Make AI part of how your team ships.

One shared way to work with AI coding agents, taught inside your own repository and measured against your own baseline.

  • AI Engineering Method
  • 12-week programme or embedded coaching
  • AI policy under AI Act art. 4

What is the AI Engineering Method?

The AI Engineering Method is the product and engineering track of our AI transformation work. It teaches your team one shared way to work with AI coding agents, inside your own repository and tools, as a 12-week intensive programme or as embedded coaching (about two days a week for around 90 days). We baseline your work-item history, then measure cycle time and the work that passes review and test gates at week 1, mid-way and at the end.

Two kinds of AI the method covers

Generative AI
Drafts and code suggestions that a person reviews before they are used.
Agentic AI
Agents that take steps inside limits we write down, with a person approving what ships.

The method at work in client teams

  • +500%

    Productivity at a healthcare software company, first week of adoption, against the team's own prior weekly average.

    Read the healthcare case
  • 1 week

    Software company: the product designer shipped an app-wide redesign in one week, guided by a frontend engineer, using AI patterns first built with the engineering team.

    Read the software company case

AI pilots and licences that never became the way your team works.

Each engineer found a personal way to use the tools, so the result depends on who picks up the ticket, and AI-written pull requests wait in review because nobody knows how they were produced. Without a baseline, no one can tell the board whether the licences paid off. What changes this is a shared practice, measured before and after.

Where teams usually are when they call us

  • Everyone on the team uses AI tools in a different way.
  • AI-written pull requests wait in review because nobody trusts them.
  • The licences are paid for and the way the team works hasn't changed.
  • AI works on a demo, then gets lost in the legacy code.

Two formats: a 12-week programme or coaching inside your team

Both install the same practice and leave the same assets in your repository. They differ in rhythm, and in how much technical leadership you need alongside.

CriterionIntensive programmeEmbedded coaching
Length12 weeksAbout 90 days
Rhythm3 full-day intensives with the whole team, plus a weekly kitAbout 2 days a week, working alongside your engineers
Where it happensYour repository and your tools, with the intensives as dedicated daysInside your issue tracker, repository and team chat
Best whenThe team can set aside three full days and wants the practice and the assets in place within a quarterDelivery can't pause for full days, or you also need technical leadership
What you end withThe full asset set, adopted and documented, plus the end measurementPatterns built with your engineers and carried to other teams, plus the same measurements
How it's pricedFixed price for the programmeMonthly fee
What can followEmbedded coaching or a fractional CTOYour team keeps running the practice, with the assets it owns

Which one fits you

Pick the intensive programme if you want a fixed start and end, a budget agreed up front and a team ready to block three days.

Pick embedded coaching if delivery can't pause, or if AI adoption is one of several technical problems.

What your team keeps after the programme

Everything lives in your repository and your accounts. Each agent, instruction file and template moves from draft to pilot to adopted, and what the team doesn't use is dropped.

  • An AI agent harness in your repository

    Project instructions and role-based coding agents (a specialist per product area, a database reviewer, a knowledge-base curator), with context engineering: each agent loads only the instructions and codebase knowledge its task needs.

  • A product and technology knowledge base

    What the software does, kept apart from how it is built. Agents read it, so domain facts aren't pasted into every prompt.

  • A knowledge graph of the codebase

    Built deterministically: no source code is sent to an LLM to build it. Custom extractors cover the parts standard tools miss.

  • Test gates on AI-written changes

    Tests, linting and security scans that every AI-written change passes before a person reviews it.

  • An analyst agent

    Turns a change request into a spec pack: functional and technical spec, expected behaviour, implementation plan and task order.

  • An AI policy under AI Act art. 4

    Which tools, on which data, with which approvals, and how the team keeps its AI literacy current.

  • An asset catalog and decision log

    Every agent, instruction file and template tracked from draft → pilot → adopted, with a progress dashboard generated from the files.

  • The weekly kit

    Templates for the weekly session, so the practice keeps running after the handover.

How do we know it's working?

We count how work moves through your team, using data you already have. Lines of code and prompt counts say nothing about delivery, so we leave them out.

  1. Baseline

    Pulled from your team's work-item history before the programme starts.

  2. Week 1

    First comparison against the baseline, on the work done with the harness in place.

  3. Mid-programme

    The same measures, plus which assets are in pilot and which are adopted.

