Every other case study here ends in a number. This one does not, because nothing is in front of real users yet. When that changes, this page changes.
Finance teams already have the answer somewhere. It is in the ERP, or a PO, or a GL code, or an email thread nobody can find. The work is not deciding what to do. It is assembling enough context to decide anything at all.
Neo is one of four solution lines at Neoflo, alongside procure-to-pay, order-to-cash and record-to-report. The other three run workflows. This one answers questions about them.
Batch ingestion across the systems finance already uses: the ERP, AP inboxes, invoices, payments, purchase orders, GL codes, vendor and client records, chats and PDFs. A knowledge graph links vendors to invoices to approvals to payments to contracts. A vector index sits alongside it.
Questions get answered in natural language, grounded in the company's own documents, with citations and an audit trail. Outdated information is superseded rather than left to compete with the current version.
A write-capable token is not made safe by a prompt telling the model not to write. Read-only is enforced at the tool layer.
Rejected prompt-level restriction entirely.
An integration gets no more implicit trust than a document. This is the whole ballgame for an assistant that reads your email.
Rejected treating an external source as more trustworthy than an internal one.
Show the most recent and flag the conflict.
Rejected silently picking one, which is how a confident wrong answer gets made.
Stricter means more I don't know, at higher precision.
Rejected fixing that trade-off in the product, because workspaces do not share a risk tolerance.
