The method

Evidence first. Then a baseline.

Five stages, run in order, on a sprint of two weeks or a program of six months. The last stage feeds the next engagement's second stage, which is why we draw it as a loop rather than a line.

Scope a sprint →

  1. STAGE 1

    Evidence

    What is actually happening — in the data, in the contracts, and in what people do when nobody is presenting. We interview, sample and read before proposing anything.

    A written picture of the current state, with its gaps named.

  2. STAGE 2

    Baseline

    Measure it before we change it: time, cost, error rate, review effort. A claim of improvement is worthless without the number it improved on.

    A baseline measurement both sides agree on.

  3. STAGE 3

    Design

    Decide the approach and where a human has to be. Retrieval, routing, review points, privacy and IP positions — signed off by the people who own the risk.

    A design your architecture and risk leads have approved.

  4. STAGE 4

    Ship

    Build and release in slices, against an evaluation set, with guardrails and a way back. Something real is in use before the engagement ends.

    Working software or a live process, in production.

  5. STAGE 5

    Prove

    Report against the baseline, including the parts that did not improve. Then decide what is worth doing next, on evidence rather than enthusiasm.

    A result measured against the baseline, and an instrument you keep.

Prove feeds the next Baseline.

Delivery rules

Three rules that don't move

  • Nothing ships without a human-review design

    Before anything reaches a user, we write down who checks the output, on what basis, and what happens when it is wrong. If we cannot answer that, it does not ship.

  • You own all prompts, code and documentation

    Everything we write is yours, in your repositories, with no license back to us. You can end the engagement and keep working.

  • Every engagement leaves an instrument behind

    A measure, a test set, a scorecard, a register — something that keeps working after we leave and lets you check the next claim yourself.

How we build AI

"AI programmes fail on data readiness and evaluation — not on modelling. So we gate both before a single agent reaches a user."

For a firm handling privileged material, that discipline isn't optional. We build AI-native legal applications with Claude Code and the Azure AI stack — evaluated on a held-out set you control, human-in-the-loop, and auditable end to end.

  1. PHASE 1

    Use-case & value case

    Score candidates on value, feasibility, data availability and risk. Sponsor approves the top three and how success is measured.

  2. GATE

    PHASE 2

    Data & platform readiness

    Readiness scorecard, lineage, and access controls. Below the agreed threshold, no build starts.

  3. PHASE 3

    Responsible-AI design

    Retrieval strategy, model routing, human-in-the-loop, plus bias, privacy and IP review — signed by your architecture and risk leads.

  4. GATE

    PHASE 4

    Build & evaluate

    Iterative build against an evaluation harness — accuracy, groundedness, latency, cost — and red-teamed for prompt injection and data leakage.

  5. PHASE 5

    Deploy, operate & scale

    Guardrails, observability, cost caps and rollback. Drift monitoring and version control keep it honest in production.

What you receive

  • Ranked use-case backlog with a business case per candidate
  • Data-readiness scorecard and remediation plan
  • The evaluation harness and its results — on a held-out set you control
  • Operating runbook, monitoring dashboard and benefit tracker

Responsible-AI controls

  • No client data to public models without consent
  • Governed LLM access — model routing and allow-lists
  • Prompt and output logging, with training-data exclusion by contract
  • IP in all outputs assigned to you

We report on

  • Evaluation pass rate against a held-out set
  • Cost per transaction
  • Share of outputs requiring human correction
  • Time to first production use case, and adoption
  • Surface AIProcess discovery
  • AI GatewayGoverned LLM access
  • AgamiKnowledge management

Delivery accelerators included at no licence cost.

Start here

Start with one sprint

Two to three weeks, one approver, a deliverable you keep. Tell us the problem and we'll come back within one business day with a scope, a date and a fee.

Scope a sprint →