Making AI work for the business, not the other way round

The approach

Adoption is led by a dedicated adoption team and runs as a staged programme. Each stage builds the confidence and skills needed for the next, and each is designed to be repeatable across business units, so success in one area can be rolled out to the next.

1. Stage 1: Build comfort across the organisation

   It starts with an introduction to AI for everyone, delivered at scale through live sessions, recorded content and leadership briefings: what AI can and cannot do, how to use it well and how to use it safely. Senior leaders go first, so the message is seen to come from the top. The aim is to remove uncertainty so people feel comfortable trying it.

2. Stage 2: Establish distributed AI governance

   In a large organisation, one central council cannot manage AI rollout everywhere, so governance works at two levels:

   • A central AI council brings together technology, security, legal, data protection and HR. It sets the AI policy, the approved tools register, data classification rules and the guardrails every function works within.
   • Function-level AI leads take delegated responsibility for rollout in their own area. Each owns local adoption, prioritises its function's use cases, manages local risk within the central guardrails and reports progress back to the council.

   This keeps standards consistent while putting decisions close to the work, where the people who understand it can act quickly. Governance is set up to enable, not block: clear rules give people permission to experiment and move scattered personal use onto approved enterprise tools.

3. Stage 3: Build a champions network

   The adoption team recruits and trains AI champions in each function, working alongside the function's AI lead. Champions run small-group lunch and learn sessions built around the work their colleagues actually do, with each group picking its own use cases. This spreads adoption far further than a central team could alone, and keeps it rooted in real tasks rather than generic demos.

4. Stage 4: Grow capability step by step

   The programme then grows in ambition as confidence builds:

   1. Chat tools for everyday tasks such as drafting, summarising and research.
   2. Workflow productivity, connecting AI into the way tasks flow from one step to the next.
   3. Building their own tools, creating simple assistants and automations for work that repeats, within agreed guardrails.
   4. Working with coding agents, first to build personal productivity tools and then tools that help whole teams, with a clear route for the best ones to be reviewed and supported centrally.

   By the final step, people are no longer just users of AI. They are shaping how their own work gets done.

The approach

• Run a use-case portfolio, not a list of pilots. Ideas come in from the champions network, the AI council and process owners, and are managed as a single portfolio with clear stage gates from idea to production.
• Start with a proof of concept, then productionise. The best ideas often come from someone close to the work who builds a working proof of concept. Once it proves a better way of working, the full process is mapped, and the solution is rebuilt to the organisation's engineering standards, connected securely to approved tools and hosted securely.
• Prioritise ruthlessly. Candidates are weighed on business value, feasibility, risk and readiness of the data and systems involved. Some processes should be automated, some should be assisted, and some should be left alone. Knowing the difference matters.
• Keep humans where judgement matters. Good AI workflows have clear points where a person reviews, approves or corrects, so quality, accountability and compliance are built in rather than bolted on.
• Prove one step, then scale. Delivery is agile: build one fully working step, put it in front of real users, learn, and extend. A working pipeline teaches more than a long plan.
• Bring security and architecture in early. Data access, integration with enterprise systems and risk controls are agreed up front, so promising work is not stopped at the last gate.
• Manage the people side of the change. Roles, responsibilities and ways of working change when AI takes on part of a process. Affected teams are involved from the start, trained on the new process and supported through the transition.

Measuring success

• Baseline before building. The current process is measured first: time, cost, error rates, volumes and customer outcomes. Without a baseline, there is no way to prove the change worked.
• Agree the measures up front. Each workflow has clear success measures tied to the business case, agreed with the process owner before delivery starts.
• Make performance visible. Dashboards track how each AI workflow is performing against its measures, including quality, human review rates, exceptions and benefits realised. They give process owners and leadership a shared, live view of value.
• Track benefits over time. Benefits are tracked and reported through programme governance, so the portfolio keeps investing in what delivers and stops what does not.

Running and refining the service

• Refine as the models change. AI models and tools improve and change quickly. Each change is treated as a release: tested against the agreed measures before it reaches users, so a model update improves the process rather than quietly breaking it.
• Tight change control. Changes to prompts, models, rules and integrations go through formal change control, with clear versioning, approval and rollback.
• Service management wrapped around. Every AI workflow is transitioned into business as usual as a managed service, with clear ownership, incident and problem management, monitoring and support, so the lights stay on.
• Continuous improvement. Insight from dashboards, incidents and user feedback feeds a regular cycle of refinement, keeping the process improving long after go-live.

The approach

• Find the service worth productising. The starting point is the work customers already pay for that follows a repeatable pattern, relies on expertise that can be written down, and is limited by how many people are available to do it.
• Capture the expertise. The knowledge that lives in experienced people's heads, the rules, standards, preferences and exceptions, is captured and structured so the product can apply it consistently.
• Design the product around the customer. The product is shaped by the outcome the customer wants, not by the old internal process. Each stage is redesigned for what AI does well, rather than automating the way the work used to be done.
• Build human review into the product. Human-in-the-loop checkpoints are a designed feature, not a safety net. They protect quality, build customer trust and create feedback that improves the product over time.
• Capture what people do. Every review, correction and judgement is captured as part of the product's knowledge. As the process is fully measured, people shift from producing the output to overseeing it, focused where their judgement adds most.
• Build it as a pipeline of agents. The product is broken into clear stages, each handled by a focused agent or automated step, with defined inputs, outputs and hand-offs. This makes it easier to test, improve and replace individual parts as AI capability moves on.
• Ship early, learn fast. Delivery is agile and product-led: a minimum viable product goes to real users quickly, and each release is shaped by how it is actually used.
• Plan for the commercial and operational change. Pricing, customer onboarding, support, roles and team structure all change when a service becomes a product. These are planned alongside the technology, not after it.