Make AI useful, governable and operable.

The gateway creates one consistent access layer across models sourced from AWS, Azure and GCP. People and applications use a supported interface, while the organisation can apply identity, access, data, cost and operational controls in one place.

It provides a controlled internal environment in which people can experiment and work safely. Keeping internal inference on that governed route gives users freedom to explore while preserving the protections and operational visibility the organisation needs.

I led the programme across the decisions needed to move it from an idea into production: the platform approach, provider integration, security review, access model, team and budget structure, FinOps attribution, onboarding and the work required to make it an ongoing service.

That involved working across Cloud Operations, engineering, security, finance, governance and provider teams. No one of those groups had the complete answer alone. The platform needed their concerns to meet in one design.

Production launch is the beginning of a platform's life.

Models change, providers introduce new capabilities, client tools behave differently and costs move as people find new ways to use the service. A useful AI platform needs normal operational ownership: monitoring, support, incident handling, upgrades, model onboarding and retirement, documentation and clear routes for change.

The runbooks around the gateway are evidence of that work. They cover practical issues such as provider differences, model compatibility, caching, excessive tool context, guardrail stability, onboarding and service recovery. The point is that AI needs the same operational discipline as any other important enterprise service.

I also helped move the gateway from a one-off delivery programme into a continuing service model, with an owned backlog, planned releases, operational work, platform improvements and a way for new requirements to enter the roadmap.

Governance works best when it helps people reach a sound answer quickly.

I contribute to the AI governance operating model as well as the technical controls beneath it. That includes risk-proportionate intake, data classification, model and supplier review, recorded decisions, review points and ongoing assurance after something has been approved.

The governance forum should not treat every request as equally risky or stop at a binary yes or no. Ordinary use should follow a clear supported path. A request involving sensitive information, a new provider or greater autonomy should bring in the relevant expertise and make the additional decisions visible.

This is the same Centre of Excellence principle I use elsewhere: give the requestor access to people who understand architecture, security, operations, cost and the business outcome. Return a route forward, not an unexplained refusal.

People will use the governed route when it helps them succeed.

That means useful model choice, understandable onboarding, support when something does not work and a route for legitimate needs that fall outside the normal pattern. It also means learning from repeated questions and improving the platform instead of answering the same ticket indefinitely.

My developing earned-capability model extends that principle. Everyone should have a useful starting point. Access to more capable models, larger budgets or greater autonomy can then grow as people demonstrate that they can use them effectively and responsibly. The controls create a route to more capability rather than acting only as restrictions.

AI is moving from generating an answer to taking action. Connecting models to enterprise tools through standards such as MCP creates useful possibilities, but it also changes the control problem.

The question is no longer only which model can see which information. It is also which identity an agent acts as, which systems it can reach, what actions it may take, where human approval is required and how the outcome can be audited or reversed.

I am applying the gateway and governance experience to that next stage: governed tool access, proportionate autonomy, clear ownership and controls that follow the action as well as the model request. This work is developing; the objective is to make agentic capability operable before it becomes fragmented across individual tools and teams.