Roughly 80% of AI projects fail. That’s RAND’s number, not mine. And when BCG traced where AI value actually comes from, the split was stark: about 70% from people and process, only about 10% from the algorithm itself.
Put those two findings side by side and the diagnosis writes itself. We fund the 10% and neglect the 70%. Same tool, opposite outcomes, and the difference is almost never the model. It’s whether anyone managed the change.
That’s the observation. Here is my recommendation: a five-step change-management plan, completed before deployment, not bolted on after the failure review.
1. Name the problem
Not “deploy AI.” A specific problem, owned by a specific team, tied to a number someone is accountable for moving. If you can’t say whose Tuesday this changes, you’re installing technology, not solving anything.
2. Ready the data
The system will run on what you feed it. Inventory the data the workflow depends on, fix what’s broken in it, and decide what the AI may and may not see. Skipping this step doesn’t save time. It relocates the cost to production, where it’s far more expensive.
3. Redesign the workflow
Don’t drop AI into the old process and expect new results. That’s how you hit the Stage 2 Ceiling (opens in a new tab): individuals get faster while the business stands still. Redesign the work around what the machine can do. Decide which steps disappear, which change hands, and which now demand more human judgment, not less.
4. Set graded governance
Before go-live, decide task by task how long the leash is. What is fully delegated? What needs a human in the loop? What is never automated? Sizing control to risk, instead of one flat rule for everything, is the heart of the judgment layer (opens in a new tab), and it is far cheaper to decide now than after the incident.
5. Train, then measure adoption
Training is not a launch email. Teach people the redesigned workflow, then measure whether the work actually changed: cycle time, error rates, decisions per person, not logins. Usage charts tell you people showed up. Only outcome metrics tell you the 70% is being captured.
What to take into the room
One question: for each of the five steps, who owns it, and is it done? If the honest answer is that the budget and the calendar are all pointed at the algorithm, you have inverted BCG’s math: spending on the 10%, hoping about the 70%. My prediction is unremarkable and well-supported: that project joins RAND’s 80%. The organizations that escape it aren’t the ones with the best model. They’re the ones that treated deployment as a change program with software attached, not software with a change program attached.