What the AI-DLC (AI-Driven Development Lifecycle) is
The AI-driven development lifecycle (AI-DLC) is the delivery method this example was built with. AI does most of the drafting, generating and checking, and humans make the decisions, own the design and sign each gate. The work is organized in three phases, each closed by a gate that someone is accountable for.
Inception: intent, scope and plan
Intent
Business requirements and targets (see overview) written as user stories with acceptance criteria.
Design
ADRs (Architecture Decision Records) drafted with the AI proposing options; the engineering lead decides and signs them.
Plan
Work broken into bolts of one to two days, each with its own evaluation criteria.
Gate: business owner and engineering lead sign the scope and the ADRs.
Construction: build in short, verified loops
Specify
Each bolt starts with failing tests and an evaluation set: expected extractions, match outcomes and cited answers.
Build
AI pairs on implementation; the human reviews every diff and owns the design.
Verify
Unit tests, contract tests for MCP (Model Context Protocol) tools, DAG (Directed Acyclic Graph) tests with stub tools, and evaluation runs on the golden sets.
Gate: all tests and evaluations pass against the thresholds in the bolt; security review for any new tool.
Operations: run, observe and improve
Release
Canary release by vendor cohort; model or prompt changes ship behind the same evaluation gate.
Observe
Dashboards for cycle time, exception rate, extraction confidence, gateway denials and cost per invoice.
Learn
Reviewer corrections become new evaluation cases; drift in exception rate triggers a bolt.
Gate: monthly control review with finance and audit on exceptions, overrides and access changes.
Who does what
| Role | AI does | Human decides |
|---|---|---|
| Business owner | Summarises stakeholder notes and drafts user stories | Scope, priorities, approval limits |
| Engineering lead | Drafts ADR options and reviews diffs for consistency | Architecture, accepts ADRs, merges |
| Developer | Writes tests and implementation from the bolt spec | Design within a bolt, what to reject |
| AP (Accounts Payable) / finance reviewer | Explains exceptions and drafts resolutions | Every approval and posting decision |
| Security and audit | Scans for risky tool scopes and prompt-injection paths | Access grants, control exceptions |
The rule is simple: the AI may recommend anything, but nothing reaches production or moves money without a named human decision.
Artifacts each phase produces
Inception
User stories, requirements table, ADRs, bolt plan
Construction
Tests, evaluation sets, tool contracts, DAG definition, prompts as versioned files
Operations
Runbooks, dashboards, cost reports, monthly control review notes
Always
Traceable links from requirement to ADR to code to test to evaluation result
How we know the lifecycle is working
- Bolts close on their evaluation thresholds, not on reviewer optimism.
- Defects found after release trace back to a missing test or evaluation case, which is then added.
- Each ADR's consequences are checked at the monthly review; a consequence that did not materialise is written down.
- Cost per invoice and exception rate are tracked from the first pilot, so changes can be judged on evidence.
Build this for your business
This example uses a generic invoicing system, but the same pattern — orchestrated steps, governed tools, grounded knowledge and a documented lifecycle — fits most back-office processes. We can scope a pilot on your own documents and systems.
Book a discovery call