Cash Forecast Accuracy Is a Data Problem Disguised as an AI Problem

Cash forecast accuracy is not mainly a model problem. It is a data-trust problem first.

If you sit in a CFO or treasury seat — or you run cash position, recon, or AR timing day to day — smarter tools can help. Sequencing decides how useful they are. I'm not anti-AI. I'm pro-foundation. Agents amplify a joined-up cash picture. They don't invent one.

Why it matters: when a 13-week forecast misses, the post-mortem rarely starts with "the algorithm was weak." It starts with fragmented actuals — bank cut-offs that don't match ERP, intercompany in a private tab, AR timing the finance team knows has moved but the forecast hasn't, unmapped entities that only show up when actuals won't reconcile. If finance can't reconcile actual cash to a single owned definition across systems, even a strong model will struggle on a number you'd take to a lender or a board.

What the operating numbers show

EY India's An Agentic AI Adoption Playbook for CFOs and Treasurers (3 September 2026) puts numbers on what many finance teams already feel. Treasury often spends 60–70% of its bandwidth on manual and low-value work. Spreadsheet-led forecast variance often exceeds 20%. Mature treasuries may run 50–100 interconnected spreadsheets. More than half of corporates still rely on manual reconciliation. That's capacity trapped in stitching, not a lack of ambition.

Workday's ANZ work on AI agents in finance sharpens the same point: 89% say data isn't ready for agents; 87% plan to retain human-in-the-loop; 77% aren't seeing ROI; and 95% put risk and compliance among their top AI priorities. Govern the stack, then scale agents where outcomes are measurable.

Soft on vendors: plenty of tools can help. They earn their seat after sources, recon, and ownership are named — not before.

Before you scale the next forecast model, do this

1. Name the sources. ERP cash, AR, AP, and intercompany by entity. Bank statements including project accounts. Facility schedules. Material contracts that drive timing. Market feeds only where they change a decision you actually make.
2. Make recon boring. Bank-to-ERP with named exception queues. An entity and account master that matches how cash is reported. A forecast-to-actual variance log with drivers — not just a percentage.
3. Write ownership down. Who owns the cash definition for board and bank reporting. Who owns bank masters. Who owns 13-week assumptions. Who can change a mapping — and who must approve it.
4. Start where trust compounds. Daily cash position you can defend. Recon auto-match on high-volume items with humans on exceptions. Forecast hygiene — actuals locked on a schedule, assumption owners named — before a new model.
5. Keep propose vs approve clear. Agent proposes; named human approves when cash moves, masters change, or the output goes to board, bank, covenant, or audit.

What good looks like

One owned cash definition. A daily position that doesn't need five files. A 13-week cadence with assumption owners and a variance log that names drivers. Soft automation on the boring matches; humans on the exceptions. AI where it earns its place — after the foundation holds.

If you can't clear those, you may have a nicer model. You don't yet have a forecast you'd defend.

Forecast accuracy gets sold as an AI upgrade. In practice it is decision design: join the sources, name the owners, lock the cadence, then put models where they amplify trust.

Hygiene before heroics — put the joined cash picture into a weekly cash cadence with owners and variance drivers.