How AI could cut policy tracking from months to days
For a life-sciences client, our team built a prototype AI assistant that tracks changes in insurers' drug rules, with people in control. It is a working prototype, built and tested on synthetic data only.
At a glance
- Status
- Working prototype, built and tested on synthetic data only*
- Client
- A life-sciences client, a drug maker
- Problem
- Tracking payer drug rules across hundreds of health plans
- Today
- Analysts read every document by hand
- Principle
- The AI never has the final say. A person decides anything uncertain or high-impact.
The problem
In the US, a patient's insurer, called a payer, decides whether it will pay for a medicine and on what conditions. It may require approval first, ask the patient to try a cheaper drug first, or limit the quantity. Payers write these rules in long policy documents and change them often.
A drug maker has to track these rules across hundreds of health plans, because they affect it directly:
Sales: reps tell doctors what is covered. Outdated information leads to rejected prescriptions.
Patient access: support teams help patients get approvals, and they need the current rules to do it.
Rebates: the drug maker pays payers in return for good coverage. If it misses a new restriction, it keeps paying for access it no longer gets.
Today analysts read every document by hand. Confirming a single change takes 3 to 6 months, and two analysts can read the same rule differently. That is the manual baseline.
What we built
We built an AI assistant that does the reading so analysts can focus on decisions. For each policy document, it:
- Reads the document and pulls out the rules that matter.
- Sorts each rule into one standard format, so everyone gets the same answer.
- Compares it with the previous version and highlights exactly what changed.
- Asks a person whenever it is unsure, or when a change has a big impact.
- Alerts the right people, each through a dashboard designed for their role.
What it could change
The prototype works end to end on synthetic test data. These are the targets it was designed to reach, not yet measured results.
| Measure | Today (manual) | Target (with AI) |
|---|---|---|
| Reading one policy document | Hours | Minutes |
| Confirming a policy change | 3 to 6 months | Days |
| Consistency of rule labels | 75 to 85% | Over 90%, with human review |
The figures in the right-hand column are targets, not yet measured results.
People stay in control
The AI never has the final say. Every result carries a confidence score, and anything uncertain or high-impact waits for a person.
The bigger shift is the team's job. Analysts move from finding and reading documents to reviewing what the AI flags and acting on it.
* Working prototype, built and tested on synthetic data only. The figures in the table are targets, not yet measured results.
