There are eight jobs in our catalog and we could have started with any of them. Month-end close is bigger. Cash forecasting is more strategic. Payables touches more transactions. We started with collections, and it was not a coincidence. Collections is the most agent-shaped job in finance, and understanding why tells you what to automate next.
It is unambiguous
An invoice is either paid or it is not. It is either overdue or it is not. The outcome — cash in the bank — is measurable to the rupee and to the day. There is no debate about whether the agent did the job. DSO went down or it did not.
Compare that to reporting, where 'good' is a matter of taste, or planning, where you will not know if the forecast was right for a year. An agent's first job should have an outcome you can put on one line.
Everyone already hates the current number
Ask a founder what their overdue receivables are and watch their face. Nobody is happy with it. Nobody thinks it is being handled well. There is no incumbent process to defend — usually the process is 'someone remembers to send an email' — so there is nothing to displace except a bad feeling.
The best first job is the one the customer is already embarrassed about.
It requires judgment, but bounded judgment
Pure rules fail at collections quickly. 'Send a reminder at day 7, 14, and 30' breaks the moment a customer replies 'we're processing it' or 'your invoice has the wrong PO number' or 'call me'. Each reply needs to be read, understood, and acted on differently. That is judgment, and it is why collections has resisted automation.
But the judgment is bounded. There are perhaps a dozen kinds of reply. The decisions are: send, wait, escalate, ask a human. A language model handles the reading and classifying well; the agent handles the sequencing; the human handles the small number of real decisions. The shape of the work fits the shape of the tools.
The blast radius is small
The worst a collections agent can do, inside our guardrails, is send a reminder that was slightly too firm or slightly too early. That is recoverable with an apology. It cannot move money, cannot write off a balance, cannot offer a discount, and cannot contact an account you have marked as sensitive. A first agent should fail softly, and collections does.
It teaches the agent your customers
This is the reason that matters most for what comes next. Running collections means learning, customer by customer, who pays when and why. That knowledge — the judgment layer — is exactly what the Cash Agent needs to forecast, what the Close Agent needs to match receipts, and what the Reporting Agent needs to explain a DSO variance.
- Collections learns that your largest customer pays on day 12. Cash uses it to forecast.
- Collections learns that a customer disputes every third invoice over a PO mismatch. Close uses it to expect the short-pay.
- Collections learns which accounts are relationship-critical. Every agent respects the hold.
Start with the job that builds the most useful knowledge, and every subsequent agent starts smarter.
What this says about the order of the rest
Payables is next for the same reasons in mirror image: unambiguous, universally disliked, bounded judgment, and it teaches the agent your vendors. Close follows because it needs both. Cash needs all three. Reporting and planning come last, not because they are less valuable, but because they are only as good as the ledger and the judgment beneath them.