  4. End

    The final comparison, written up for your management with the baseline next to it.

What we measure: cycle time, work items completed through review and test gates, and change failure rate.

Guardrails that keep delivery safe while the team learns

These rules go into the AI policy in the first weeks and apply to every agent and every person on the team.

  • Agents get narrow access

    Each agent reads and changes only what its job needs; anything else needs approval. Unlimited agent access is a named risk in the OWASP Top 10 for LLM Applications (LLM06, Excessive Agency).

  • A person approves every merge

    Agents propose changes. An engineer on your team reviews and merges them, after the test gates have run.

  • No sensitive data in any AI tool

    Agreed in writing before work starts. In healthcare, that means no patient data in any tool, prompt or asset.

  • A shadow track before agents take tasks

    On a real change request, agents produce a spec pack beside your team's usual approach. Your tech lead compares the two and decides go or no-go.

  • Delivery comes first

    The pilot runs beside your release plan. If a client deadline is at stake, the pilot waits.

Week by week, in the intensive programme

A typical plan: the exact weeks are fixed with your tech lead at kick-off. Embedded coaching follows the same steps over about 90 days.

  1. Weeks 1-2: baseline, rules and harness

    We pull the baseline from your work-item history, write access limits and data rules into the AI policy, install the harness in your repository and run the first full-day intensive. First measurement at the end of week 1.

  2. Weeks 3-5: knowledge base, graph and shadow track

    We build the knowledge base and the codebase graph. On a real change request, agents produce a spec pack beside your team's own approach, and your tech lead decides go or no-go.

  3. Weeks 6-9: agents take real tasks

    After a go, agents pick up tasks behind the test gates and a person approves every merge. Second intensive, agents and instruction files move from draft to pilot, mid-programme measurement.

  4. Weeks 10-12: adoption and handover

    Third intensive. Assets the team uses are marked adopted, the rest are dropped. End measurement, final AI policy and handover with the weekly kit.

Built for the codebase you have, legacy included

We run it on long-lived enterprise codebases, healthcare among them. Greenfield teams get the same method with less archaeology.

A good fit if

  • Your team pays for AI coding tools and delivery looks the same
  • AI-written pull requests pile up in review
  • You run a long-lived codebase and want AI to work on it safely
  • You work in a regulated sector and need written data rules and an AI policy
  • You want the practice to stay after we leave, in your repository and owned by your team

Not a fit if

  • You want a licence rollout or a one-off workshop
  • The goal is to replace engineers with AI
  • Nobody on your side can own the practice once the programme ends
  • The team has no room in the next quarter for intensives or weekly sessions

Questions about the AI Engineering Method

The intensive programme has a fixed price for the 12 weeks, agreed before it starts. Embedded coaching is billed monthly. The How we engage page explains how we scope and price our work.

The programme is planned around your releases. The intensive format takes three full days out of twelve weeks; the rest happens inside normal work. Agents take tasks only after a shadow track on a real change request and a go from your tech lead, and if a client deadline is at stake, the pilot waits.

The ones your team already has or chooses, for example Claude Code, Cursor or Copilot. The harness, the agent instructions and the knowledge base are files in your repository, so they stay with you if you change tool. The AI policy records which tools are allowed on which data.

Yes. We run it on long-lived enterprise codebases. The knowledge graph is built deterministically, with custom extractors for the parts standard tools miss, and no source code goes to an LLM to build it.

We see it early. The week-1 and mid-programme measures, together with the asset catalog, show which assets the team actually uses. Unused ones are reworked or dropped, and the owner on your side decides what stays. If the team needs more hands-on help, embedded coaching can follow.

You do. Everything lives in your repository and your accounts, documented, so your team can change it without us.

This is our reading, not legal advice. AI Act art. 4 has applied since 2 February 2025; as amended in July 2026 by Regulation (EU) 2026/1744, it asks providers and deployers of AI systems to take measures to support the AI literacy of their staff. The written AI policy and the programme itself are measures of that kind. Check your specific obligations with your counsel.

The person who approved the merge, as with any other change. Agents propose, an engineer on your team reviews and merges after the test gates, and each agent has written limits on what it can read and change.

See it in practice

Book a 30-minute call to set your baseline.

We look at your work-item history together and tell you which format fits, or that neither does yet.

Let's talk