Credit control is the finance process I would hand to an always-on AI agent first. The reason is simple. Most late payers are not refusing to pay. Their payment is stuck. The invoice went to someone who left in the spring. It is missing a purchase order number. It is waiting for an approval, or it is quietly disputed and nobody on your side knows yet. Clearing those blockages is reading, matching and chasing, which is exactly what an agent does well. It watches the ledger, chases every overdue invoice in your tone of voice, logs promises to pay, spots disputes the day they are raised and reports the cash position every morning. A named person still decides anything with consequences: write-offs, payment plans outside the written rules, account stops and legal action. The return is cash in sooner, and a credit control function that keeps up with growth without another hire.
Late payment is mostly a process problem
Ask UK businesses why they pay their own suppliers late and the answers are mostly about process. In a survey of 300 businesses for the Department for Business and Trade, 36% blamed administrative errors, 31% disputed invoices and 23% technical issues such as lost invoices. Only 1% admitted to paying late on purpose (DBT and IFF Research, 2024). Seen from the other side, 24% of suppliers put the late payments they receive down to administrative errors, rising to 36% among large firms, which the researchers link to stricter processes such as "no purchase order, no pay".
Atradius, the credit insurer, sees the same pattern in its 2026 survey of 210 UK firms. Around two thirds report late payment from business customers, affecting about a quarter of invoiced B2B turnover. Three of the four most common reasons are banking delays, internal approval delays and complex payment processes. And 73% of overdue invoices are less than 30 days late (Atradius Payment Practices Barometer UK, 2026).
Customer cash flow is still the top reason in that survey, at 51%, and no agent fixes a customer who cannot pay. It can spot one sooner, though. Much of the rest of the overdue ledger is a queue of small, fixable blockages, and a queue is what agents are built for.
What a day of DSO is worth
Days sales outstanding (DSO) is how long, on average, you wait to be paid after you invoice. It is one of the first numbers a lender, a board or a PE operating partner looks at when the conversation turns to working capital, and the arithmetic is worth doing on your own figures.
Take a business turning over £40 million a year. Each day of DSO holds roughly £110,000 of its cash (£40 million divided by 365). Take five days out and about £550,000 comes back onto the balance sheet, and stays there for as long as the improvement holds. That is cash you do not have to borrow. The numbers here are illustrative; a baseline measured on your own ledger replaces them.
Five days is not a fantasy. Finning, the world's largest Caterpillar dealer, which operates in the UK, Ireland, Canada and South America, cut DSO by five days and its past-due share from 50% to 24% in four months after bringing AI into collections and disputes (HighRadius, vendor case study). Freight forwarder OVRSEA, which pays carriers before its own customers pay it, took DSO from 38 to 27 days in its best month and averages 34 days overall, with 70% of its collection actions automated (Upflow, vendor case study). Both are the vendors' own stories, so weigh them accordingly. Both point the same way.
What a credit control agent does all day
Every process I put on agents follows the same seven steps. In credit control they look like this.
| Step | What the agent does in credit control |
|---|---|
| Watch | Monitors the sales ledger, the credit control inbox and the bank feed, continuously |
| Read | Reads customer replies, remittance advice, dispute emails and promises to pay |
| Match | Ties each reply to the right customer, invoice, purchase order and contact |
| Decide | Applies your written chasing rules: timing, tone, who to copy, when to stop |
| Act | Sends the chaser, re-sends the invoice with the missing PO, logs the promise to pay, notes the account |
| Escalate | Hands disputes, broken promises and anything outside the rules to a named person, with the full history |
| Prove | Reports cash received, cash promised, aged debt and every action taken, every morning |
The middle rows are where the value sits. A reminder sequence sends the next email on a schedule. An agent reads the reply. When the customer writes "we never received the PO", the agent finds it, re-sends the invoice with it attached and sets the next check, instead of firing the third reminder in a sequence. When the customer writes "we'll pay on the 28th", it logs the promise, stops chasing, and checks the bank feed on the 29th. I describe how the new platforms make this cheap in always-on AI agents are here.
The cash is in the disputes
A disputed invoice is the most expensive kind of late invoice, because nothing moves until someone notices it. Finning started with more than 4,000 open disputes worth $32 million and an average resolution time of over 56 days. Once gathering the claims and backup documents was automated, resolution time fell to 29 days (HighRadius, vendor case study).
