Manage cash: what software and AI can take over, and what still needs a person

Software can already match bank activity to the ledger, build the cash position, generate payment files and post routine entries. People still need to approve payments, decide on investments outside policy, judge the forecast and negotiate with banks. None of it works until bank data, account mappings and remittance details are clean.

Steps software can run now

Reconciliation is the clearest candidate. Bank statement feeds arrive in standard formats, and matching rules can clear most items against the ledger without anyone touching them. The cash position then builds itself from those balances and from known payments already scheduled.

Receipt matching sits close behind. When a customer payment carries a usable reference, a rules engine links it to the open invoice and the payer account. Where a payment covers several charges, the system can apply amounts in a set order, for example penalties first, then fees, then interest, then principal, provided someone has written that order down. Returned checks and failed debits can also be reversed automatically once the bank's return codes are mapped.

Payment file creation is mechanical. Approved payments go out through the bank channel in the right format, with status messages flowing back. Sweeps into money market funds or similar instruments can run on standing instructions when balances cross agreed thresholds.

The accounting entries that follow cash movement, such as bank charges, interest earned and transfers between own accounts, are repetitive enough to post on rules. Fee analysis also suits software. Bank analysis statements can be compared line by line with the agreed fee schedule, and variances flagged.

Steps that still need a person

Releasing money is the obvious one. Automation can prepare a payment run, but a human with authority should approve it, and the approver should not be the person who set up the payee. Changes to vendor bank details deserve a call back to a known contact, because fraud attempts target exactly this moment.

Unidentified receipts need investigation. A deposit with no reference, a partial payment or a collection that belongs to an agreement nobody recognises takes phone calls and judgement. Software can queue these and suggest likely matches. It cannot confirm them.

The forecast is a judgement product. Systems can project from history and from open payables and receivables, yet the large items that move the number, like a tax payment, a capital purchase or a delayed customer, come from conversations. Someone has to collect those and decide how much to trust them. Reporting large expected deposits and disbursements to treasury in advance depends on the same human gathering.

Investment choices outside standing policy, borrowing decisions and anything touching covenants belong with a person. So does the banking relationship itself. Negotiating fee schedules, disputing charges and deciding to consolidate or open accounts are commercial conversations.

Where AI helps and where it does not

Machine learning earns its place on the messy edges. It can read free text remittance notes, learn which customers pay under a parent company name and propose matches that rigid rules miss. It can spot unusual payment patterns worth a second look. Forecast models can improve the baseline for routine flows.

It should not be the final word on releasing funds, on clearing an unexplained difference to a suspense account or on a forecast that feeds a funding decision. Those need someone accountable who can explain the call.

What has to be true about the data first

Every bank account must be mapped to a single ledger account, with an owner. Accounts that nobody claims break reconciliation before it starts.

Bank feeds need to be complete and arrive in a consistent format. Missing statements or manual uploads undo the benefit.

Customers and vendors need clean master records, including verified bank details and stable identifiers. Remittance references must actually reach the cash team, not stop in an email inbox somewhere in sales.

Transaction codes from each bank should be translated into a common set, so a fee at one bank looks like a fee at another. The order for applying mixed payments and the rules for writing off small differences must be agreed and written down. Without that, automation simply hard codes whatever one person happened to do.

Questions to ask the people who run the process

• When a payment arrives with no reference, what do you actually do with it, and who do you ask?
• Which matches do you clear by hand even though the system offers a suggestion?
• Are there bank accounts you check outside the main system, and why?
• Where do the numbers for the forecast really come from? Which ones do you adjust before sending?
• How do you hear about a large payment that is coming? Does it ever arrive as a surprise?
• Who can change a supplier's bank details, and how is that change checked?
• When the bank charges something unexpected, who notices, and does anyone go back to the bank?
• Which spreadsheets would stop the process if they disappeared tomorrow?
• What workaround would you be embarrassed to show an auditor?

APQC's Process Classification Framework® (PCF) is an open standard developed by APQC, a nonprofit that promotes benchmarking and best practices worldwide. To download the full PCF or to view definitions and measures, please visit www.apqc.org/pcf.