Context
A pre-call account summary for reviewers at a business in regulated finance, where the facts about an account are spread across several kinds of record.
The problem
A reviewer has about 30 seconds to size up an account before a call, with the facts spread across agreements, invoices, payment history, and filings. Asked to summarize that, an AI will confidently invent a dollar figure, and one wrong number is enough for a reviewer to stop trusting the whole page.
Constraints
- Readable in about 30 seconds, before the call starts.
- Every amount and date had to be verifiable against the system of record.
- Where data was missing, the summary had to say so rather than fill the gap.
Approach
The assistant cannot be handed facts. It reads data only through tools that query the system of record, so every figure traces to a real source. The rule is strict: every amount and date must appear word for word in the data, and “no data” beats a guess. It also looks for contradictions, such as an account marked current that still carries a past-due balance, and flags them instead of smoothing them over.
Key engineering decisions
Facts only through tools
The model reads the system of record through query tools instead of being given data in its prompt, so every figure has a traceable origin.
Verbatim values, or “no data”
An amount or date that does not appear word for word in the source is not stated, and a gap is reported as a gap.
Surface contradictions
Records that disagree, like a current account with a past-due balance, are flagged for the reviewer to check first rather than silently reconciled.
Result
A quick, source-backed read a reviewer can trust at a glance, with a clear “check this first” note wherever the data disagrees with itself.