The situation
The writing was quick. The looking-up was not.
The fund's written-enquiry team answered questions about contributions, insurance and rollovers. A good reply took a couple of minutes to write, and several more to assemble: the member's account in the registry, recent transactions, the insurance cover in force and the right product disclosure wording.
We timed it. Most of the handling time went to research, not to writing. That was the part worth automating.
What we built
- 01
An MCP server over the member registry
Read-only, with field-level permissions agreed with the fund's privacy officer. Every call logged with the staff member and enquiry it served.
- 02
A drafting agent
Classifies the enquiry, gathers the member context through MCP and drafts a reply, with each fact linked to the record it came from.
- 03
An evaluation set
Six hundred historical enquiries with the replies actually sent, used to score every model and prompt change before release.
Measured in discovery
Where the minutes went, before
- Find the account1.6
- Check transactions2.1
- Check insurance1.4
- Find wording1.8
- Write reply2.2
| Find the account | 1.6 |
|---|---|
| Check transactions | 2.1 |
| Check insurance | 1.4 |
| Find wording | 1.8 |
| Write reply | 2.2 |
The draft is right most of the time. When it isn't, the citations show us why in seconds.
What changed, and what did not
Changed
- Research happens before a person opens the enquiry
- Replies cite the record they rely on
- New staff reach full speed sooner
Unchanged
- A person reads and sends every reply
- No access to anything outside the agreed fields
- Data stays in Australian regions
