If you're in a Brisbane CFO or treasury seat — or running 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. Cash forecast accuracy gets framed as a model problem. On the desk, it's a multi-system trust problem first — and an AI opportunity once that trust is in place.

What the operating numbers actually show

EY India's An Agentic AI Adoption Playbook for CFOs and Treasurers (3 September 2026) puts numbers on what many of us 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. Controllers and AR leads feel it as much as the C-suite.

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.

Where variance actually comes from

When a 13-week cash 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 desk 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. The EY sequencing point lands here: build the joined-up treasury data picture across ERP, banks, contracts, and feeds first. Agents amplify that foundation — they don't replace it.

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

Minimum viable foundation (checklist)

You don't need a five-year platform programme. You need named ownership.

Sources: ERP cash/AR/AP/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.

Reconciliations: Bank-to-ERP with named exception queues; entity/account master that matches how cash is reported; forecast-to-actual variance log with drivers, not just a percentage.

Ownership: 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.

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.

Agent proposes; named human approves when cash moves, masters change, or the output goes to board, bank, covenant, or audit. That split matches how Australian desks already think about controls.

CTA: Wrestling the same cash-data stack? Compare notes at financesignal.ai — practical thinking from the finance desk, without the funnel.