Use AI across your workflow without losing control of it
AI can speed up every stage of delivery, and it can also move a mistake from your editor to production faster. For each stage, this page names where AI helps, the safeguard that keeps you accountable, and what you should not hand over.
Step 1 of 6
Plan
AI summarises feedback and drafts acceptance criteria and edge cases. A person turns the vague ask into a ticket with a checkable right answer and decides priorities.
Automation takes the mechanical checks, and people keep the judgement calls.
Stage 1
Plan
Summarising feedback and tickets, drafting acceptance criteria, listing edge cases, and enumerating design trade-offs for a person to choose between.
Safeguards
- Turn a vague ask into a ticket with a checkable right answer before a coding agent sees it
- Name a human owner for every decision, including those an assistant helped draft
- Check the data classification before choosing which assistant or model to use
- Write down what the change must never do, as input to later red-team tests
Do not
- Let an assistant decide priorities or the definition of done
- Paste customer or confidential data into an unapproved tool to save time
- Treat an AI-generated edge-case list as complete
Stage 2
Build
Gathering context across a large codebase, drafting an implementation and tests from a well-scoped ticket, and following repository conventions.
Safeguards
- Keep tickets small and scoped so the agent has a bounded, checkable target
- Exclude secret files and data exports from the assistant's context, and scan every commit for secrets
- Give coding agents a sandbox and their own limited credentials, not your login
- Check any new dependency the assistant adds before it is installed
Do not
- Let an agent run with production credentials on your machine
- Accept a large diff you have not read because the summary looks tidy
- Connect an MCP server you have not read to an agent holding real access
Stage 3
Review
A first-pass review that flags likely bugs, missing checks and style or consistency issues before a person opens the pull request.
Safeguards
- Resolve the automated review comments first, so human time goes to domain judgement
- Write a pull request description that states intent and what was verified
- Make approval mean the reviewer read the change, and mark AI-assisted changes
- Look for cross-cutting bypasses, such as a second code path that skips a new filter
Do not
- Approve from an AI-generated summary alone
- Let an AI review comment unblock a merge, since only a human approval should
- Dismiss a scanner finding without a logged decision from a named owner
Stage 4
Test
Generating unit tests and edge cases from a signature or specification, and pointing out untested branches.
Safeguards
- Run the tests yourself and in CI, on infrastructure the agent does not control
- Break the code on purpose to confirm generated tests would fail
- Add an evaluation regression gate for any change to prompts, retrieval or model versions
- Add red-team cases that plant instruction-like text where the system reads
Do not
- Trust an agent's report that all tests pass
- Count generated tests that assert nothing meaningful as coverage
- Run the red-team set once at launch and never again
Stage 5
Ship
Cohort-based rollout with automatic rollback triggered by a quality or cost metric, not only an error rate.
Safeguards
- Roll out progressively and watch the specific metric the change could affect
- Keep manual rollback authority with a person, whatever the automated logic says
- Pin model versions and review any version change like a code change
- Confirm the new path goes through the gateway, the policy check and the audit log
Do not
- Float a model version on latest in production
- Skip the canary because the change looks small
- Ship an agent whose kill switch has not been tested
Stage 6
Operate
Summarising traces, correlating a rise in bad answers with a recent change, and drafting an incident timeline for a person to verify.
Safeguards
- Trace requests end to end and alert on cost, denial rate and approval-review rate
- Keep logs immutable and queryable under each agent's own identity
- Treat AI-assisted correlation as a hypothesis, and decide the root cause as a person
- Feed incidents and evaluation failures back into requirements and tests, with a follow-up ticket
Do not
- Let an assistant auto-remediate a production incident without a person
- Turn every one-off failure into a permanent rule
- Wait for the invoice to find a cost problem
Turn this into a checklist for your setup
Pick where AI runs for you and the risk tier, and get the controls that apply.