A pragmatic playbook for generative AI in the enterprise
Why 80% of enterprise GenAI pilots stall — and the six moves that separate the leaders from the rest.
Most enterprises do not have a generative AI capability problem. They have a prioritisation problem. Teams run a dozen small pilots, each technically impressive, none tied to a number a finance director would recognise. The result is motion without progress.
The first move is to pick problems where the output is checkable. Drafting, summarising, classifying and retrieving are all tasks where a human can verify quality in seconds. Start there, not with autonomous decision-making.
The second move is to fix your data access before your model choice. Retrieval quality, not model brand, decides whether an assistant is useful. If your documents are scattered, unlabelled or contradictory, no model will rescue the experience.
Third, design the guardrails as part of the product, not as a compliance afterthought. That means scoping what the assistant may see, logging every prompt and response, and giving reviewers a clear path to correct bad output.
Fourth, measure adoption honestly. Licences issued is not adoption. Track weekly active use by role, the tasks people actually bring to the tool, and time saved per task with a before baseline captured in advance.
Fifth, plan for the change, not just the build. The teams that succeed rewrite the standard operating procedure around the new capability and train managers to expect a different workflow. The teams that fail hand over a login and hope.
Finally, keep a kill list. Pilots that cannot show a measurable outcome within one quarter should be closed rather than quietly extended. Discipline about stopping is what frees budget for the two or three ideas that genuinely scale.
