How we work
Six phases, designed to ship. Each phase has specific inputs, specific deliverables, and an explicit decision gate before the next phase begins. Engagements that skip phases are the ones that stall.
Discovery
Understand the problem before proposing the answer.
We start by framing the engagement as a contract between leadership and delivery — not a brief and not a proposal. We map the business problem, the deployment surface, the stakeholders whose sign-off matters, and the constraints that will actually bind the work (regulatory, commercial, organisational). The output is a one-page engagement thesis both sides sign.
- Stakeholder mapping and interview series
- Problem framing and hypothesis set
- Success criteria defined in business terms
- Risk register first pass
Assessment
Measure readiness against the actual use cases, not in the abstract.
An honest assessment answers four questions for each candidate use case: is the data in the right form, are the platforms and MLOps discipline in place, does the workforce have the capability to operate it, and is the governance framework sufficient for the risk tier. Dimensions that score low become pre-conditions of the roadmap — not things the programme will discover the hard way in month nine.
- Data readiness: lineage, freshness, access rights, quality
- Technical readiness: platforms, MLOps, observability, security
- Workforce readiness: literacy, fluency, role coverage
- Governance readiness: policy, risk tiering, owner model
Strategy
Prioritise, sequence, and commit.
We score candidate initiatives on value, feasibility, time-to-value, and risk, and cut the ones that fail any of those tests. We then specify the operating model — how AI work will be funded, staffed, governed, and measured — and sequence a 12–24 month roadmap with named owners, dependencies, and decision gates. The roadmap is written to be approved, not just presented.
- Use-case prioritisation and portfolio design
- Operating model and funding structure
- 12–24 month roadmap with named owners
- Board-ready investment narrative
Build
Engineer the systems as production software.
Build starts with the evaluation harness, not the model. Real tasks, real data, graded rubrics — so promotion decisions are measured rather than guessed. Around the model we build the tool contracts, policy guardrails, observability, and integration discipline that determine whether the system survives contact with production. Our AI Agent & Automation Engineering practice handles this phase directly.
- Evaluation harness and graded task set
- Tool and integration design with rollback discipline
- Guardrails and policy enforcement layer
- Observability, logging, and incident triggers
Deploy
Shadow, supervised, autonomous — with explicit promotion criteria.
Deployment is phased. Shadow mode runs the system alongside the current process with no customer impact. Supervised mode promotes it to active use under human oversight. Autonomous mode is granted only when the evaluation harness, policy layer, and monitoring all pass defined criteria. We train the users and operators during this phase so the system lands into capability, not into confusion.
- Shadow → supervised → autonomous promotion gates
- Runbooks, on-call, and incident playbooks
- User and operator enablement
- Monitoring dashboards signed off with the client team
Govern
Operate the system under continuous governance.
Production is not the end — it is when the clock starts. Models drift, data changes, policies evolve, and regulators ask questions. We design the governance operating cadence that keeps the system accurate, safe, and auditable: drift monitoring, retraining triggers, incident response, and the evidence packs external audit will ask for. Alignment is to the frameworks that actually bind the organisation — EU AI Act, NIST AI RMF, ISO 42001, and NDPA 2023.
- Drift and performance monitoring tied to business KPIs
- Retraining schedule and triggers
- Incident response and post-mortem discipline
- Audit evidence: model cards, DPIAs, policy attestations
Why this shape
AI initiatives fail at predictable points. They skip assessment and discover in month nine that the data cannot support the use case. They treat governance as a launch checklist and fail the first external audit. They promote to autonomous mode without an evaluation harness and lose a customer over a hallucinated answer. The six phases exist so those failures have somewhere to surface — early, cheaply, and with an owner.
We have made every one of these mistakes before, in previous lives. The methodology is what we wish we had had then.