Your Cash Forecast Got Precise. So You Locked Up the Buffer.

A founder I work with runs a $12M distribution business with real cash swings. His treasury agent got good this spring. Thirteen-week forecast, refreshed nightly, landing within a couple points of actual most weeks. So he did the rational thing. He stopped leaving $1.8M idle in the operating account and built a T-bill ladder plus a money market sweep to earn on it. The board loved the yield line.

Then May happened. A customer's $340K payment slipped eleven days, and a supplier pulled a deposit early the same week. His forecast had flagged neither. He needed $400K that wasn't in checking, and the cash was sitting in a 4-week bill that matured in nine.

Precision became permission

Accuracy is the whole pitch on AI treasury tools, and the numbers behind it are real. Visa reported this year that AI cut cash-flow uncertainty for a set of firms from 68% to 17%. Vendors now claim better than 92% accuracy on the 13-week horizon. Roughly 82% of midsize companies have started pointing agents at working capital, and CFOs increasingly treat idle cash as a growth lever rather than a safety margin.

What follows is a move almost nobody names out loud. A sharper forecast feels like permission to run a thinner buffer. If the model lands within two points, why hold three weeks of float doing nothing? So the cushion gets swept into ladders and sweeps that pay 4% instead of zero. Real efficiency. What's new is the exposure it creates.

Where the other 8% lands

A 92% forecast is wrong 8% of the time, and those misses aren't random noise you can average away. They cluster on exactly the events a model can't see. One customer decides to stretch you. A supplier changes terms. Then a chargeback batch lands. Those arrive without warning, and they tend to arrive together, because the same pressure that makes one customer slow makes three slow.

Your buffer was never inefficiency. It was the shock absorber for the 8% the forecast will always miss. Move it into a 4-week bill and you haven't removed the risk. You've guaranteed the cash is least reachable on the day you need it most.

Forecast accuracy tells you what will probably happen. Liquidity is about what happens when it doesn't. Those are different jobs, and a yield sweep only does the first one.

Keep the buffer liquid, sweep the rest

None of this means leaving $1.8M earning zero. It means separating the two piles on purpose. Size a liquidity buffer against your worst realistic miss, not your average one. For the distribution client we set it at the largest single receivable plus one payroll, held in same-day money market with no lockup. Everything above that goes into the ladder.

Then stress the forecast, not just the point estimate. If your top two customers each paid ten days late in the same week, could you cover payroll without breaking a bill early? If the answer depends on a maturity date lining up, your buffer is too thin.

The 92% is a genuine upgrade. Just don't spend the 8% you can't see to fund a yield line you can.