A 9-step manual payables process, read and matched by an agent.
A premier Philippine real estate conglomerate ran accounts payable by hand: billing countered at a desk, encoded into a weekly spreadsheet, checked, entered into SAP, and routed for approval on paper. Volumes were on course to quadruple. The reading was the bottleneck, so that is what we pointed the agent at.
A premier Philippine real estate conglomerate.
High-rise luxury residential developer, 35+ years of operations, managing properties across Metro Manila and expanding nationwide. Unnamed here under a mutual NDA.
Seven minutes a transaction, and the judgement left where it was.
Figures below are the client’s own, measured against their pre-project baseline and agreed for release. Where a number is a projection it says so.
Volume was the reason, not cost.
Transactions were on course to quadruple in two years. At 20 minutes of reading each, the pile grows faster than the team can be hired against — which is what makes reconciliation the right place to put an agent, and the wrong place to put more people.
Bars are scaled against the 2026 manual figure. The tilde on 2026 is deliberate — it is a projection, and every projected number on this page wears one.
Nine steps became seven, and three of them are still a person.
The steps that disappeared are the ones where somebody retyped what a document already said. The review step was kept on purpose: exceptions are handled by the team, not by the agent.
- 1Counter billing documents
Paper-based receipt - 2Encode to weekly payable monitoring
Internal Excel - 3Review and validation of documents
Manual checking - 4Generate goods receipt
SAP system entry - 5Prepare request for payment
Manual RFP prep - 6Encode billing to RCDT
Excel monitoring - 7Routing for approval
Manual - 8Scanning of RFPs
Manual - 9Transmittal to finance
Manual
20 minutes a transaction at the outset, error-prone, and bottlenecked at scale. Nine steps, and a person performs all nine.
- 1Vendor uploads documents via portal
Vendor · PO · SI · GRN · DR - 2Fields extracted and validated
Agent · three-way match - 3Team reviews flagged items
Person · only what was flagged - 4Request for payment auto-generated
Agent · pre-filled from extracted data - 5Encode billing to RCDT
Person · internal Excel monitoring - 6Routing for approval
Person · system-guided workflow - 7System transfers to finance
Agent · payment voucher auto-generated
7 minutes a transaction, 85%+ extraction accuracy, tested to 6,000 transactions a month.
A two-week proof first, because the three-way match had to be shown.
Nobody had to take the extraction accuracy on trust. The first two weeks tested it on their own documents against a stated floor, and production only started after that number was met.
The honest version, including what the agent did not fix.
Published because the difficulties are the useful part of a case study. Every one of these changed how the product works.
- !Vendors submitted combined PDFs, which had to be unbundled before a three-way match could run at all.
- !Transaction volumes more than doubled mid-engagement, raising the scalability bar after the design was set.
- !Dual-entity support required isolated workflows on shared infrastructure.
- !SAP integration was deferred, so the Excel-based RCDT monitoring had to be bridged as an interim.
- !Document formats varied across 58+ vendor types, which broadened what the extraction had to handle.
- ›A proof of concept first de-risked the production investment and built confidence before budget was committed.
- ›AI reduced effort per transaction, not operational complexity. Headcount was redeployed, not eliminated, and saying otherwise would have been a false promise.
- ›Human review at the exception stage is not a transitional measure. It is what makes the throughput safe.
- ›Phased delivery let scope be refined against feedback instead of guessed at up front.
- ›Separating vendor, materials and AP access cut error propagation significantly.
Their AWS account, their documents.
Deployed on AWS with the document processing on Amazon Bedrock. The portal, the workflow engine and the dashboards are ReconcileAI; the account and the keys are the client’s.
Bring us one exception type.
This engagement started with three document triads and a stated accuracy floor. Yours can start the same way.