How Far AI Should Go in Commercial Collections

AI can already do most of what a collection call requires. Systems built for the job can place the call, respond to an excuse in real time, propose a payment plan, and follow up on a broken promise without a person touching the account. Some collection operations are being built entirely around that capability, research, contact, negotiation, and resolution running through a machine with a person stepping in only when something breaks the pattern. Whether AI can collect is no longer the interesting question. The one worth asking on any given account is whether the reason it is still unpaid is something more automation can actually fix, or something better execution alone may not solve. JSD builds around that second question, and it is why we keep judgment and debtor contact with a person rather than scaling them the way information scales.

The boundary that matters

Commercial collections runs on two different kinds of work, and they scale differently. One is gathering information, reading documents, checking a company's current legal and financial standing, tracking an account after it is placed. The other is deciding what to do with what was gathered, when to push, when to concede, whether an excuse holds up, whether to escalate. The first kind of work gets better the more processing power is thrown at it. The second does not scale in the same way.

JSD would automate as much of the first kind as the technology allows, including:

• Document extraction and reconciliation
• Contact and business status verification
• Public record and legal exposure research
• Account preparation ahead of the first call
• Ongoing monitoring after placement
• Timeline construction across invoices, deliveries, and signatures
• Discrepancy detection across documents

None of that is manual labor a collector should be doing when a machine can do it faster. The boundary shows up somewhere else. It appears the moment a system stops reporting what it found and starts deciding what to do about it. Reporting a discrepancy is automation. Deciding what the discrepancy is worth is a decision, and that decision stays with a person.

What document review earns back, at intake and later

Take document review first, since it tests what human-led actually means. A creditor places a batch of accounts, and each one arrives with a stack of paperwork nobody has fully read, invoices sitting against purchase orders that were never checked line by line, delivery receipts nobody matched to the right shipment. OCR reads through that pile automatically, matching invoice numbers and dates against the purchase orders and the placement list, and flags where the numbers disagree before a collector ever opens the file. Human-led does not mean a collector reading four hundred pages before making the first call. It means a person deciding what a flagged discrepancy is worth once the machine has already found it.

The same processing pays off again later, usually in a dispute. A debtor claims the goods never arrived. The creditor has a signed delivery receipt, but it is a photo a driver took years ago, buried in a folder of hundreds of PDFs left over from an old accounting system. Because OCR already ran against the full documentation set, that receipt turns up in seconds instead of a manual search through the folder, with the signature and delivery date already matched to the invoice.

Finding the receipt doesn't resolve the dispute. It means a collector, or later an attorney, already has it in hand when the argument starts.

Researching a debtor before the first call

Before a collector can have a useful conversation, they need to know who they are actually calling, whether the company is still active, and who currently controls payment, since that person is rarely who signed the original contract. Checking all of that one source at a time can run 20 to 40 minutes on a single account, and even then it usually stops at whatever a person had time to look up before the call.

The machine can know far more about an account than a collector could reasonably assemble by hand. A business status check against state filing records can run alongside a search for who currently holds the accounts payable or controller title, while a look at the legal record and a scan for financial distress signals run behind them at the same time. The outputs merge into a single briefing a collector reads in a couple of minutes, and the collector benefits from everything in it without having spent the morning finding it themselves.

The financial signals category is worth being careful about. A departing CFO or a lapsed domain is worth checking. Neither proves the company cannot pay, and a system that turns a signal like that into a conclusion about ability to pay is guessing, a guess that should never reach a collector unlabeled.

The limit disputed invoices expose

Take a quantity dispute. The purchase order authorizes 500 units, the invoice bills 550, and the signed delivery receipt shows 520 received. AI reads all three in seconds and reports the conflict, which matters more than it sounds, because a dispute that reaches an agency has often been open for months precisely because nobody on the creditor side ever put the three documents next to each other.

AI can put the purchase order, the invoice, and the delivery receipt side by side and show exactly where the numbers disagree. It can build a timeline of what was invoiced, delivered, and signed, and when. Determining legal liability is different work. Whether the extra fifty units are owed depends on whether the master agreement carries a delivery tolerance or whether a verbal change order preceded the shipment, and on which document governs when the terms conflict, questions that sit on a body of contract case law a language model has no standing to apply. Asked to answer them anyway, a model can produce a confident legal conclusion that has not been properly reviewed, which is worse than producing nothing.

The harder judgment has nothing to do with the documents themselves. Some disputes that reach an agency exist purely to buy time. A genuine dispute often appears earlier, has a documented basis, and stays relatively consistent when it is challenged. A tactical one more often shows up late and shifts its grounds as each version gets answered, though timing alone does not prove which is which; a real dispute can surface after months of silence, and a manufactured one can be raised on day one. Reading that pattern takes judgment about a specific customer and account, not a rule a document could contain.

AI earns its place on a disputed account before the conversation starts, making sure the conflict is visible and the timeline is already built by the time a collector picks up the phone. Processing the account and owning it are different claims. Reading three documents and reporting where they disagree is processing. Deciding what the disagreement means, and whether the customer believes their own argument, is owning the account, and that stays with a person regardless of how good the reading gets.

If AI can collect autonomously, why stop here?

An AI-first collection operation is a reasonable model for a lot of receivables, and not a strawman. A system like that can go live immediately, work every account around the clock, follow up on the same schedule every time, and carry a portfolio far larger than any team of collectors could touch. For a high volume of small, standardized balances, consistent follow-up matters more than the judgment behind any single call, and a model built to never let an account sit untouched has a real advantage there.

