Inference Public

Applied AI · Software delivery

AI in how public services build, test and release software

How public-sector engineering teams can adopt AI coding assistants and agents with a measured baseline, clear boundaries and evidence of what actually improved.

An engineer working at her desk in a bright office

Many public-sector engineers already use AI coding assistants, sometimes with approval and sometimes informally, and leaders are under pressure to show that the tools save time. The harder part is knowing whether delivery has actually improved, because AI shifts effort from writing code to reviewing and verifying it, and faster releases do not establish improvement if review quality, reliability or service outcomes deteriorate.

Ask about an engineering assessment

What better looks like, and how you would measure it

A credible measure starts from the team's current delivery record. DORA's five software delivery metrics are a widely used starting point: - change lead time - deployment frequency - failed deployment recovery time - change fail rate - deployment rework rate.

Because AI assistance changes where effort goes, we add the time spent reviewing changes and the time spent verifying them. Where a team cannot measure these today, putting the measurement in place comes before any claim of improvement.

Start with the platform you already have

Government guidance on AI coding assistants makes the point that the risk depends heavily on the platform underneath: where a production service is developed, maintained and deployed from a single environment, it warns that AI coding assistants may introduce unacceptable risks, whereas a team with separate environments, automated tests and a pipeline that records what was released and why has much less cause for concern. Strengthening that foundation is often the most valuable step, with or without AI. Our articles on treating the build pipeline as a production system and pipelines that produce their own evidence explain why.

The options, in plain terms

  • An assistant in the editor, which suggests code and tests while engineers remain in control of every change.
  • Agents in the delivery workflow, which can make multi-step changes, open pull requests or run tasks, and therefore need tighter permissions and review.
  • A specification-led way of working, where requirements and designs are written first and AI works from them, as in AWS's published AI-Driven Development Lifecycle method and the Kiro development environment.
  • Strengthening the platform first, where testing, environments or pipeline evidence are not yet strong enough for AI-generated change.
  • No change, where the evidence does not show a benefit worth the risk and cost.

Who does what

What AWS providesWhat we doWhat you retain
Kiro, AWS's agentic development environment; Amazon Bedrock for custom agents; pipeline and developer services. AWS has announced that Amazon Q Developer IDE plugins and paid subscriptions reach end of support on 30 April 2027, with Kiro as the suggested successor; Q Developer in the AWS Console is not affectedEstablish your delivery baseline; review the evidence your pipeline produces; propose boundaries for what AI may read, change or run; compare tools and migration options; design a measured trial; implement and integrate when separately agreedEngineering standards, security policy and sign-off, tool and licence choices, release authority, and the decision to trial, adopt or stop

Controls that matter

The NCSC advises that agents should have only the access they need for as long as they need it, with short-lived credentials and named people who can stop them. Some departments set their own rules: the Home Office engineering standard, for example, requires AI-assisted code to meet the same security expectations as human-written code. AI-generated tests deserve particular care, because a test written by the same tool that wrote the code can pass while checking the wrong behaviour, as our article on verifying AI-generated code and tests explains.

What government has found so far

The Government Digital Service ran a cross-government trial of AI coding assistants from November 2024 to February 2025. Of 2,500 licences made available, around 1,900 were assigned and about 1,270 activated. Participants estimated that they saved 56 minutes a working day on average, a self-reported figure that the report itself qualifies, rather than a measured delivery outcome. GDS's AI Engineering Lab has since offered public-sector teams limited, time-bound access to coding tools with support, so eligibility is worth checking before buying licences. Neither programme is Inference Public's work.

When we would recommend something else

We would advise against adopting new AI tooling where the delivery platform needs strengthening first, where government-provided access already covers the need, where the team's existing tools are adequate and better controls would close the gaps, or where the evidence does not show a benefit. Our assessment can end in any of those conclusions.

For central government

Departments working to the Technology Code of Practice and the Service Standard will want technology choices, costs and risks to be traceable to a decision. Some departments, such as the Home Office, publish their own standards for AI-assisted engineering, and the assessment works to whichever apply to your team.

For councils

Councils often rely on suppliers for parts of their software estate, so the first question is who writes and releases the code, and under what contract. Where suppliers use AI tools on your behalf, the same boundaries and evidence should apply to their work as to your own team's.

For agencies and regulators

Agencies with small engineering teams may get more from strengthening review and pipeline evidence than from new tools, and those operating regulated or sensitive services should agree the boundaries for AI-generated change with their security lead before any trial.

A sensible first step

Our AI-assisted engineering assessment takes four weeks for one team and one service. It produces six deliverables: a delivery baseline, a review of pipeline evidence, proposed AI use boundaries, tool and migration options, a measured trial plan and a decision record. The trial and any implementation are scoped separately.

Ask about an engineering assessment

Further reading