Finance teams are buried. AP clerks spending three hours a day matching invoices against POs. AR teams chasing payment confirmations across email threads and WhatsApp messages. Month-end reconciliation happening in spreadsheets at 11pm because the ERP data doesn't reconcile itself.
The tools exist — ERPs, RPA bots, dashboards — but they don't communicate, they break on exceptions, and they were built for the 2010s. Every workflow has a human in the middle, duct-taping systems together.
CFOs don't need another dashboard showing them where the inefficiency is. They need the inefficiency removed. That's the gap Neoflo is closing.
Neoflo is an AI-native back-office automation platform. Not a copilot that suggests actions. An agent that executes them.
The focus is purchase-to-pay (P2P) and order-to-cash (O2C): invoice ingestion, three-way matching, exception routing, payment approvals, and reconciliation. The workflows CFOs hate most, now handled end-to-end.
The wedge: non-trade invoices. These are the invoices that arrive via email, helpdesk tickets, and supplier portals, outside the structured ERP data model. They're typically processed manually, have the highest error rates, and are the first thing finance teams ask to fix. We start there and expand.
I joined Neoflo in November 2025 as Head of Product, the first PM in the building.
My work spans three tracks:
Product definition: translating finance team pain into specs. What does "3-way matching" actually mean when non-trade invoices don't have a formal PO? What's the minimum viable confidence threshold before the system auto-approves vs. escalates? How does the UI need to work for a 50-year-old AP manager who's never used AI tools?
Client delivery: owning onboarding, UAT and go-live across our enterprise deployments. I built the UAT framework, defined the test suite with deliberate error cases, and own the go-live criteria.
Roadmap strategy: working directly with the founders to define what we build next. Which workflow gets productized, which ERP integrations unlock the most clients, how we price the platform.
Our deployments run with enterprise clients across Southeast Asia and India, automating non-trade invoice workflows that arrive through helpdesk queues rather than the ERP.
The process before: invoices arrive via helpdesk → AP team manually reads each one → cross-references against the ERP → routes to the right approver → files after payment. Full cycle averaged 4–6 days per invoice. Error rate from manual data entry: ~12%.
With Neoflo: invoice lands in the helpdesk queue → extracted and classified automatically → matched against ERP purchase orders via our connector → exceptions routed based on configured rules → approved invoices queued for payment.
The result across those deployments is a 50%+ efficiency gain. That's a measured number in production, not a projection. Cycle time dropped from days to under 24 hours on the clean path.
The next phase is pushing that to 70–80%. That one is a target, not a result. Getting there means automating further into the exception path, which is the hard half.
You cannot automate a process you're not allowed to see. This is the one nobody warns you about. The SOP is out of date or was never accurate. The system logs tell you what happened, not why. The real process lives in one person's head, and getting it out means watching them work: their inbox, their screen, their side spreadsheets. That is exactly the access a security review exists to prevent, and both sides of that conversation are correct. Most of the discovery work is finding a way through that without asking anyone to hand over their mailbox.
ERP data is messy. Enterprise ERP data at this scale isn't clean. PO numbers aren't standardized, vendor names have variations, and cost centres drift over time. The matching logic needed fuzzy matching, not exact matching, with tunable confidence thresholds per rule.
Exception handling is the real product. 80% of invoices are straightforward. The 20% that aren't are where the value is, and where most automation tools break. We built a structured escalation path: the system flags what it can't resolve, packages the context, and routes to the right human with a pre-filled form. The human makes one decision, the system learns.
Buyer confidence in AI decisions. CFOs trust AI less than AP managers do. The product had to show its work. Every match includes the source data, the confidence score, and the rule it matched against. Not a black box. An auditable trail.
I'd start with the ERP connector earlier. We built the helpdesk ingestion layer first because that's where the invoices entered, but the blocking dependency was always the ERP connector. The integration work was longer than scoped and gated everything downstream. Next time: build the data connections first, UI second.
I'd also build the error taxonomy before writing a single spec. The first version of our UAT suite was too broad. It caught big errors but missed the subtle ones that would actually show up in production (vendor name mismatches, duplicate invoice numbers with different dates, split PO lines). The test suite got smarter but it cost us time mid-UAT.
Back-office automation isn't a technology problem. The technology exists. It's a trust and change management problem.
Finance teams have been burned by automation promises before. RPA bots that broke every quarter, BI dashboards nobody used, ERP implementations that took two years and delivered half the spec. Neoflo's job isn't just to automate. It's to be the first automation they actually trust.
That means showing your work, handling exceptions gracefully, and making the human who's still in the loop feel more capable, not replaced. Get that right, and the efficiency gains follow naturally.
Building AI products since 2020, from spatial computing to finance automation.
