CFOtech Australia - Technology news for CFOs & financial decision-makers
Australia
What Happens When Intelligence Automates Itself

What Happens When Intelligence Automates Itself

Wed, 18th Nov 2026 (Today)
Rob Harvey
ROB HARVEY Chief Product Officer Sidetrade

A decade ago, automation meant wiring a rule. An invoice lands in an inbox and moves to accounts payable. A payment arrives and cash application matches it against an open receivable. Useful, and mindless by design.

AI agents change that equation. Give an agent an objective and it breaks the problem into steps, picks a method, acts, watches what happens and adapts. This is intelligence automating automation, and it is a genuinely new thing to manage, which is exactly why so many capable finance teams are circling it carefully rather than diving in. Their caution is good judgement looking for a safe way in. Finding it starts with being honest about what agents can and cannot do.

Here is a thought experiment I keep coming back to. Could two people, armed with AI agents, rebuild your entire Order-to-Cash operation in two to three months?

On the technology, they could get surprisingly far. What they could not rebuild is everything your team has decided over the years. What to automate and what to keep human. Which data to trust. How to handle the customer who always pays, but never on time. Those thousands of decisions are written down nowhere. They live in your people and your processes, and they are precisely the asset agentic AI needs to be any good.

The calls that live in your team's heads

Think of the credit manager who extends a limit for a strategic account at quarter-end because the relationship justifies it. The collector who holds back a firm reminder because a customer is mid-renewal and sales has asked for a light touch. . None of it sits in the credit policy, yet each call moves a number you report, from credit exposure and DSO to the bad debt provision and the cash forecast.

Payment data, especially at scale, can tell an agent how a customer is likely to behave. It cannot tell it how your business has chosen to respond, or why. An agent needs both.

Those calls are getting more expensive

Those calls decide how quickly cash comes in, and the cost of getting them wrong is rising. Rates are climbing again on both sides of the Atlantic, with the Bank of England back at 4% and the US Federal Reserve raising rates for the first time in three years, so every day of DSO now costs more to finance. On £500 million of annual revenue, one day of DSO ties up roughly £1.4 million of cash.  In North America, The Hackett Group found that receivables are now the biggest source of trapped working capital at the largest companies, a $773 billion opportunity. In the UK, the government puts the cost of late payment at £11 billion a year.

Deciding whether to charge late-payment interest to a valued customer, or how firmly to chase an account that is also a key supplier, carries real commercial and reputational weight. It is exactly the kind of decision that lives in your team's heads.

For the CFO, that reframes the business case. The value is not headcount saved in collections. It is a faster, more predictable cash conversion cycle. Getting theremeans capturing the judgement your team already has and building it into how the system works, so your people spend their time on the calls that need them most.

Use the first phase to write it down

The early phase of an agent deployment, where the agent drafts and a person sends, is the best capture tool most finance teams will ever have. Each time a collector rewrites a draft or overrides a suggested action, they are recording a decision they have made thousands of times without explaining it. Ask them to add a line saying why. After a few months, you have something no process manual has ever managed, a working record of how your best people make their calls. It is also an audit trail your auditors can read.

Give the agent authority one step at a time

Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Handing an agent authority before anyone has evidence it has earned it is a quick way to join them.

The pattern that works is unglamorous, and it starts in that same drafting phase. Let the agent complete work on its own only in narrow, reversible task classes,  inside limits you set by balance, risk score and action type, as you would in any delegation of authority. Reminders on low-risk accounts come first. Credit holds, payment plans and write-offs come last, if ever. 

Decide before the pilot starts what good looks like. Track how often drafts are accepted unchanged, how heavily the rest are edited, and how often a person overrides the agent's choice. Track outcomes too, comparing DSO on agent-managed accounts against a control group.  Set the threshold for widening its mandate in advance, so enthusiasm does not make the call. If acceptance falls or overrides rise, the agent steps back a level until the numbers recover. A good controller would treat a new hire the same way.

Every correction teaches the system

Run it that way and the loop starts to compound. The agents propose actions and handle the routine ones inside their mandate. Your people make the calls that matter, and every correction they make teaches the system. 

Could two people and a stack of agents rebuild your operation in three months? For most companies, the honest answer is closer than you would like on the plumbing, and nowhere near on the judgement.

For a CFO, your team's judgement is an asset that never appears on the balance sheet. It is also the part a competitor cannot copy. The teams that get the most from AI will be the ones that find a safe way to put it to work, one earned step at a time.