Buyers forward me these decks. The shape is always the same: a six-figure engagement with a consultancy, a deck long enough to feel substantial, and a stock photo of a brain on most pages.
Here is what is never in one.
Not one slide named a model. Not one slide named a tool. Not one slide had an eval criterion. Not one slide named a person on the buyer's team who would own a specific output. Not one slide proposed a measurement.
The deck did contain: a maturity matrix with five tiers, a roadmap with three horizons, a stack diagram with eight boxes labeled in management consulting english, a competitive landscape with logos, and twelve case studies of companies that had successfully transformed with AI. The case studies named no people, no tools, no models, and no numbers.
Here is what an AI strategy looks like.
It names a workflow. The customer support triage workflow. The financial close reconciliation workflow. The content review workflow. Some bounded process where humans currently do bounded work that is mostly pattern matching against documents.
It names the model. Not 'an LLM.' Claude Opus 4.6. Gemini 3 Flash. Specific.
It names the eval. Three hundred examples. Hand-graded. Six axes. A regression suite that runs in CI. A holdout set the team has never seen until acceptance.
It names the integration surface. The model gets called from the case management system at this hook. The output goes into this field. The fallback when the model declines is this human queue.
It names the failure modes. What happens when the model hallucinates. What happens when the model refuses. What happens when the input is adversarial. What happens when the input is in a language the model handles poorly.
It names the cost ceiling. Per call. Per month. With circuit breakers.
It names the human in the loop. By role. With escalation criteria.
If you have those eight things, you have something a senior engineer can build. If you have a deck of brain stock photos, you have bought permission to keep procrastinating.