ScalePoint Research · Process map · Cluster

Reconciliation

Matching what should have arrived with what did, and working out why the two differ. Ten processes on the map come down to the same job, from a merchant's card income to an EMI's safeguarding account and a stablecoin firm's on- and off-ramps.

Vendor scan: Next Model test: Next Agent fit: Agent workflow 10 processes · 6 types of firm

The processes in this cluster

Click a process to see what the work involves.

Card settlementMerchant

Match the acquirer's settlement files to the money received in the bank and the sales in the ledger, and explain each difference.

Scheme settlementAcquirer · Issuer

Check the card schemes' settlement reports against the firm's own clearing records, including the interchange and scheme fees charged.

APM settlementMerchant

The same job for payment methods other than cards, such as wallets, bank transfers, buy now pay later and local methods, where references and timing differ from cards.

Cash applicationMerchant

Match payments received from customers to the invoices they pay.

Merchant payoutsPlatform · Acquirer

Check that each payout to a merchant equals its transactions less fees, refunds, chargebacks and reserve movements.

Split payoutsPlatform

Check that each payment split between sellers, the platform's commission and fees adds up, and that every seller was paid the right amount.

SafeguardingEMI

Each day, check that the money held in safeguarded accounts covers what is owed to customers, and explain any shortfall.

Nostro accountsEMI · Cross-border

Match accounts held at banks in several currencies against the firm's own ledger, including the FX conversions between them.

On- and off-rampsStablecoin

Match money coming in and going out in normal currency against the stablecoins bought, sold or sent on the blockchain.

Stablecoin reservesStablecoin

Check that the reserves backing a stablecoin match the number of coins in circulation.

The work, step by step

Reconciliation has five steps, and AI helps with some more than others. Matching records (step 02) is already done well by software rules. Most of the time goes on the differences left after matching, called breaks. Finding and valuing the breaks takes one model call. Tracing each break to its cause takes several steps across different systems, which is where an AI agent helps.

StepTodayWith AIFit
01
Collect the records
Settlement files, bank lines and the ledger for the period.
Today
A person downloads files from provider portals, or a scheduled export drops them in.
With AI
Little to add. Scheduled exports and APIs already do this. AI only helps where records arrive as PDFs or emails that need reading.
Mostly no AI needed
02
Match
Pair up the records that agree, by amount, date and reference.
Today
A matching engine, or a spreadsheet at smaller firms.
With AI
Nothing to add. Rules do this well and cost almost nothing per run.
No AI needed
03
Find and value the breaks
Spot every difference left over and put a number on it, ignoring timing differences that will clear.
Today
A person works through the unmatched items.
With AI
One model call over the unmatched items. We'll be testing this step in an AI model benchmark, coming soon. About the test
Single step, tested
04
Investigate and explain
Trace each break to its cause, such as a missing settlement, a fee at the wrong rate or a chargeback debited without a case.
Today
A person looks each break up across two or three systems.
With AI
An agent chases each break through the records until it has a cause, and writes it up.
Agent workflow
05
Adjust or escalate
Post the adjustment, or raise a claim with the provider.
Today
A person drafts and posts the journal, or emails the provider.
With AI
An agent drafts the journal or the claim. A person approves before anything is posted or sent.
Agent-assisted

Note: Step 03 is shaded because it's the one step we test AI models on. The other steps are covered by the vendor scan.

Agent several steps on its own, checked at the end Agent-assisted drafts, a person decides Single step one model call No AI needed rules already do it

Why we rate reconciliation as agent work

Every process on the process map is rated against the same five tests. Reconciliation passes all five, which is why it's rated Agent workflow.

Several steps across systemsYes

Each break needs a look in two or three places: the settlement file, the bank line, the ledger and sometimes the contract.

Many similar casesYes

Breaks turn up every day and most fall into a few recurring types: timing, fees, refunds, chargebacks and missing settlements.

A result you can checkYes

The month balances, and every break has a cause and an amount. Either the agent got there or it didn't.

Leftovers that rules missYes

Matching rules (step 02) clear the records that agree. Working out why the rest don't agree (step 04) takes judgement, which rules can't do.

Mistakes can be caught before money movesYes

AI agents still make mistakes, so a process suits them better when a person can check the result before it costs money. Here the agent proposes adjustments and claims, and a person approves them before anything is posted. Little in reconciliation is real time, so that check doesn't slow things down.

The rating holds across the cluster with two exceptions. Cash application, matching customer payments to the invoices they pay, is rated Agent-assisted. Existing software already matches most payments on amount and invoice number, so an agent only helps with the few that don't match. Stablecoin reserve checks are a single step: two data sets compared once a period.

Key findings

Coming soon

Two or three findings from the vendor scan and the model test, such as where single products stop for each type of firm and which steps firms are still left to build themselves.

Vendor landscape

Coming soon

The scan covers every vendor selling reconciliation into any of the ten processes: reconciliation platforms, AI-first start-ups, modules in ERP and finance suites, and features built into PSP and acquirer platforms. For each vendor it records which processes and which steps the product covers, and how far its AI goes:

None: no AI in the product.Copilot: the AI suggests, and a person does the work.Agent in production: the AI takes a break from start to finish, from spotting it to proposing the fix, and paying customers already use it.

No vendor pays to be included. If you sell reconciliation software and want to be covered, email me at martin@scalepointpartners.com.

Model test

Coming soon

A month of an invented merchant's settlement files, bank lines and ledger in one go. Each model has to find and value every break. Models are scored on money missed and false alarms, against price per run.

How the testing works →

Related clusters

Disputes and claims

Chargebacks debited without a case show up as reconciliation breaks first.

Payment exceptions and tracing

The same investigate-and-explain step, on failed and missing payments.

Statement and contract reading

Fees at the wrong rate are found by reading the statement against the contract. Read the paper →

I can run the same scan and test on your own reconciliation: your files, your volumes and the vendors you're talking to.

Email me