Delivery Playbook · Quality & AI
AI engineering
How AI supports the whole engineering workflow, who is responsible for what, and the rules that keep it safe.
Owner: CTO · Last reviewed: · Next review: 28 April 2027
AI tools make good engineers faster. They also make it easier to produce code nobody understands. Our approach keeps the speed and keeps people accountable for what ships. We want to be paid for results, not for typing.
Three levels of AI adoption
Level 1 · Where it starts
AI-assisted engineering
Individual engineers use approved AI tools to speed up specific tasks, such as drafting code and tests, explaining unfamiliar code or summarizing requirements. The engineer reviews and owns the result.
Level 2 · Where we are building toward
AI-orchestrated delivery
Controlled AI-agent workflows handle bounded engineering tasks, for example a refactor or a test suite. Each has defined permissions, automated verification and clear rules for when to stop and ask a person.
Level 3 · The longer-term direction
AI-enabled operations
AI helps monitor running systems, diagnose incidents, spot recurring defects and technical debt, and suggest improvements. People decide what changes in production.
These are capabilities and working patterns we apply step by step, agreed with each client. They are not a claim that every capability is already in use on every project.
Where AI helps across the workflow
AI can support the whole engineering workflow, not only code writing. In every case a person stays responsible for the result.
| Activity | What AI can do | What stays with people |
|---|---|---|
| Requirements analysis and clarification | Summarizes notes, spots gaps and conflicts, drafts questions and acceptance criteria | Designer, Tech Lead and client confirm what is actually needed |
| Architecture exploration | Lays out options and trade-offs, drafts diagrams | Tech Lead decides and records the decision |
| Code generation and modification | Drafts and changes code in small steps | Engineer reads, understands and owns every change |
| Automated test generation | Proposes test cases, data and edge cases | Engineers check that tests fail when the code is wrong |
| Code review assistance | Flags common defects, risky patterns and missing tests | A human reviewer approves the change |
| Documentation | Drafts docs, release notes and runbooks from the code and tickets | Owners check accuracy before publishing |
| Legacy-code analysis and modernization | Explains old code, maps dependencies, suggests refactors and test coverage | Architect decides the approach; tests protect existing behavior |
| Defect investigation | Traces likely causes from logs, code and recent changes | Engineer confirms the cause and the fix |
| Production incident analysis | Correlates alerts and logs, summarizes timelines and drafts post-incident notes | On-call engineers decide and act; consequential changes are approved by a person |
| Technical debt identification | Finds duplication, outdated dependencies and hotspots | Tech Lead and client agree what to fix and when |
Who does what
AI
Generates, analyzes, summarizes and recommends
Never approves, accepts or releases anything.
Engineers
Review, validate, understand and own changes
If an engineer cannot explain it, it does not ship.
Automated controls
Test, scan, validate and enforce defined policies
Failing checks block the merge or the release.
Human decision-makers
Approve architecture, security exceptions, releases and consequential production changes
Named people, recorded decisions.
Confidentiality, access and data handling
Your code and data are yours. These are the rules we follow, and we confirm the details in the engagement's AI-use policy at kickoff.
Client code and data
Agreed at kickoff in the engagement's AI-use policy. Only approved tools touch client code or data.
Access control
AI tools and agents get the minimum access they need, on specific repositories and environments, and nothing more.
Secrets and regulated data
Credentials, secrets and regulated data stay out of AI tools unless the policy explicitly allows it.
Auditability
Changes live in version control with the author, the reviewer and the checks that ran. We record when AI assisted a change where the policy requires it.
Human approval
Architecture, security exceptions, releases and consequential production changes are approved by named people.
Tool terms
Each AI tool has its own data-retention and training terms. We check them before approving a tool, and we tell you which tools we use. We do not promise anything beyond what a tool's terms and our controls support.
Rules we follow
- Only tools approved in the engagement's AI-use policy are used on client code or data
- No secrets, credentials or regulated data are placed in AI tools unless the policy explicitly allows it
- Every AI-assisted change is reviewed by a human engineer who understands it
- AI-generated tests are checked to ensure they fail when the code is wrong
- We tell clients how we use AI on their project, at kickoff and on request
- We review in retrospectives how AI affects quality and speed, and adjust
Responsibilities by role
| Role | Responsibility for AI use |
|---|---|
| Client Sponsor | Agrees which tools may touch their code and data |
| Delivery Manager | Makes sure the policy is followed and visible |
| Tech Lead | Sets where AI-generated code is acceptable and how it is reviewed |
| Engineers | Understand, test and own every change they commit |
| QA | Applies extra scrutiny to AI-assisted changes |
| CTO | Defines approved tools and data-handling rules, and is accountable for the behavior of AI features delivered to the client |
Related
Quality & AI
Delivery with and without AI
How our development approach differs with AI, stage by stage, and what stays exactly the same.
Quality & AI
Quality gates
Four gates a change passes between idea and production, and what each one checks.
Roles
Engineering roles
Solution Architect / Tech Lead and Software Engineers.
Governance
Metrics and reporting
What we measure across delivery, quality, AI and business outcomes, how we report it, and how the numbers lead to decisions.
