Book a Health Check

Give AI a useful place in delivery

A bounded task, reviewable changes and evidence beside the work make AI assistance easier to assess.

Originally published . Original article by Hugo Mendiguchia; editorial update by OTTRA.

Practical guidance, with sources and trade-offs.

The engineering decision

Evaluate a real workflow from request to reviewed change. Measure the review effort and quality of the result, as well as the time taken to draft it.

On this page

Start with one useful task

AI alignment has two practical tests: is the work serving a useful organisational goal, and is the agent acting within the intent and boundaries people agreed? Keep both questions visible when choosing the task and reviewing the result.

Choose work that has a clear owner and a result someone can inspect: explain a failed pipeline, draft tests for an existing function or propose a small dependency update. Record the permitted repository, tools and actions before you give an agent access.

GitLab’s Duo Agent Platform documentation describes feature-specific chat, development and review capabilities. Availability depends on the feature, offering, version and commercial arrangement. Begin with the capability your team can actually use.

Follow the change through delivery

  1. Intent. An engineer asks for a dependency check without changing the existing build. They define the checks and approval required.
  2. Draft. AI prepares a proposed change and an explanation. An engineer checks the relevant files, assumptions and permissions.
  3. Evidence. The pipeline runs the configured tests and checks; the reviewer examines the findings alongside the diff.
  4. Decision. An authorised person approves, rejects or asks for more work. The change record shows the decision and the evidence used.

This is an illustrative workflow, not a report of a customer result. The useful question is whether your team can inspect and repeat it. The OTTRA Method makes those responsibilities explicit.

Trace the data route

Do not infer an AI data boundary from where the Git repository runs. Identify the selected feature, model provider, context sent, retention/logging settings and access controls. GitLab’s data-usage documentation distinguishes model-provider processing, chat/workflow history and optional logging; policies are not identical for every model and deployment.

Treat model output as a draft that can be incomplete or wrong. A human still needs to inspect the change and the configured checks must support the decision.

Evaluate the result your team receives

  • Review effort: how much investigation or correction was needed?
  • Quality: did the proposed change satisfy the acceptance criteria and existing tests?
  • Rework: what failed later or had to be undone?
  • Operating fit: can the team control access, diagnose failures and see relevant usage/cost?

Compare a few representative tasks before expanding access. A faster draft is useful only when the whole delivery path remains understandable. Check the applicable usage model before scaling an AI workflow.

Build the working practice

Agree a small task catalogue, review ownership, escalation route and place to share examples. Revisit tasks that create more review work than they save. Training should leave the team able to explain what the tool can do and where their judgement is required.

OTTRA helps teams configure these boundaries and put them into practice through GitLab Duo Enablement and Claude Enablement.

Primary sources

Source scope and product terms can change. Checked on 7 November 2026.

Partnerships & assurance

Scroll to see allUse the left and right arrow keys to scroll.