Most AI strategies in regulated businesses fail the same way. Someone picks a model, then looks for a problem, then meets the compliance team. A year later there is a pilot nobody can put into production. This note sets out how we choose where AI is worth using in a bank, an insurer or a healthcare provider, and where it is not.
Start with a decision someone already makes
Every model worth building answers a question a person already answers today, by hand, with a spreadsheet or a rule book. Name the decision, who makes it and how often. If you cannot, there is nothing for the model to improve.
- Who makes the call now, and with what information?
- What changes if the answer arrives a day earlier, or a point more accurate?
- Which systems already hold the signal, and who owns them?
Rule out what a regulator will not accept
Some decisions cannot be automated, whatever the model’s accuracy. A credit decision needs a reason the customer can read. A clinical triage step needs a named clinician who signs it off. Write the constraint down before scoping. It rules out half the candidate use cases in an afternoon and saves a quarter of wasted build.
Three questions that settle it
- Can the decision be explained to the person it affects, in plain words?
- Does a human sign it off, and do they have the time to?
- Is the training data held on terms that allow this use?
The cheapest model is the one you never had to build because a rule did the job.
Elev8 delivery team
Prove it on one workflow before you buy a platform
A platform decision made before the first live use case is a guess. Run one workflow end to end, in production, with real volume: the intake, the model, the sign-off and the audit record. That takes six to ten weeks with the client’s own data. What you learn about data quality and hand-offs is what the platform then has to support, so buy it second.
use_case: claims triage
decision_owner: claims-ops
human_sign_off: required
data_terms: policyholder consent, retained 7y
go_live_target: 8 weeks