The agent's job on disputes is narrow and valuable. Spot the dispute the day the customer raises it, even when it is buried in the third paragraph of a reply about something else. Gather the evidence: the order, the delivery note, the email where the price was agreed. Route it to the person who can resolve it, with all of that attached. Then keep the clock visible, so a dispute cannot sit in someone's inbox for six weeks. Most of what is actually agreed in a B2B relationship lives in email, which is why I call email your most underrated data source.
If you sell to larger companies, the timing may soon matter legally too. The Commercial Payments Bill, introduced to Parliament in May 2026, would cap payment terms for large firms paying smaller suppliers at 60 days, make interest on late payment mandatory at 8% above Bank of England base rate, and give suppliers a right to a fixed sum where a customer raises a dispute late or without enough information (GOV.UK). If it passes as drafted, a dated record of when each dispute was raised, and what it said, becomes evidence.
What the agent never does
The rule is simple. The agent chases. A named person decides. It never writes off a debt, never agrees a payment plan outside the limits you have written down, never stops an account, never claims the statutory interest you are entitled to, and never mentions legal action. Each of those is a commercial judgement about a relationship. It belongs with the credit manager or the finance director, who gets the full history in front of them before deciding.
Tone matters as much as the rules. The same government research found the most common reason businesses do not formally pursue late payment is that they do not want to damage the customer relationship (33%). That instinct is right, and the agent should share it. Good chasing in B2B is polite, specific and persistent: the invoice number, the PO, the amount, how to pay, and an offer to help if something is wrong. OVRSEA's finance team describes its automated reminders, some sent before the due date, as friendly rather than aggressive. Write your tone into the rules, approve the templates, and read a sample of what went out every week.
Why credit control projects stall
When a collections project disappoints, the technology is almost never the cause. The cause is the records the agent reads and the rules nobody wrote down. Chasing an invoice the customer paid last week, because the cash arrived but was never allocated, is the fastest way to annoy a good customer. So is chasing an invoice that a credit note already cancelled, or emailing a contact who left months ago.
So the records come first: unallocated cash cleared, credit notes posted, customer contacts and accounts payable inboxes current, purchase order numbers captured on every invoice. Then the rules: chase timings by customer group, who is copied at each stage, what counts as a dispute, and which customers are never chased automatically. That last list matters more than people expect. Your ten largest accounts may be better served by a person on the phone, with the agent preparing the call.
None of this is exotic. It is your finance team's own knowledge, written down. Do it before the agent goes live, because an agent that works overnight can also be wrong overnight.
How I would roll it out
The method is the one I used at one of Europe's largest insurance brokerages, where the agentic system I led resolved 67% of customer service cases without a human. That was customer service, not credit control, and I will not pretend otherwise. What transfers is the discipline: the data sorted out before any model was built, a rollout that started at 10% of volume, and an audit trail on every case.
In credit control it runs in four steps. First, a baseline: DSO, the past-due share by age band, dispute age, promises kept, and the hours your team spends chasing. Second, shadow mode, where the agent drafts every chaser and logs every promise for a person to approve. Third, live chasing on about 10% of accounts. I would start with the long tail of smaller customers, who get the least attention today. Fourth, scale as the numbers justify it. A first process like this typically takes six to eight weeks as a Process Build.
What the finance director gets
The morning report is where you see it working: cash received, cash promised and by when, disputes open and their age, and every action taken on every account. The value is cash in sooner and a credit control team that can absorb growth without hiring ahead, not a smaller team. The people who chase today can spend their time on disputes, large accounts and the conversations that need a person. In a mid-market business that is growing, that is the point: as volume rises, the agent absorbs the routine chasing and your credit controller keeps the judgement.
Measure it on five numbers: DSO, past-due share, the share of promises to pay that are kept, average dispute age, and cost per invoice chased, including the agent's running cost. Put that last one on the dashboard from the first week. Agent spend is a variable cost and deserves a budget and an owner like any other.
If your overdue ledger has grown faster than your credit control team, I would be glad to hear about it. A Process Audit measures your baseline in two weeks and tells you honestly whether an agent would pay, or whether the fix is simpler. For the other finance processes that suit agents, with the published evidence for each, see what AI agents can run. Or just get in touch.
Related: Always-on AI agents are here. The agent is now the easy part · Email is your most underrated data source · Why grounded AI agents are cheaper to run · Uber burned through its token budget by April. Your business will be next
