How an AI Agent Works a Reconciliation Break

Reconciliation is a useful process for seeing how an agent works, because most of the job is already automated and the part that isn't is easy to describe. Software matches the bulk of settlement lines to the bank and the ledger on amount and reference. The time goes on what's left over, called breaks. Each break has to be traced to a cause, valued and fixed, and that means opening several systems.

If you're new to the terms model, copilot and agent, the first piece in this series explains them.

The break

The example is a UK restaurant group that takes card payments through one acquirer. Each day the acquirer pays the group its card sales for the day, minus the acquirer's fees and minus any chargebacks. A chargeback is a payment a cardholder has disputed, which the acquirer takes back from the merchant. The group's ledger records the day's sales and the fee the contract says the acquirer should charge, so the ledger knows how much cash to expect.

For 12 October, the matching software flags that the acquirer paid £1,240 less than the ledger expects. Worked by hand, an analyst downloads the settlement file, searches the ledger, pulls up the contract and may end up emailing the acquirer. Here's what an agent does with the same case.

1. Read the break

The matching software puts the case in the agent's queue. The model reads it: which acquirer, which date, how much is missing.

2. Decide what to look up

The model picks a tool and the inputs to give it. The tools are small pieces of code the firm wrote: fetch the settlement lines, search the ledger, read the fee schedule in the contract, get the bank line. Each tool answers one kind of request, and the model is told what each tool does and which inputs it needs.

The model doesn't send the tool a sentence. It sends a short structured request that names the tool and fills in its inputs, such as get_settlement_lines with batch 4471. Code runs the request and passes the result back to the model. The model chooses which tools to call and in what order.

3. Read what comes back and pick the next step

The settlement file shows two things the ledger doesn't have. The acquirer took back three chargebacks totalling £1,050, and a ledger search finds no entry for them. The fee is also higher than expected. The model reads the contract, which sets the fee at 0.25% of sales, and sees the acquirer charged 0.30%. Code works out the difference: 0.05% on £380,000 of sales is £190. The chargebacks and the extra fee together make up the shortfall.

The gap is £1,240: £1,050 of chargebacks the ledger doesn't know about, plus £190 of fee above the contract rate.

4. Explain the cause and the amount

The model writes up the break in plain language. Each number points back to the record it came from, so a reviewer can check the explanation in a minute or two.

5. Draft the fix and queue it

The agent drafts a journal to book the three chargebacks and a £190 fee claim to the acquirer. Both go into an approval queue, and nothing is posted or sent until a person approves it.

Steps 2 and 3 repeat until the model has causes that add up to the break. A simple break might need two or three lookups, and a messy one ten or more.

What the model did and what code did

The model did the reading and the deciding. It read a settlement file laid out the way this acquirer lays it out, linked the fee line to the clause in the contract that sets the rate, chose what to look up next and wrote the explanation. Rules struggle with this part, because every provider's file and every contract is a bit different.

ScalePoint Research plans its model tests around the skills this kind of work needs. Reading the settlement file is extraction, checking the fee against the contract is rule application, and tying the file, the ledger and the contract together is reasoning across sources. These are three of the nine skills for testing AI models in payments, and the method page shows which test covers which skill.

Code did the line matching, the sums, the data access and the limits. The matching software paired the settlement lines with the bank and the ledger before the agent started, so the agent only got the lines that didn't match. The £190 was calculated in code, because models still make arithmetic mistakes. The model only saw what the tools returned. And the tools were read-only, so the agent couldn't post the journal or send the claim itself.

Where a person comes in

The approval queue is where the firm keeps control. A reviewer sees the break, the explanation, the records behind each number and the drafted fix, and approves, edits or rejects it. Every lookup the agent made is logged, so an auditor can follow the case later.

This is still an agent, even with a person approving each fix. The difference from a copilot is who does the investigating. With a copilot, the analyst downloads the file, searches the ledger and reads the contract, and the model suggests a likely cause along the way. With an agent, the agent does all of that and the analyst checks finished work. Once the results are good enough, the firm can also let small fixes through without approval. What changes when an AI agent works the queue covers how to set those approval points.

The firm sets these limits, and the model can't change them. Hence an agent can be given a lot of room to investigate, across files, ledgers and contracts, while money only moves within limits the firm has approved.

What drives the cost

An agent calls the model once for each lookup and again for the write-up, and every call reads the case so far. A break that takes eight lookups costs several times as much in model fees as one that takes one or two.

Model fees are only part of the cost. Two checks catch an agent's mistakes, and both cost more when the model is weak.

• Checks in code. The causes the agent finds have to add up to the break. If the chargebacks and the fee came to £1,100 against a £1,240 break, code would send the case back to the model to keep looking, and pass it to a person after a set number of attempts. Each extra round is another set of model calls.
• The reviewer. A person reads every explanation before approving the fix. A weaker model writes more explanations that are wrong or hard to follow, so reviewers spend longer per case and reject more, and each rejected case goes back to an analyst. Reviewer time often costs more than the model fees.

The most expensive case is a wrong explanation that a reviewer approves, because money is then booked or claimed wrongly. Hence the full cost per case is the model fees, plus reviewer time, plus the cost of mistakes that get through. On that basis a cheap model can come out dearer than an expensive one. The unit economics of AI in payments and finance operations works through the full cost per case with figures from the tests.

ScalePoint Research tests models on payments tasks like this one, and scores each model on the money it gets wrong and what it costs to run. It's starting with a first round of 11 one-step tests, where the model gets all the files at once and no tools, and plans to test agents end to end after that. Test 2 is the closest so far to this break: nine models reconciled a month of card and payout settlements for an invented merchant group. The best model missed about $13 a run, and most cheaper models missed hundreds of dollars. The tests page has the latest results.

Evidence so far

Each ScalePoint Research test adds a row to this table, with what that test shows for the point of this article.

Where reconciliation agents are today

Several reconciliation vendors started selling agents in 2026. BlackLine made its Verity Prepare agent generally available in July 2026, and Trintech launched an exception management agent on 29 September 2026. Others offer copilots that suggest causes for a person to check. ScalePoint Research plans to review reconciliation vendors in a series of articles, starting with BlackLine and Trintech. Sign up on the Research page to get an email when they're out.