Inference Public

Applied AI · Offer

Can AI improve how your engineering team delivers software?

An AI-assisted engineering assessment for one team and one public service, grounded in the evidence your delivery pipeline already produces.

AI coding assistants and agents can help engineers explore unfamiliar systems, generate code, prepare tests and work through routine tasks, but bringing them into an established delivery team also changes what people must review, which information the tools can reach and how the organisation shows that a change was safe to release. The question for a public-sector engineering leader is whether a new way of working will improve delivery for their own team, while keeping the controls and professional judgement on which their service depends.

Inference Public offers a four-week, bounded assessment for one engineering team working on one service. We examine how software currently moves from requirement to release, the evidence behind each release, where AI assistance might add value and which restrictions would need to be in place before a trial. The engagement ends with a recorded decision: proceed to a controlled trial, fix weaknesses in the delivery foundation first, keep the current approach or stop.

Discuss an engineering assessment

Who this is for

The assessment is intended for heads of engineering, digital and technology leaders, service owners and platform teams in central government departments, local authorities, agencies and other public bodies that are considering AI-assisted development or want to bring existing tool use under stronger engineering control.

It is particularly useful where engineers already use coding assistants informally, where leaders have been asked to show productivity benefits, where delivery depends on pipelines built by a supplier, or where the team must decide whether agentic development tools are justified. It also helps when tool subscriptions or vendor support are changing, because the assessment establishes which tools the team actually relies on before any change is made.

What we examine

We work with the team, its service owner and its security lead, using agreed read-only access to repository, pipeline and delivery records wherever that is practical. The assessment covers the development workflow, code review, automated and manual testing, approvals, deployment and recovery, and how source code, requirements, credentials and operational information would be handled if AI tools were used.

For delivery performance we use the five software delivery metrics published by DORA: change lead time, deployment frequency, failed deployment recovery time, change fail rate and deployment rework rate. Because AI assistance shifts effort from writing code towards checking it, we add two measures DORA does not cover: the effort spent reviewing changes and the effort spent verifying them. Missing, inconsistent or inaccessible data is reported as an evidence gap and never converted into estimated figures, and no measure is chosen in a way that rewards faster releases at the expense of reliability or service outcomes.

We take account of the standards your organisation works to. For central government these usually include the Technology Code of Practice and the Service Standard, and some departments set their own rules: the Home Office engineering standards, for example, require AI-assisted code to meet the same security expectations as human-written code. Local authorities and agencies apply these frameworks in different ways, so we agree at the start which standards bind your team and which serve as good practice.

What you receive

The work produces six named deliverables, each written for review by the people responsible for the service.

1. Delivery baseline

A measured account of the current delivery process, drawn from an agreed period of the team's own repository, pipeline and incident records, with the definition and limitations of each measure stated. Where reliable historical data does not exist, we record the gap and set out the instrumentation needed to establish a baseline from now on.

2. Pipeline evidence review

A practical assessment of what each release proves, covering traceability from change to approval, test results, security checks, the deployment record and the route to recovery. It shows where the pipeline already produces dependable evidence and where assurance still relies on manual reconstruction or individual knowledge.

3. AI use boundaries

A proposed set of rules for what AI tools and agents may read, draft, change or run, with particular attention to sensitive information, credentials, third-party code, untrusted content in repositories, tool permissions and human approval. The rules follow the principles in the NCSC's published advice on agentic AI, including least privilege, limited scope and short-lived credentials, and they set out how AI-generated code and tests are checked independently and which decisions stay with authorised engineers. The proposed rules are presented to your security lead for review. Delivering the assessment does not depend on that review being completed, but no trial should proceed without the approvals your organisation requires, and approving and applying the rules remains your organisation's decision.

4. Tool and migration options

An assessment of the team's existing tools and realistic alternatives, considering developer experience, identity and data handling, fit with the current platform, the ongoing ownership burden and the cost of changing direction. Where relevant, it covers AWS's announced end of support for Amazon Q Developer IDE plugins and paid subscriptions on 30 April 2027, and what a move to Kiro or another approved tool would involve, without treating an AWS product change as a reason to migrate. It also notes any government-provided access to tools that the team may be eligible for, which your organisation would need to confirm at the time.

5. Measured trial plan

A bounded proposal for evaluating one or two candidate practices within the same team, using comparable work and explicit measures of quality, review effort, reliability, security and developer experience. The plan names the permissions, responsibilities, observation period, stop conditions and the evidence needed before adoption is extended. The trial itself is not part of this assessment.

