AI drafts the transaction. A human with approval authority releases it. The audit trail records which was which.
Inbound email or document is classified, screened for injection and PII, then parsed into structured line items. Non-RFQ mail is filtered out before it ever reaches the pipeline.
Buyer is matched to an existing account — a new address on a known company is flagged as suspect. Free-text product lines are matched to the catalogue by salt, strength, form and pack.
Vendors are scored on five factors, prices outside their validity window are excluded, and a landed-cost sheet is assembled with freight, duty and FX.
Priced in the customer's own currency with margin computed per line. Where costs are missing, the run parks and produces a partial quote rather than guessing.
No agent can release a customer-facing document. Every AI output lands as a draft that a named approver must accept.
Once a person edits a field, agents stop overwriting it. Your correction is final.
Every mutation is written to a hash-chained audit log, attributed to a human or a specific agent run.
Each revision snapshots the full quote, so you can prove exactly what was sent, when, and at what margin.
Admin, Operations, Approver and Viewer. Only an Approver can clear a human-in-the-loop gate.
Every row carries an organisation ID enforced at the database level, not just in application code.
Workflows are checkpointed graphs. A run waiting on your approval survives a restart and resumes where it stopped.
The system learns your pricing habits and customer preferences as evidence-backed memories, each with a confidence score.
Hybrid retrieval over your own catalogue and history means every drafted figure can be traced to its source record.