The workflow before
When Exactera onboarded a transfer pricing client, subject matter experts read the client's raw documents, PDFs and Excel workbooks, to find the legal entities and financial figures, then mapped them by hand into Exactera's models. The work was slow, varied by analyst, and left little record of why each mapping was made, which matters when the analysis has to stand up in an audit.
What we built
In a four-week sprint we built an LLM-based agent that reads incoming client documents, extracts the key entities and figures, and proposes how each maps into Exactera's data model. Documents pass through an ingestion module and OCR before the model sees them.
Experts review the agent's work in a human-in-the-loop interface, confirming or rejecting each suggestion on screen. Every decision is written to a relational database with the source location, the AI suggestion and the reviewer's call, so each mapping has an audit trail for transfer pricing defense.
Before and after
| Measure | Before | After |
|---|---|---|
| Mapping a client's financial entities | 2 to 3 days of manual review | Under 1 day, AI draft plus expert review |
| Audit trail for mapping decisions | Partial, in analyst notes | Every decision logged with source and reviewer |
What's next
The next step is to extend the agent from entity mapping into the rest of onboarding, such as flagging missing documents and drafting comparables analysis, with experts still approving the exceptions.
