Takeaways
What to remember, in ten points, for a leadership team deciding what to do about generative AI.
What to remember
- Generative AI is the latest threshold in eighty years of work, with two collapses along the way. Before accepting a claim, ask whether the task is discriminative (classifying what exists) or generative (producing something new).
- "AI" is a family of techniques. A small model running on infrastructure you control is often the right first choice rather than a fallback.
- The question has moved from whether to invest in AI to how to make the spending pay. Set the conditions before you spend.
- Read vendor messaging as positioning. The signal about the technology comes from your own pilots, measured against a baseline.
- Stress-test every business case on volume and pricing model, not only on price, and include a scenario where AI spending slows.
- The same technology produced success, failure and damage. How the problem was chosen, who led, how outcomes were measured and whether failures were stopped explain most of the difference, in my reading of the cases. Treat reported gains as claims until they are measured.
- Every use case needs a business case that would stand without AI. Governance then rests on six pillars, each with an owner and a minimum form from the start. A sanctioned tool good enough to be preferred is the first control against shadow AI.
- Results are easiest to measure in retrieving and extracting. Acting comes last, and Fly only where actions are bounded and reversible.
- Choose the first project for how quickly it will show whether it works. The most useful test is that checking its output takes much less time than producing it.
- Start this week without a budget: an owner, visibility of current usage, an interim policy and a sanctioned tool. Then run two tracks, safe use for everyone and one first project, and decide on the evidence at ninety days.
In one line
Generative AI did not change what good execution looks like. The organisations that benefit will apply the ordinary discipline, without shortcuts, to an extraordinary tool.