In short
- Pilots stall on ownership, not on models.
- Pick 3 use cases with a named P&L owner each. Stop the rest.
- Fix data in the order the use cases need, not alphabetically.
- Measure adoption weekly; value monthly; publish both.
The decision problem
Walk into most organisations and ask who can approve an AI use case for production. You’ll get 3 answers and a meeting invite. That isn’t a technology gap – it’s an operating-model gap, and no amount of model evaluation closes it.
The organisations that scale have done something unfashionable: they’ve written down who decides, what evidence they need, and how long they get. Usually one page. Usually boring. Always the difference.
“Pick 3. Name an owner for each. Then say no to everything else for a quarter.”
3 fixes, in order
1. Run an honest inventory. Every pilot, who owns it, what it costs, what it has changed. Expect the list to be longer than anyone admits – yes, including the spreadsheet nobody will own.
2. Fund by value, not by enthusiasm. Rank on 2 axes: value if it works, and distance to production. Fund the top 3. Park the rest with a date to revisit.
3. Fix data in use-case order. Data readiness programmes that start with ‘everything’ finish with nothing. Start with the fields the first use case needs and widen from there.
AI office
Sets the rules: approval, risk, value tracking.
AI enablement
Builds the capability: data, platforms, people.
What this looks like in practice
A health-tech scale-up we’re working with asked for a data governance diagnostic, not an audit – one that told the board where the exposure was and what to fix first, in time for a funding round. That’s the pattern: decide what matters, then build only what that needs. Read the case study →
