Engagements run from a few weeks of diagnostic to eighteen months of transition leadership, in Paris, London, Brussels, Amsterdam and Casablanca, and remotely for Dubai and Riyadh. Every engagement ends with a hand-over package the team can run without us.
You have more AI initiatives than decisions. Vendor-embedded AI, experiments and fundable projects sit in the same list; value, regulatory exposure, social impact and environmental cost are assessed by different people, if at all. The work is to install a method that produces defensible recommendations — and to publish the evidence, including the uncomfortable numbers, before the headline metrics.
A five-axis diagnostic without a composite score; a Machine Learning Canvas at exploration; six-dimension dossiers with three-colour verdicts and non-blocking reservations; a value × feasibility × factorisation grid for programme design.
A leading French mutual insurer: about 70 tracked items, 21 AI-heavy projects in a €9.3M 2026 budget, an acceleration programme and an AI Manifesto validated by the board. A general scoring framework validated for all predictive models, with run governance owned by the business. An Agentic Data Office framework and a six-week proof of concept with every criterion passed.
An architecture framework, a qualification method the teams run without you, a programme with explicit risk and ROI claims, and governance that decides rather than defers.
The Data & AI office has become the place where everything escalates: sources, flows, incidents, business-unit roadmaps, AI questions. The mandate is to narrow it to what it should own, give the rest a named owner, and fund it on commitments rather than recharged problem-solving.
Thirty days of evidence — systems, budget, platform, incidents, interviews — then a decision-rights matrix, a second-line office for standards, monitoring and challenge, three accountable heads, spans of five to eight, five monthly operating gates.
A global electronics manufacturing group (2026): diagnostic, decision-rights matrix and target structure delivered as the hand-over package; 2027 budget built on service- and consumption-based recharges and a no-work-without-funding rule. A global food & beverage group (2022–23): Enterprise Data Management operating model designed, proved on carbon accounting, and transferred to the permanent lead.
A diagnostic the organisation recognises itself in, decision rights with names and dates, a budget that funds ownership, and a successor in place.
A platform has been rolled out and the field has stopped trusting it; a programme has consumed its budget and the legacy it was meant to retire is still running; three tools need to become one and no entity wants to move. These are decision problems before they are delivery problems.
A taxonomy of what fails and why, regional tables with owners and dated targets, relaunch from the strongest region on evidence, adoption tied to managers' own indicators, stabilisation and evolution funded separately, technical debt protected in the plan, end-game and decommissioning run as separate costed projects.
A national water utility: subscription pass rate 39% → 86%, adoption 42% → 66%, support backlog 12 weeks → under 2; an IoT platform taken to 100% of 18,200 terminal units and transferred to operations. A Global CRM consolidation scoped to a Go in principle with five entity heads. A six-party telematics insurance launch delivered on the motor-show date.
A platform people use, a run organisation that detects before the field does, and a closing plan that survives a budget cut.
Some problems need the model to be built, not described. The prototype is developed personally — data reconciled, hypotheses tested, operating point set on the business's own error costs — then industrialised with the engineers who will run it.
Explainable alerts landing in the CRM against a named owner; thresholds chosen on a confusion matrix with asymmetric costs; deterministic training, byte-tested artefacts and parity certification as merge-blocking gates; an agentic engineering studio with ADRs, automated re-review and human-only merges.
B2B churn across six countries and 30+ sources; 20,000 industrial sensors with anomaly detection and MLOps; a customer-satisfaction data platform with NLP across a dozen languages; a signal-to-execution trading platform — 400,000+ lines, 1,000+ pull requests in twelve months — that withheld capital when the economics said no.
A working prototype, the economic case for its operating point, the pipeline and monitoring to run it, and the discipline to stop when the numbers say stop.