is built, not after it is shipped.
The Given-When-Then Format
The Given-When-Then (GWT) format, also used in behaviour-driven development (BDD), structures acceptance criteria as scenarios that describe the context, an action, and an expected result. Given describes the starting state or context. When describes the trigger or action the user takes. Then describes the expected outcome. For example: Given a user has uploaded a PDF document to the analysis queue, When the analysis job completes, Then the user is shown a summary of the document's key points with each point linked to the relevant page number. This format makes acceptance criteria testable by design: for every criterion, you can write a test that sets up the Given state, executes the When action, and asserts the Then outcome. It also makes criteria unambiguous: 'the document is analysed correctly' is not an acceptance criterion. 'The user is shown a summary with key points linked to page numbers' is. The GWT format forces precision at the right moment, before development begins.
Writing Acceptance Criteria for AI Features
AI features require acceptance criteria that address output quality, not just functionality. Functional acceptance criteria for an AI classification feature might specify: when the user submits a support ticket, the system returns a predicted category within 2 seconds. Quality acceptance criteria add: the predicted category is correct for at least 85% of test cases in the evaluation dataset. Edge case acceptance criteria add: when the ticket content is ambiguous or insufficient to classify, the system returns an 'unclassified' category rather than a low-confidence prediction. Failure handling criteria add: when the AI service is unavailable, the system queues the ticket for manual review and notifies the user that classification will be delayed. Each of these criteria addresses a different dimension of AI feature completeness: speed, accuracy, robustness, and resilience. Without all four dimensions addressed in the acceptance criteria, a sprint can close with an AI feature that is technically complete but not actually ready for production use.
INVEST: Criteria for Good Acceptance Criteria
Just as user stories can be evaluated against the INVEST criteria (Independent, Negotiable, Valuable, Estimable, Small, Testable), acceptance criteria can be evaluated against their own quality attributes. Good acceptance criteria are specific: they describe a concrete, observable outcome rather than a vague quality. They are testable: there is a clear method for determining whether the criterion is met. They are non-redundant: each criterion adds a distinct condition not already covered by another. They are complete: together the criteria cover the most important dimensions of the feature's expected behaviour. And they are agreed: the product owner, developer, and tester all understand and accept the criteria before development begins. Common failures in acceptance criteria writing include criteria that are too vague to test (the AI response should be helpful), criteria that specify implementation rather than behaviour (the system should use GPT-4o to classify the ticket), and criteria that omit the most common edge cases because they were not thought of during story writing.
Acceptance Criteria and Compliance Requirements
For UK and EU products in regulated sectors, acceptance criteria serve a dual purpose: they define functional completeness and they document compliance implementation. GDPR compliance for an AI feature can be partially expressed through acceptance criteria. For a feature that processes personal data, relevant criteria might include: the system does not store personal data from user inputs beyond the session, the user is shown a clear disclosure that their input is processed by an AI system, and the user can request deletion of all stored personal information through the account settings. FCA consumer duty requirements for AI-generated financial information can be expressed as acceptance criteria: all AI-generated content that could be interpreted as financial advice is accompanied by a regulatory disclaimer. ICO guidance on AI transparency can be operationalised as: users interacting with an AI chatbot are informed at the start of the session that they are interacting with an AI system. Expressing compliance requirements as acceptance criteria makes them testable and auditable, which supports both internal compliance management and external regulatory review.
Acceptance Criteria at SpeedMVPs
At SpeedMVPs, acceptance criteria are defined during the discovery sprint for every engagement, before a line of code is written. For AI features, we include quality thresholds measured against representative evaluation datasets, edge case handling specifications, and error feedback requirements alongside functional behaviour criteria. The acceptance criteria document is the primary scope contract for the engagement: if it is in the criteria, it is in scope. If it is not, it is not. This discipline protects delivery timelines and prevents the scope drift that is the most common cause of AI MVP delays. Acceptance criteria are included in the delivery documentation so clients have a complete, testable record of what was committed to and what was built. Projects from GBP 8,000, 2-3 week delivery, full code and documentation ownership on completion.