AI agents · Explainer · October 2026
Rules, Machine Learning or a Model: Who Does What in Payments Operations
Most payments and finance processes already run on rules, and some on machine learning. A language model earns its place in the steps those two leave. Here's how that splits across the 51 processes on the ScalePoint Research process map.
- Rules do the exact work: matching on amount and reference, fee and FX sums, limits, time limits and routing tables. They cost almost nothing per transaction and give the same answer every time.
- Machine learning scores and ranks at volume: fraud scores, alert scores, likely matches when references are messy, and credit risk.
- A language model reads files and contracts in whatever layout they come in, works the cases that rules and scores leave behind, and writes them up for a person to approve.
- Hence the useful question for each process is which step needs a model, and what that step costs per case.
When a vendor pitches an AI agent for reconciliation or disputes, it's easy to picture the model doing the whole job. In practice the model sits on top of two older layers that already do most of the work, and the choice of model only matters for the part they leave.
This piece uses three terms. Rules means instructions someone wrote, such as "if the amount and reference match, pair the lines". Machine learning means a model trained on past data to give a score or a ranking. A language model is the kind of AI that ScalePoint Research tests, such as Claude, GPT or Gemini. Machine learning and language models are both AI, and rules are not.
What each layer is good at
Rules
Rules handle anything that can be written down exactly: pairing a settlement line with a bank receipt on amount, date and reference, working out the fee a contract says should have been charged, checking a chargeback is inside the scheme's time limit, or sending a payment down a set route. They're cheap to run, easy to audit and give the same answer every time. They break when the input doesn't look the way the rule expects, such as a missing reference, a new file layout from an acquirer or a contract clause nobody coded.
Machine learning
Machine learning handles volume and fuzziness that rules can't, such as scoring every card payment for fraud in a few milliseconds or suggesting which invoice a payment with a garbled reference probably belongs to. It gives a number without a reason. It also needs a good history of labelled cases to learn from and someone watching for drift as patterns change.
A language model
A language model reads text and tables it hasn't seen before, follows a written policy and explains what it found. That makes it good at the leftovers: a settlement break that needs the file, the ledger and the contract read together, or a dispute where the evidence is a mix of emails, receipts and scheme rules. It costs far more per call than a rule, it can be wrong with confidence, and it shouldn't do arithmetic in its head. Hence code does the sums and a person approves anything that moves money. How an AI agent works a reconciliation break walks through one case in detail.
How the split plays out across the map
The ScalePoint Research process map lists 51 payments and finance processes in 11 clusters and rates each one for agent fit: 22 as agent workflow, 14 as agent-assisted and 15 as single step. The ratings follow how much work is left once rules and machine learning have done their part. The table shows the typical split for the main clusters.
| Cluster | Rules | Machine learning | Language model |
|---|---|---|---|
| Reconciliation (10 processes) | Match on amount, date and reference. Work out expected fees from the contract. | Suggest likely matches when references are missing or garbled, as in cash application. | Work each break: read the files, find the cause, value it and draft the fix. |
| Disputes, claims and complaints (6) | Reason codes, scheme time limits, refund thresholds. | Score claims for likely fraud, such as APP scam claims. | Read the evidence and draft the response or decision letter. |
| Alert review (6) | Sanctions list checks and monitoring scenarios. | Score alerts and match names that are spelt differently. | Gather the evidence and draft the case note for an analyst. |
| Onboarding and KYC (7) | Check documents and data against policy. | Credit and risk scores. | Read company documents, list what's missing and summarise the case. |
| Statement and contract reading (4) | Check totals and compare rates with the contract. | Little needed. | Read statements and contracts in whatever layout they come in. |
| Pricing, treasury and payment performance (9) | Pricing tables, billing runs, cash positions, routing rules. | Forecasts, decline analysis and routing choices. | Single steps, such as explaining a decline pattern or drafting a deal summary. |
Where the work ends in a written case for a person to decide, the map rates it agent workflow, because the model has to look several things up before it can write the case. Where the work is mostly a calculation or a decision from a table, rules and machine learning cover it, and a model adds a single step at most. Pricing, treasury and payment performance sit there.
Where the cost sits
The model's cost lands only on what's left after the other two layers. In reconciliation that's the breaks, a small share of all the lines. Hence the number to compare is the cost per case on that step, against what a wrong answer costs.
The two tests so far show why that matters. In Test 1, reading card statements, GLM 5.3 got 0.48% of fees wrong at 1.4 cents a statement, against Claude Opus 5.5 at 0% and 15.5 cents. For a job that is mostly extraction, the cheaper model was the better buy. In Test 2, reconciling a month of settlements, GLM 5.3 cost 65 cents a run less than Opus and missed $1,381 more, on files with 17 planted errors. On a step that needs judgement, the cheaper model's misses far outweighed the saving. Nine skills for testing AI models sets out which skills each test asks for.
The same logic works the other way. Putting a model on work that rules already do well adds cost and a new way to be wrong.
What the tests measure
Each ScalePoint Research test takes one task from the map and gives every model the same prompt and files. The aim is to test the step a model would actually do. Test 1 did that: reading statements is the model's job in that cluster. Test 2 asked models to reconcile a whole month in one reply, matching included, which is a harder version of the job than a model gets in practice. ScalePoint Research is building a version that tests the breaks only, with the matching done first.
Questions to ask before you buy or build
- Which steps does the product do with rules, which with machine learning, and which with a language model?
- For the model steps, which model is it, and what does it cost per case at our volumes?
- How much would simple rules catch on their own, before any model is involved?
- Who approves what the model writes before money moves or a customer hears from us?
This is one of five pieces on AI agents in finance and payments operations. The series, and the research map behind it:
- AI models, copilots and agents: what's the difference
- How an AI agent works a reconciliation break
- Rules, machine learning or a model: who does what in payments operations
- Will AI agents be standard in finance operations by 2031?
- What changes when an AI agent works the queue
- The process map: 51 payments and finance processes rated for agent fit