Designing an AI-first operating model

We start with observed work rather than the organisation chart. Examples include a claims team re-keying information between systems, a support queue dominated by repeated questions, or a compliance review delayed because one specialist must read every case. Slow hand-offs, duplicated decisions, and information bottlenecks are candidates for process change, whether or not that change uses AI.

Not every bottleneck requires AI. Re-keying between systems is often better solved with an integration; an approval that exists only for audit reasons may need a workflow change, not a model. The assessment should record these as conventional changes and compare them with AI approaches against the same outcome measures.

This stage produces a short, prioritised list of changes. Each connects a technical investment to a measurable outcome, such as hours returned, shorter cycle time, lower error rates, or additional capacity. If a proposed initiative cannot name its measure, it is not ready to be built.

When AI takes over a class of decisions, the surrounding roles change whether anyone plans for it or not, so it is better to plan for it. Consider a support operation where an assistant drafts responses and resolves routine cases. The frontline role shifts from answering everything to reviewing drafts and handling complex cases. The queue needs new routing rules, and quality assurance must assess the combined human and model workflow. None of that is a technology task, and all of it determines whether the technology delivers.

Process ownership has to be explicit. Every AI-assisted process needs a named owner who is accountable for its outcomes, empowered to tune or pause it, and responsible for the human workflow around it. Without an owner, model behaviour and input distributions can change, workarounds can accumulate, and operating performance can decline without a clear response.

Expect the workload to change in composition as well as size. When a system handles routine cases, the remaining cases are often more complex. The people reviewing those cases may therefore need more expertise and judgement. A staffing plan should measure this change instead of assuming that automation produces a proportional reduction in headcount.

An AI-first operating model is defined as much by what it refuses to automate as by what it automates. Decisions with legal weight, obligations of fairness, or serious consequences for individuals may need a human decision-maker as a matter of policy, not just prudence. Calculations that must be exactly right, such as pricing, tax, and eligibility rules, belong in deterministic code, where the model can at most prepare inputs or explain outputs.

Writing these boundaries down protects the organisation from consequential failures. Explicit authority limits also give teams a clear basis for deciding when and how to use the system.

Governance provides standing decisions for recurring questions. These include which data may be used for which purposes, which models and vendors are approved, what evaluation a system must pass before it affects customers, and how incidents are reported and handled. Central decisions can reduce repeated review, although each use case still needs an assessment of its specific risks and obligations.

Measurement belongs in the same frame. Each deployed system should report against the outcome that justified it, and the portfolio should be reviewed on a cadence with the authority to expand, tune, or retire. An AI-first organisation is not the one running the most models; it is the one that can say, for every model it runs, what the model is for and how it is performing.

Sequence matters because operating results provide the evidence for further investment. One approach is to take one or two high-value workflows through implementation, including role changes, ownership, and measurement, before starting a larger portfolio of pilots. A completed workflow provides evidence about costs, benefits, and organisational changes that can inform the next decision.

Throughout the change, personnel need accurate information about what is changing, why it is changing, and how their responsibilities are affected. Their operational knowledge is also necessary for process design and evaluation. The intended result is an organisation that can explain what each AI system does, why it is used, and who remains accountable for its outcomes.