Contents

THE AI-NATIVE SOFTWARE DEVELOPMENT PLAYBOOK

Chapter 03

From business intent to engineering requirements

Turn ambiguous requests into verifiable engineering intent.

Make business outcomes, decisions and acceptance criteria explicit before building.

Start with the decision, not the prompt

A vague instruction creates a wide search space. A useful AI-native process begins with a clear problem statement, measurable desired outcome, stakeholders, constraints and the evidence needed to accept delivery. AI can propose questions and draft specifications, but it must not invent policy where the organisation has not made a decision.

The five-part intent brief

For ServiceFlow, a working objective might be: “Reduce untracked internal requests by making every submission visible to its owner and approver.” It is not yet a complete specification. The analyst still needs to define request types, authorisation rules, escalation windows, reporting definitions and exceptions.

Use AI to expose gaps

Give the agent the brief, current process map and known system constraints. Ask it to return a list of ambiguities, assumptions, dependencies, risks and suggested acceptance tests. Separate confirmed facts from proposals. Record decisions explicitly, including who approved them and when.

RequirementProposed acceptance evidence
An employee creates a requestAuthenticated submission creates an ID and audit event
A manager approves a requestOnly the assigned authorised approver can change approval status
A user checks progressRequester can see own requests, not those of unrelated employees
An administrator audits changesTime, actor, action and result are retrievable

A compact handoff contract

Before implementation, require a reviewed scope, acceptance criteria, data classification, non-functional expectations, dependencies and a rollback or recovery consideration. Decompose work into small changes whose effects and tests can be understood. AWS AI-DLC uses collaborative clarification and shorter work units, but a team can adopt that principle without renaming its entire Agile practice.

IN PRACTICE An agent may identify missing decisions and recommend alternatives; a named product or business owner must resolve material ambiguity.

Research basis: AWS [2–3], OpenAI [6], Anthropic [1].

Requirements are also a test design input

The most useful acceptance criteria describe externally visible behaviour and explicitly include invalid and adversarial cases. “A manager can approve a request” is not enough. Consider: what happens when the requester approves their own submission, when an unrelated manager guesses the request URL, when a role is removed during an active session, or when the same approval is submitted twice? These questions convert an attractive prototype into a controllable workflow.

Write requirements so evidence can be attached to each one. A simple traceability chain connects intent, requirement ID, implementation change, test result and release record. Not every small team needs an enterprise requirements platform; a version-controlled markdown file and linked pull request may be sufficient, provided the links are reliable and decisions are accessible.

ServiceFlow: an explicit decision record

Decision SF-ADR-003: A request cannot be approved by its requester. Rationale: separation of duties. Alternatives considered: self-approval below a value threshold; delegation to an alternate manager. Consequences: approval endpoint must check authenticated identity against request ownership; the UI alone is not a security control. Owner: product manager and security approver. Status: approved before development. This brief record prevents an agent from inventing an approval policy.

top