Direct answer: Trust reconciliation totals should be calculated deterministically from explicit source records using decimal-safe or integer arithmetic. AI may assist with narrow tasks such as suggesting a transaction category or summarizing an exception, but its output should be labeled, reviewable, and unable to approve, move money, alter source evidence, or turn an unresolved legal judgment into a clean close.
Key takeaways
- Use deterministic math for balances, differences, and report generation.
- Treat AI classification as a suggestion with confidence and provenance.
- Require human review for uncertain or consequential classifications.
- Measure classifier performance and preserve overrides.
What deterministic means here
The same complete source bundle and rules should produce the same reconciliation result. Currency calculations should avoid binary floating-point drift. Matching rules and cutoff logic should be testable, versioned, and explainable.
Determinism does not guarantee that source data is correct. It guarantees that the calculation is reproducible, which makes source errors and rule changes easier to investigate.
A narrow role for AI
Unfamiliar descriptions and messy exports can create classification work. An AI model may propose that a transaction is a bank fee, transfer, deposit, or unknown item, but the system should retain the source text, proposed label, confidence, model or rule version, and reviewer decision.
Low-confidence or high-impact cases should default to unresolved. The product should never fabricate a client mapping just to make the totals agree.
Control the model, not only the prompt
Test against labeled examples, track override rates, monitor changes after model updates, and provide a complete non-AI fallback for calculation. Keep sensitive data minimization and vendor handling terms in scope.
TriBalance's product boundary is explicit: deterministic reconciliation, assisted classification for edge cases, and human approval. AI does not make legal judgments or finalize the close.
Frequently asked questions
Should an LLM calculate the three-way totals?
No. Exact balances and differences are better handled by deterministic code and decimal-safe or integer arithmetic.
Can AI automatically resolve an exception?
It may suggest a classification or explanation, but a controlled workflow should require appropriate human review and preserve the decision evidence.
Sources and further reading
This article is operational education, not legal advice. Trust-account rules and retention requirements vary by jurisdiction. Confirm the requirements that apply to your firm and accounts.