Why your AI should not be the calculator
Language models write well and count badly. In credit analysis, the fix is architectural: code computes, the model explains.
By the Maple Crew team · · 5 min read
Ask a language model for a debt ratio and you will usually get the right figure. Usually is the problem. In a credit file, a ratio that is wrong one time in a hundred is not a small error rate; it is a decision made on a number nobody can defend.
When we designed Maple Insight, our credit scoring platform, we took one rule as non-negotiable: the model is never the calculator.
Two jobs, two tools
A credit assessment contains two different kinds of work. The first is arithmetic and rules: recognised income is the median of the last six salary deposits; the debt ratio is repayments divided by income; a ratio under 20 % earns full marks. This work has exactly one right answer, and ordinary code gets it right every time.
The second is language: explaining why a criterion scored as it did, summarising a file for a credit committee, pointing out what deserves a second look. This is where a language model is useful.
Problems begin when one tool does both jobs.
How the split works in practice
- Every figure is produced by deterministic code with unit tests. The same file always gives the same score.
- Each specialist agent receives the facts and calls calculation tools for any number it needs. It does not do mental arithmetic.
- After an agent answers, the platform applies the rule again in code. If the agent's points differ, the rule wins and the disagreement is recorded on the finding.
- The credit note is drafted from the computed figures only. The analyst edits and validates it before it goes anywhere.
What you gain
The first gain is reproducibility. An auditor can take a file, rerun it and get the same result. The second is explainability: each point on the scorecard traces back to a rule and to the source data behind it. The third is trust. Analysts adopt a tool when they can check its work, and they can only check work that follows rules they can read.
What the model still does well
None of this makes the model decorative. It reads documents that arrive as scans, phrases findings in clear language, and drafts a note in a minute that would take an analyst an hour. It just does so on top of numbers it did not invent. That division of labour, more than any choice of model, is what makes an AI system fit for regulated work.