Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
53 changes: 27 additions & 26 deletions examples/agent_demo/scenarios/insurance/PROMPT.md
Original file line number Diff line number Diff line change
@@ -1,39 +1,40 @@
You are an analyst on the customer retention team at Maple Mutual, a
Canadian insurer. In this directory are two raw exports from the policy
system:
You are an analyst on the group retention team at Maple Mutual Group
Benefits, a Canadian insurer that sells group health, dental, life and
disability plans to employers. In this directory are two raw exports
from the plan administration system:

- `renewals_2025_export.csv`: last year's renewals, with
`Status at Renewal`: did the policyholder renew or cancel?
- `renewals_due_export.csv`: policies whose renewal is due in the next
three months. Nobody knows yet who will cancel.
- `renewals_2025_export.csv`: last year's plan renewals, with
`Renewal Status`: did the employer renew its plan with us or leave?
- `renewals_due_export.csv`: plans whose renewal is due in the next
three months. Nobody knows yet who will leave.

You have tabular foundation models (the tabctx tools) for predictions
from tables. Use them, not a model you write yourself.

1. Clean both exports with one Python script, `clean.py` (pandas is
installed), and run it. Write `renewals_2025.csv` and
`renewals_due.csv` with snake_case column names (`Policy ID` becomes
`policy_id`), one row per policy, consistent categories, numbers as
numbers, missing values left empty, and `years_as_customer` at
renewal in place of the two dates. In the history, `cancelled` is
Yes or No. Keep every policy in `renewals_due.csv`. List each problem
you found and what you did about it.
2. Find out how well cancellation can be predicted from the clean data,
`renewals_due.csv` with snake_case column names (`Plan ID` becomes
`plan_id`), one row per plan, consistent categories, numbers as
numbers, missing values left empty, and `years_with_us` at renewal in
place of the two dates. In the history, `terminated` is Yes (the
employer left) or No. Keep every plan in `renewals_due.csv`. List each
problem you found and what you did about it.
2. Find out how well leaving can be predicted from the clean data,
compared with a standard gradient-boosting model, and pick the better
tabctx model.
3. Predict every due policyholder's chance of cancelling.
4. Find which factors drive cancellation.
5. Take the three policyholders most likely to cancel. For each, test
two or three things the retention team could do (for example, an
advisor check-in call, which counts as one advisor contact, moving
them to pre-authorized debit, or a smaller premium increase) and how
much each moves their chance of cancelling.
6. Write `report.md` for the head of retention, in plain language: what
was wrong with the data and how it was fixed, how accurate the model
is, the 10 policyholders most likely to cancel (id and probability),
the main drivers, the what-if results, and one recommended retention
3. Predict every due plan's chance of leaving.
4. Find which factors drive employers to leave.
5. Take the three plans most likely to leave. For each, test two or
three things the retention team could do (for example, an account
review meeting, which counts as one account review, a smaller renewal
increase, or resolving their open service issues) and how much each
moves their chance of leaving.
6. Write `report.md` for the head of group retention, in plain language:
what was wrong with the data and how it was fixed, how accurate the
model is, the 10 plans most likely to leave (id and probability), the
main drivers, the what-if results, and one recommended retention
program with its expected impact. Say plainly where the what-if
results are patterns the model learned, not proven effects.

Keep the predictions CSV the tools write: it will be scored against who
actually cancelled.
actually left.
Loading
Loading