§ Practice Three registers, one contract Updated May 2026

02 —We define, build,
and operate.

Most firms doing one of these things well call themselves an AI practice. We do all three — under a single delivery contract, with a senior team that owns the outcome, not just the workstream.

§ A · Register one
Define — AI strategy & portfolio

AI strategy & portfolio

The right bets, in the right sequence, with the right hypothesis.

Most AI strategy engagements produce a slide deck. Ours produces a funded roadmap — with sequenced initiatives, staffed teams, and a business case the CFO can sign. We do this by bringing two models simultaneously: a model of the business (where value lives, what friction costs, which workflows can absorb change) and a model of the technology (what AI can actually do in production, in 2026, at your data maturity level). The output of this register is not a recommendation. It is a decision.

A-1
Value-case modelling
Quantified ROI hypotheses for each initiative, with sensitivity ranges, confidence levels, and the assumptions that govern them.
A-2
Capability diagnostic
An honest assessment of your data infrastructure, ML tooling, team readiness, and where the real gaps are — versus where leadership believes they are.
A-3
Vendor & build strategy
Which components to buy, which to build, which to wait on. A model-selection framework grounded in your use case, not vendor relationships.
§ Deliverables
AI opportunity mapValue-case workbookSequenced roadmapCapability gap reportVendor scorecardSteering committee deck
§ B · Register two
Build — Agentic systems & platforms

Agentic systems & platforms

Working systems, shipped. Not prototypes. Not proofs of concept.

We engineer the working system. Model selection, evaluation harnesses, agent frameworks, retrieval architecture, observability, and the data plumbing underneath. Every release is governed by an evaluation harness — a set of tests that represent what the system is supposed to do, run on every push, visible to the business, not just the engineers. We have never shipped a system without one, and we never will. This is what delivery discipline looks like in an AI-era practice.

B-1
Agentic architecture
System design for multi-agent workflows — task decomposition, tool use, memory, orchestration, and failure modes. Built to be debugged, not just deployed.
B-2
Evaluation & safety
Harness-first development. Red-team before release. Every system has a defined set of evals that govern deployment — human-in-the-loop where the risk profile demands it.
B-3
Data & retrieval
The plumbing that makes the model useful. Chunking strategy, embedding selection, retrieval architecture, context window management, and the quality gates that govern all of it.
§ Deliverables
Production-deployed systemEvaluation harnessArchitecture documentationRed-team reportRunbookObservability dashboard
§ C · Register three
Operate — Run & evolve

Run & evolve

We run what we build. The work does not end at launch.

A shipped AI system is a living system. Model providers update APIs. User behaviour diverges from training data. Business rules change. The evaluation harness we built in Register B becomes the operating instrument of Register C — a continuous feedback loop between what the system does and what the business needs it to do. We staff this with the same engineers who built it, which means the people debugging production are the people who understand the architecture.

C-1
Managed operations
On-call coverage, incident response, release management, and the operational wrapper that makes the business confident in a system it did not build.
C-2
Telemetry & quality
A living dashboard of model performance, citation quality, hallucination rates, latency, and cost — reported to business leadership in plain language on a defined cadence.
C-3
Cost & capacity
Inference cost management, capacity planning, and the model-version strategy that keeps the system current without introducing regression. We own the tradeoff.
§ Deliverables
Monthly performance reportQuarterly business reviewContinuous eval resultsCost & capacity reportImprovement roadmap
§ P · How We Work
The engagement model — four phases, one contract
§ The model

An engagement that unfolds in four phases. One contract throughout.

We do not re-pitch between phases. We do not hand off from strategy to delivery teams. The firm that defines the roadmap builds and runs the system — which is why our diagnostics have teeth and our operations have context.

§ Phases
  1. P-01
    The diagnostic
    A small senior team. Two to four weeks. ROI hypotheses, capability gaps, the shortlist of bets worth making. Deliverable: a funded roadmap, not a slide deck.
    Duration2 — 4 wk
    Team4 — 8
    Detail
  2. P-02
    The blueprint
    Architecture, data strategy, evaluation design, vendor decisions. A buildable plan with cost and risk underwritten. The team that writes the blueprint builds the system.
    Duration4 — 6 wk
    Team6 — 14
    Detail
  3. P-03
    The ship
    Integrated delivery. Weekly demos to a business owner, not an IT committee. An evaluation harness governing every release. Monthly steering with the executive sponsor.
    Duration12 — 36 wk
    Team12 — 40
    Detail
  4. P-04
    The steady state
    We run what we shipped. Telemetry, quality gates, cost management, and a quarterly business review that moves the metric — not the slide. Same engineers. Better outcomes.
    Duration12 mo +
    Team6 — 24
    Detail

§ How we work — operating principles

01 / Senior-led, always

No bait-and-switch.

The partner who runs the diagnostic is the partner who runs the program. The senior engineer who architects the system is on the team that operates it. We do not win business with partners and staff it with analysts.

02 / Evaluation-first

Evals before everything.

Every system we build has a defined evaluation harness before it has a production environment. Evals are written with the business, reviewed by the executive sponsor, and run on every deployment. They are the release gate — not an afterthought.

03 / Plain language

One version of the truth.

Business leadership sees the same telemetry engineers see — in plain language, on a defined cadence. We do not manage perception. We report what is happening, explain what it means, and say what we are doing about it.

04 / No stranded work

We run what we build.

The incentive to build a maintainable, well-documented system is very different when you are the team that will operate it. We have never handed off a system to a client team and walked away. We do not believe in that model.

§ R · The record

Selected programs. Real metrics, named carefully.

REC / 01418 months
Global industrials · FTSE 100

Service operations, AI-native rebuild

A 22-month program reorganising two business units around shipped agent systems.

38%Handle time ↓
2.1×First-contact res.
14Production agents
REC / 02212 months
Tier-1 banking · North America

Document intelligence at portfolio scale

Retrieval and evaluation harness for a 1.4M-document workflow, replacing a manual review chain.

9d→11hCycle time
99.2%Audit pass rate
$48mAnnual. value
REC / 0298 months
Life sciences · Top 10 pharma

Clinical operations copilot

Evaluation-first build for trial-protocol authoring and review, with regulator-grade traceability.

4.2×Author throughput
100%Citation trace
3 moTo first release
§ Begin · Start a conversation

The board wants AI results,
not AI announcements.

Begin an engagement Read the field notes