The tradeoff shows up as the accounts get less standardized, but the deeper reason JSD stops before full autonomy has less to do with account size than with what is actually broken on a given account. Automation is built to solve an activity problem. It is worth asking, on any account that has gone unpaid, whether an activity problem is what is actually happening.

A large share of nonpayment is exactly that. Follow-up is inconsistent. Contact information is stale. An invoice went to the wrong inbox and nobody resent it. The right AP contact was never reached. A promise to pay quietly expired and nobody noticed. The account is not unpaid because the customer refuses to pay it. It is unpaid because the process working it has gaps, and closing those gaps is precisely what a system built to never let an account sit untouched is good at.

Sometimes the account has already moved past that. The customer has the invoice, knows what is owed, has been contacted repeatedly, and has decided not to pay it, at least not yet and not without something changing first. Running the process again does not touch that kind of account, because the obstacle was never a lack of activity. What is left to weigh is usually something more efficient execution does not resolve, a genuine dispute against a tactical one, an inability to pay against an unwillingness to, a competing creditor, a relationship the debtor wants to preserve, or a settlement decision that depends on all of those at once. That is where the value of a person goes up.

Picture ten thousand small unpaid invoices for a subscription service against a single large account gone quiet for the same reason, nobody ever reached the right person, and the balance sat untouched. On both, the bottleneck really is activity, and automating it is the practical answer, probably the right one for a lot of the industry, regardless of which one has more zeros on it. Now picture that same large account, except the customer already has the invoice, already knows JSD wants it paid, and has decided to dispute a delivery term buried in a master agreement from years ago. The size did not change. What changed is that the question is no longer whether anyone followed up. It is why the customer is refusing and what the creditor should do next, and no version of faster follow-up answers that.

AI-driven collections and AR automation both attack the same target extremely well, making sure an account never sits untouched. Once a debtor has already been reached and decided not to pay, neither one solves the problem simply by executing the same process more efficiently. We make the same case from the AR side in how AR automation and commercial collections actually fit together.

Passive monitoring after placement

Research does not stop after the first call. The same information-gathering role continues after an account is opened, watching for a new lawsuit or a sudden change in who runs the company. Neither of those is a decision. Each is a change worth a collector's attention, surfaced automatically rather than discovered by accident three months later.

No collector is going to set up Google Alerts on every account in their portfolio and check them daily. A monitoring layer can watch hundreds of accounts continuously and surface only the changes that matter, freeing a collector to spend their attention deciding what a change means instead of hunting for whether one happened at all.

What the full workflow looks like

Put it together and the AI-assisted backend on a single account looks like this:

1. Account placed with supporting documentation.
2. Document processing extracts and reconciles the paperwork at intake, flagging discrepancies for human review.
3. Research runs at the same time, business status and current contacts alongside the legal record and any financial signals worth checking.
4. Outputs merge into a structured briefing well before a collector would have pulled one together by hand.
5. A collector reviews the briefing, corrects anything AI got wrong, and decides on first-contact strategy.
6. Collector makes contact. The conversation is human end-to-end, with no bot or auto-dialer standing in for a person and no AI-drafted letter going out unreviewed.
7. Monitoring continues in the background, flagging the collector when something material changes and leaving the interpretation to them.
8. If a dispute surfaces or the account moves toward escalation, the documentation and timeline are already built, ready for a person to interpret instead of rebuilt from scratch under time pressure.

Everything on that list reports what was found. Nothing on it decides what happens next, because that work never belonged to the machine.

What we do not delegate to AI

None of this is a claim about what AI can technically do. Systems already exist that can negotiate, draft outbound communication, and decide when to escalate an account without a person in the loop. At JSD, these are the functions we choose to keep off AI's plate, regardless of how capable the systems handling them become.

• Debtor communication. A negotiation carries the weight of an ongoing business relationship on both sides, and that is a call we want a person making, not a system executing a script well.
• Unreviewed outbound letters. The last check on what leaves the building under a client's name should always be a person, not a policy about how good the draft usually is.
• Escalation decisions. Legal referral is a call about what the relationship and the recovery are both worth, and we keep that with a person who can weigh both at once.
• Dispute resolution. AI surfaces the documentation on a disputed account. Deciding whether the dispute is real, and what to do about it, stays with someone who can read the account's history, not just its paperwork.
• Settlement authority. What a creditor will accept, and what a debtor can realistically be held to, is a negotiation JSD keeps with a person on both sides of it.

AI owns the information a collector works from. People own every decision made with it, and people are the only ones who ever talk to a debtor. JSD chose that boundary. Every one of these systems could clear that bar eventually, and that is exactly why the line gets drawn now instead of waiting to find out.

Where JSD sits on this

We are not a software company and this is not a pitch for an AI product. Some of what is described above already runs in our process in some form. Some of it describes where the technology is heading and how far we expect to take it. Either way, the boundary comes first and the tool comes second. We want our own collectors surrounded by as much useful automation and information as the technology can provide, and we are not opposed to agencies automating far more of the process than we do, especially on a large volume of standardized receivables.

The automation stops at the point where information becomes a decision, or where a decision becomes contact with a debtor. The future of commercial collections may end up heavily automated either way. What decides it is whether AI supports the collector or becomes the collector, and JSD is built around the first answer.

The line separating information from judgment in how JSD approaches B2B collections is a choice JSD made, and it will hold as the technology gets better rather than loosen. If you have past-due B2B receivables and want a person working the account from the first call, place it directly.