# Concept build: Calder & Rowe: AI-drafted replies with a human sign-off, for an insurance broker

> A concept build: a fictional brand we designed to show what this looks like. It is not a client engagement and the numbers are modelled, not measured.

A concept where routine client emails are drafted by AI from the broker’s own policy documents, and every reply is approved by a person before it is sent.

## Challenge
Account handlers spend much of the day answering the same questions about cover and renewals, retyping details found in policy documents.

## Approach
- **Draft from the documents.** The assistant reads the policy and proposes a reply, citing the clause it relied on.
- **Human approval always.** Nothing goes to a client until a handler has reviewed and approved it.
- **Corrections become training.** Every edit a handler makes improves the next draft.
- **Audit trail.** Who drafted, who approved and what was sent are recorded.

## Built
- Policy-grounded reply drafting
- Human review screen
- Citation of source clauses
- Edit-learning loop
- Audit log

Stack: LLM, Document index, Review UI, Email integration.

## Worked example: time returned to handlers (worked example)
Assumptions:
- Routine replies per day: 30
- Minutes saved per reply by drafting: 5
- Working days per year: 230
- Loaded cost per hour: £22

Arithmetic:
- Minutes saved per day: 30 × 5 = 150
- Hours per year: 150 × 230 ÷ 60 = 575
- Value per year: 575 × £22 ≈ £12,650

Modelled from the assumptions shown. Actual time saved depends on the mix of queries and how much review each draft needs.

Page: https://devraxo.com/work/calder-and-rowe/