Agentic Treasury Cash Positioning: Building a Rolling Cash Forecast Before a Shortfall Hits

The first real obstacle has nothing to do with forecasting — it's that banks don't agree on how to describe a balance. One relationship delivers a BAI2 file overnight, another exposes an open-banking API with intraday balances, a third still requires logging into a portal because no automated feed was ever set up. Each source has its own definition of a value date versus an available date, its own cutoff time relative to a wire or ACH batch, and its own handling of weekend and holiday postings. The agent's first job is mapping every one of these into a single normalized balance record per account per entity — ledger balance, available balance, and currency — so that everything downstream is working from the same clock instead of quietly comparing a same-day API balance against a next-morning BAI2 file and treating the mismatch as real movement.

A bank's available balance and the ERP's cash-account balance are almost never the same number at any given instant, and the difference is exactly the information a position needs to be accurate: outstanding checks that haven't cleared, wires initiated but not yet debited, deposits in transit, and bank fees or interest postings the ERP hasn't recorded yet. The agent pulls the ERP's cash-in-transit and outstanding-item detail and nets it against the bank-reported balance to produce a true current position per account, then rolls those up by entity and by currency. This step is what makes everything that follows trustworthy — a forecast built on an unreconciled bank balance just compounds a stale starting point forward in time.

With a true starting position established, the agent pulls the near-term items that are already known but haven't hit the bank yet: the next scheduled AP payment run and its total, payroll runs on the calendar, debt service and lease payments due, and an AR collections estimate built from open invoice aging rather than assumed on-time payment. These get laid onto the reconciled starting position day by day across a rolling window — commonly thirteen weeks, since that's far enough out to see a drawdown or shortfall coming without pretending to forecast the parts of the business that genuinely aren't predictable yet. The output is a trajectory, not a single number: today's position, and where it's projected to sit each of the next several weeks under what's already committed.

A forecast is only useful if it's tested against something concrete — a minimum operating balance the business has set for itself, an undrawn capacity limit on a revolving credit facility, or a covenant-linked liquidity threshold from a loan agreement. When the rolling projection crosses one of those lines at any point in the window, the agent surfaces it to treasury with the specific drivers behind the dip: which entity, which week, and whether it's being driven by a concentrated AP run, a large AR balance aging past its expected collection date, or a scheduled debt payment landing awkwardly against a low point in the cycle. The agent can size what a credit-line draw of a given amount and timing would do to the projected trajectory, but initiating the draw, negotiating with a customer on collections, or rescheduling a payment run stays a treasury decision — the agent's job ends at making the shortfall visible early enough that the decision isn't made under pressure.