Keep the expertise.
Change the model.
The lasting value of AI is the capability your organisation builds around it: its knowledge, its judgment, and its learning. The model is a component. It should not become the owner of the system.
Discuss an AI opportunityPractical AI, with the boundaries thought through.
Start with a useful task.
Document understanding, knowledge retrieval, and workflow assistance are concrete starting points. We identify the decision, the evidence, and the failure cost before choosing the technology.
Keep the model replaceable.
Your knowledge, workflow rules, tools, and evaluation criteria should not disappear when a provider changes. Stable boundaries make model selection a measured decision rather than a rewrite.
Make human judgment count.
Define who sets the goal, what the system may do, and where someone must review the evidence. High-impact actions need explicit authority, not confidence in generated text.
Prove the result on your work.
Representative documents and real tasks matter more than general benchmarks. Evaluate quality, failure modes, latency, and cost before expanding the scope.
Treat access as part of the architecture.
Retrieval and actions must respect the user's permissions. Data processing, retention, provider terms, and audit requirements need explicit decisions, not a generic claim of compliance.
Keep the operating model practical.
Start with metered enterprise AI services where their terms fit the requirements. Measure cost per completed task and avoid committing to expensive infrastructure before the workload justifies it.
Build a learning system. Not a dependency.
Better workflows should make your people more capable. That means preserving their role in framing the problem, challenging the output, and carrying the learning into the next task.