AI employee for collections

What is collections work when money is late?

Someone owes an amount that is already on the books. The early work is a reminder, a date, and a way to pay. The later work is judgment: hardship, dispute, settlement, legal. An AI employee is a fit for the early work if the ledger is clean. It is a poor fit for the later work.

I have sat on systems that moved money. At 10mg we have put out $3.4 million in loans to more than 6,000 providers. If a row is wrong, a clinic is in the wrong conversation. Collections on a wrong row is harassment dressed as ops. Fix the ledger first.

Write the schedule you already use. Day 3, day 7, day 14, or whatever your policy is. The employee follows that schedule. It does not invent a fourth channel because the model wants to be helpful.

What does a good collections employee look like?

The message names the amount, the due date, and the pay path that already exists. The tone is short and plain. No shame. No fake urgency. The record after a reply is a state: promised, disputed, asked for a person, hardship. A collector can open the account and know what happened.

Good also means a stop that sticks. If the person said stop, the next send does not go out. If that requires a flag in the system of record, build the flag before the first campaign. A model that “forgot” is not an excuse.

I use the same test I use for support. Can someone who did not write the workflow explain the last touch from the log? If no, you have a script in a blender. That will not survive a complaint.

• Amount and date come from the ledger, not the prompt
• Every reply writes a state, not a paragraph
• Stop flags block the next send
• Pay links are the ones you already operate

When do you stop the machine?

When they ask for a person. When they dispute the debt. When they describe hardship. When they mention counsel. When they are distressed. When your policy or the law says you have touched them enough. Those are hard stops, not suggestions.

Escalation is a state handoff. The collector gets the account, the amount, the last touches, the words that triggered the stop, and the reason. A chat export is not enough. The person who picks this up may be walking into anger. Give them the state.

A 10–15% escalate target is a useful shape once the early reminders work. Early collections should mostly close or pay. If you are escalating half of them, the ledger or the timing is wrong. If you escalate none, you are probably ignoring distress.

What should you not automate first?

Do not automate threats. Do not automate legal letters. Do not automate a new payment plan or a settlement number. Do not let it imply a consequence you will not actually take. Those are how you get a regulator and a story you cannot defend.

Do not start with the oldest, angriest book. Start with the early, undisputed balances that already get a template. If you cannot name that bucket, you are not ready.

Skip voice for collections as a first channel. Text or email with a written record is enough. Voice is easy to get wrong and hard to take back. I will not put a night voice on a late bill as a first install.

How do you measure collections without being cruel?

Cash recovered on the set the employee touched. Promise-kept rate. Time to a named collector after a stop. Complaint rate. Unsubscribe or stop-contact rate. Put recovery and complaints on the same page. If you only show cash, someone will turn the screws.

Sample the sends. Is the amount right? Is the person already on a plan? Did we write stop and then send anyway? A wrong amount is a reliability bug. I treat it like a wrong disbursement. The slide about “AI collections” will not help in that review.

If recovery rises and complaints rise faster, you failed. If both stay flat, you sent mail. If recovery rises and complaints stay flat, you may have earned the next bucket. Earn it. Do not jump to legal tone.

Why is money a hard gate here?

Because the employee can hurt someone and still post a win on a dashboard. Money, legal, and medical are the three gates I will not soften. Collections sits on the first two at once. That is why the first job is a reminder, not a chase.

At SmartComply we kept a person on the miss. Same here. The model can draft the day-7 note. A person writes off. A person approves a plan. A person sends anything that sounds like a consequence.

When we sold Insurpass and Shopl, diligence was cost, uptime, and whether we could explain the books. A collections employee that cannot explain a send is a liability in that meeting. Leave the log. Leave the stop. Leave a way to turn it off.

What does a state handoff look like when someone is angry?

Short. Account. Amount. Last three touches. The sentence that stopped the machine. What they asked for. What you already promised. The collector should not have to ask the person to retell the bill. Retelling a bill when you are already late is how you lose the room.

Do not transfer into a hold with no owner. Night or day, the next person is a name or a queue that is staffed. If you cannot staff it, say when you will call. Miss that time and you should not have offered it.

This is the same handoff I want in support, voice, and nights. Collections just has less room to be sloppy. Get the state right. Then let a person do the human part.