When an LLM should not do the maths
A practical design rule for systems that need both exact calculation and useful generated explanation.
Fluency is not a numerical contract
A language model is useful when a product needs to interpret intent, restructure information or explain a result in accessible language. Those strengths do not make it the right place to own exact business calculations.
If a number must be repeatable, testable and auditable, it should come from deterministic code or another controlled computational system. The model can describe the result; it should not quietly redefine it.
Give each component one clear job
A dependable pattern is to calculate first, assemble a constrained evidence package, and only then ask a model to produce narrative. The generated response can be checked against the same evidence package before it reaches the user.
- Calculation layer: produces canonical values.
- Evidence layer: carries values, units, assumptions and provenance.
- Generation layer: turns supplied evidence into appropriate language.
- Validation layer: rejects unsupported or inconsistent claims.
The wider principle
This is not only about arithmetic. It is about assigning authority. In any AI-assisted system, the team should be able to say which component owns a fact, which component may transform it, and what happens when the two disagree.
Clear ownership makes testing easier, provider changes safer and failures more legible. It also keeps the most persuasive part of the interface—the prose—from becoming the least accountable part of the system.