6. Decision record

A document that brings together the findings, options, likely costs and remaining uncertainties, with a recommended next step from the four set out above. It names the accountable decision-maker in your organisation and leaves space for their later decision to accept, amend or decline the recommendation, so that open questions stay visible. That decision is yours to take in your own time and is separate from completing the assessment.

What we need from you

  • One team and one service, agreed at the start.
  • Read-only access to repository metadata, pipeline runs, deployment records and incident records for an agreed period, or exports of them.
  • Regular time from the team's technical lead, two working sessions with your security lead, and a decision session with the service owner or accountable lead.
  • Short interviews with members of the team.

No changes are made to your systems, and nothing is deployed.

How the four weeks work

StageFocusOutput
Week 1: establishConfirm scope, secure access to evidence, interview the engineering, service and security leads, and book the week 3 security session and the week 4 decision sessionAgreed scope, evidence inventory and baseline definitions
Week 2: examineAnalyse pipeline and repository data, review assurance evidence and identify candidate uses of AIDelivery baseline and pipeline evidence review
Week 3: designCompare tool and migration options, develop the AI use boundaries with the security lead and calibrate measures for a trialDraft AI use boundaries, options comparison and trial plan
Week 4: decideReview findings with you, incorporate security comments and present the recommendationAll six deliverables and the decision record

Four weeks is the planned duration once scope, access and stakeholder time are agreed. If evidence or people cannot be made available as planned, we agree a revised plan with you rather than present incomplete work as finished.

What stays with your organisation

Your organisation keeps responsibility for its service, engineering standards, security policy, tool and licence approvals, release authority and any decision about whether an AI-assisted change may proceed. We provide the assessment, the analysis and a recommended trial method. Any trial, software implementation, tool procurement, ongoing operation or production change would be scoped and agreed separately.

Inference Public is an AWS-focused engineering and delivery organisation, so we can assess AWS tools in depth alongside the other approaches available to your team. The assessment does not assume an AWS purchase, and where fixing existing platform or assurance weaknesses would do more good than a new tool, that is what we will recommend.

What completion looks like

The assessment is complete when all six deliverables have been provided, any missing or unreliable baseline data is stated explicitly, the proposed AI use boundaries have been presented to your security lead for review, the trial plan includes measurable success and stop conditions, and the decision record has been presented to your accountable lead. Completion does not depend on your security review or your decision being finished. Your organisation may later accept the recommendation, ask for changes or decline it, and completing the assessment does not authorise a trial or any change.

Common questions

Does the assessment include implementing an AI coding assistant?

No. It establishes the evidence and controls needed to decide whether to begin a controlled trial. Configuration, licensing, the trial itself and any change to production can be scoped separately if you decide to continue.

Do we need a complete set of DORA metrics already?

No. We start with the evidence the team already has, check whether it is reliable enough and state clearly where measurement is missing. When existing data cannot support a credible comparison, the baseline deliverable sets out how to start measuring.

Will you review source code or production data?

We agree the minimum access the assessment needs, preferring metadata, configuration and read-only delivery records. Any access to confidential repositories or sensitive material follows your security, contractual and information governance requirements, and you do not need to send us source code or credentials to make an initial enquiry.

Does this prove that AI-assisted development is safe or more productive?

No. The assessment identifies the current evidence, the risks and the tests that are feasible, then specifies how a bounded trial could evaluate the change. It does not certify a tool or guarantee improvements in productivity, quality or security.

Is the assessment useful if we have already adopted AI coding tools?

Yes. Existing use can be assessed against current delivery evidence, approved tool use, permissions, verification practice and any upcoming product changes, and the outcome may be better controls or measurement without replacing the tools you already have.

Is this the starting point for any AI project?

It is designed for software engineering teams adopting AI across the delivery lifecycle. If you are considering an assistant that answers from your own information, the AI-ready data assessment is the better starting point, and for document extraction, casework or services for the public, talk to us about your situation.

Take the next step

Tell us which service your team maintains, what kind of AI-assisted engineering you are considering and where you are uncertain about quality, permissions or outcomes. We can then tell you whether this assessment is the right first engagement, and you do not need to share confidential project information through the contact form to do so.

Discuss an engineering assessment

Further reading

Scope, access and commercial terms are agreed in writing before work begins. The assessment is not a security certification, accreditation or legal opinion.