Enterprise platforms are building agents for collections teams. The bigger gap is businesses with no finance team at all.
By Bishr Haffar, CEO & Founder of PayFlowAI
Accounts receivable software has caught the agent bug. Over the past year the big platforms have been adding AI to collections at a steady clip. Billtrust, for example, has shipped AI-assisted email handling, adaptive collection procedures, call transcription and summaries, and, in March, AI-driven credit-line analysis. I think the direction is right. I also think most of it is pointed at the wrong person.
Look at who these products are built around: collectors, credit managers, finance teams with an ERP to integrate and a policy to enforce. Billtrust’s own product chief describes the approach as human-first, with collectors staying in the loop rather than racing toward full autonomy. That’s a sensible call for a company with a collections department. It says nothing about the freelancer, the two-person agency or the owner-run firm, where the person who did the work is also the person who has to ask for the money.
The people who need this most
The scale of the problem is easy to underestimate. Intrum’s European Payment Report 2026 surveyed 8,385 businesses across 20 countries. It found that 57% had missed growth targets because of late payments, and 62% were paying their own suppliers late because their customers paid them late. More than half (53%) expect things to get worse over the next year.
Press enter or click to view image in full size
The detail I find most telling is that agreed payment terms have barely moved in five years. What has moved is behavior. In B2B, the gap between the terms on the invoice and the day the money lands has grown from 16 days in 2023 to 20 days in 2026. Intrum also notes that smaller businesses feel it hardest, because their financial buffers are thinner.
So the pain is real, it’s growing, and it lands hardest on the companies least likely to have anyone whose job it is to deal with it.
The unit of value is the decision
Most writing about AI in finance dwells on tasks: drafting the email, matching the payment, pulling data off an invoice. Those matter, but they’re not where the money is. The money is in the small decisions between the tasks.
Press enter or click to view image in full size
Who do I chase today, and who do I leave alone? Is this a polite nudge or a firm one? Did that reply change what should happen next? Can this wait, or do I need the cash this week?
A finance team makes those calls, imperfectly, all day. A one-person business mostly doesn’t make them at all. The invoice goes out, nobody follows up, and the awkward email gets put off until the bank balance forces the issue. I send invoices myself, so I know exactly how easy it is to let that slide.
That’s the real opening for AI in receivables. Executing tasks faster is nice. Making the decision nobody was making, and then acting on it, is where the value is.
What sales learned about AI
I come from sales, where the AI conversation is a few years ahead of finance. Gartner’s sales analysts, in a recent webinar on AI in sales, argued that the projects that deliver sit where three things overlap: a clear business objective, a working process underneath, and a real plan for getting people to change how they work. Remove any one and the project stalls, usually in a pilot nobody scales.
I’d put it in plainer terms for AR. You need a goal that’s about cash, not about using AI. You need a process that already makes sense, because software will simply do a bad process faster. And you need people to trust the output enough to stop doing it the old way.
Now apply that to a small business. The goal is obvious. But the process often doesn’t exist, and nobody is available to be retrained. This is exactly why AR tools designed for enterprises don’t carry over. They assume a process to automate and a team to bring along. For a small business, the software has to bring the process with it, with sensible defaults, and earn trust quickly.
Five decisions worth handing to software
Press enter or click to view image in full size
Who pays how. Start by learning how each customer actually pays: how late they typically are, how they respond to reminders, whether a reliable payer has started slipping. Everything else builds on this. Treating a habitual late payer and a first-time customer the same way is probably the most common collections mistake there is.
Is the invoice right? A surprising amount of late payment is really an invoice the customer’s team couldn’t process: a missing field, the wrong VAT treatment, a format their system rejects. This matters more as Europe moves to structured e-invoicing. Germany has required every business to be able to receive e-invoices since January 2025, with the duty to issue starting in 2027 for companies above €800,000 in turnover and in 2028 for everyone else. Belgium made structured e-invoicing mandatory for domestic B2B transactions in January 2026. Catching errors before an invoice leaves saves weeks of chasing. A rule I’d hold to: validate against the actual specification, and don’t claim compliance you haven’t had independently confirmed.
What happens next, and when? Timing, tone and channel are small choices made constantly, and this is where an agent earns its keep. Start in assisted mode. Let it draft, approve what it sends, and give it more freedom as its drafts start matching what you would have written. Trust builds in increments.
What did the customer just say? “Paying Friday.” “Never received it.” “That amount looks wrong.” Replies are where cash gets stuck, and one unanswered message can add weeks. Treat each reply as information (a promise, a dispute, a request for a document) and act on it. Real disagreements about money or the quality of the work should still go to a person, with the context already assembled.
When will the cash arrive? The due date is a term in a contract, not a prediction. A forecast built on how customers have actually behaved is far closer to reality, and it matters a great deal to an owner deciding whether they can cover payroll or make a hire.
Mistakes I’d avoid
- Automating a process you can’t describe. If you can’t explain your reminder rules in a few sentences, no agent can follow them.
- Sending the same message to everyone. It damages the relationships you’re trying to protect.
- Skipping the baseline. Record your days sales outstanding, the share of invoices paid late, average days late and the hours you spend chasing before changing anything. Otherwise you’ll never know whether it worked.
- Counting saved hours as a return. Time saved is only worth something once it’s spent on something else.
- Removing the human tone. The aim is to end the forgetting and the awkwardness, not the relationship.
Where I think this goes
My bet is that AR splits into two tiers over the next two or three years. At the top, enterprise platforms keep deepening what they offer to organizations with the people and systems to use it. Below them, a different kind of product gives small businesses a finance function they never had: one that decides, drafts, follows up and forecasts by default.
The lower tier is the bigger market by headcount, and the less well served. It won’t be won by the best model. It’ll be won by whoever makes the process disappear and trust easy to earn.
Late payment isn’t waiting for a smarter algorithm. It’s waiting for someone to make the decisions that currently go unmade.
Bishr Haffar is the CEO and founder of PayFlowAI and writes about accounts receivable, e-invoicing and applied AI for small businesses in Europe.
Sources
- Intrum, European Payment Report 2026 (April 2026; 8,385 businesses, 20 countries)
- Billtrust, Agentic VoIP launch (January 2026) and Collections update (July 2025)
- Germany e-invoicing timeline: ClearTax overview
- Belgium e-invoicing: TrueCommerce guide to the 2026 mandate
Original content: Medium.com