A series exploring how AI agents change the daily reality of finance teams. Not theory: working systems you can try.
The overnight problem
Between the time your team logs off on Friday and logs back in on Monday, overdue balances keep aging. Thresholds get crossed. Customers that were at 29 days hit 32. The €200K account that needed a call yesterday is now three days further from resolution.
The AR Collections Agent closes that gap. It runs in continuous cycles, scanning every account, comparing against the previous cycle, and acting the moment something changes. A payment comes in at 2 AM? It de-escalates automatically. An invoice crosses the 30-day mark on a Sunday? The email is drafted and ready.
How it makes decisions
The agent follows your collection policy, but it doesn't follow it blindly.
These are the defaults. But the agent evaluates the full picture: amount at risk, number of open invoices, payment history, prior actions, and assigns a risk level per customer. A €150K invoice at 12 days with no payment history? That might trigger an early email even though the policy says "monitor." The agent explains why it deviated, every time.
Policy is the baseline. Risk assessment is the intelligence layer. Every deviation comes with reasoning.
What it actually does for your team
This isn't a dashboard that shows you data and waits for you to act. The agent acts:
- Drafts collection emails that escalate in tone: friendly reminder at 15 days, firm at 30, urgent at 60+. Ready for your team to review or send directly.
- Generates voice call scripts with text-to-speech, so your collectors have a prepared, consistent message for each account.
- Detects payments and de-escalates, no more sending a collection email to a customer who paid yesterday.
- Moves customers through a pipeline: not started, email sent, voice call, escalated, with full history at every stage.
- Produces executive summaries after every cycle: what changed, who was contacted, what's at risk.
The audit question
The first thing any CFO will ask: "Can I trace what it did and why?"
Yes. Every cycle produces a complete log: which customers were assessed, what risk level was assigned, what action was taken, and the reasoning behind each decision. If an auditor asks why a specific customer was escalated on a specific date, the answer is there. Not a summary, the actual reasoning chain the agent used at that moment.
The system also includes a Q&A assistant where anyone on the finance team can ask questions in natural language. "What's the status of Mediterranean Foods?" "Which accounts crossed 60 days this week?" "What's our total exposure above €100K?" No pivot tables. No report builder. Just ask.
Why local inference
The agent runs dozens of cycles per day, each one triggering multiple LLM calls: risk assessments, email drafts, call scripts, summaries. At that volume, cloud API costs add up fast. Running the model on local hardware makes the economics work, and as a bonus, sensitive financial data never leaves the network.
That said, this is a design choice for this specific use case. Nothing stops you from pointing the same architecture at a state-of-the-art cloud model if latency and cost aren't a concern. The agent doesn't care where the intelligence comes from, it just needs structured output it can act on.
Beyond collections
The pattern behind this agent, scan, assess, act, log, repeat, is not specific to AR collections. It's a loop that maps onto any finance process where you're monitoring state, evaluating against policy, and taking action based on thresholds:
Cash flow forecasting. Budget variance monitoring. Vendor payment optimization. Intercompany reconciliation. Audit trail analysis. Each of these follows the same cycle, and each could run with the same agent architecture.
Collections is where I started. It won't be the last.
See it work
The agent is running live with a demo dataset. You can watch it cycle through accounts, read the emails it drafts, see the risk assessments, and explore the Q&A assistant.