On this page
Language can make analytical tools easier to approach. It should not be the mechanism that decides which numerical result is true.
The difficulty hidden by a simple question
A business user should be able to ask a question in ordinary language. That does not mean the language itself is a sufficient analytical specification.
“Which product performed best?” might refer to revenue, units sold, margin, growth or customer preference. A question about change may depend on the selected periods and whether the values are comparable. Even where a calculation can be carried out precisely, the result is only useful when the system has correctly understood which calculation was intended.
Generative AI is valuable in this situation because it can help work with unfamiliar expressions, terminology and conversational ambiguity. The difficulty arises if the model is also allowed to calculate the answer or choose its own interpretation without a separately checkable contract.
The original determinIQ research question was therefore not whether AI should be used. It was where AI’s authority should end.
A boundary between intention and execution
The architecture places a governed question contract between a person’s request and the calculation engine. In broad terms, this contract establishes what analysis is being requested, which data fields are relevant, which operation is supported and which constraints must hold before execution.
The contract is not itself the calculation. Nor is it a free-form transcript of the user’s words. It is the point at which an open-ended expression must become a controlled analytical request.
If that conversion cannot be completed reliably, the correct next step may be a clarification or a refusal. An unsupported request should not be passed down to a general language model in the hope that it will invent a plausible substitute. The ability to say “I cannot establish that from the available data” is part of useful behaviour.
This also makes responsibility clearer. The intent layer concerns interpretation. Validation concerns whether the request can proceed. Deterministic execution concerns the numerical answer. Evidence-related components record what supports the result. Presentation and narrative concern how the result is communicated.
A system that allows all these functions to modify one another is much harder to test. A system with explicit boundaries can make stronger, more specific claims about what each component does.
The calculation is not the explanation
Once a supported question has been prepared and executed, the result may still need explanation. The same numbers can be introduced differently to a finance director, an operational manager or a less specialist reader. The language may change, but the underlying result must not.
This is particularly important when a user asks a follow-up question. Conversational continuity is useful, yet a follow-up should not become permission for the narrative layer to revise an earlier calculation silently. New analytical intent requires its own governed path. An explanatory sentence cannot create a new analytical fact.
Separating these responsibilities makes it possible to use AI with considerable ambition for language tasks while treating the generated wording with appropriate caution. The explanation can be helpful without being authoritative over the arithmetic.
What has been built and what it proves
The current determinIQ implementation has a governed preparation and execution architecture, with a distinct calculation authority and supporting evidence and presentation paths. This is more substantial than a prompt instructing an AI not to make mistakes. The source architecture embodies the separation as a software responsibility boundary.
The implementation alone does not establish that every natural-language request is interpreted correctly. Ambiguity, schema classification, field meanings and unsupported business terms still create failure modes that must be tested. A well-designed boundary improves the conditions for correctness and auditability; it is not an unconditional guarantee.
This is why the testing method matters. A plausible rendered answer is insufficient. The interpreted request, permitted operation, calculation and evidence need to be checked against a known expectation, including when the system should refuse to proceed.
A useful consequence
The decision to limit AI’s authority has made it easier to think about the rest of the product. Guided exploration, proactive findings, comparison and narrative presentation can be developed as distinct ways of approaching a common analytical process rather than as separate sources of truth.
This does not make the system less capable. It is an attempt to make capability accountable. Language helps people ask and understand; the governed engine establishes the supported calculation; people remain responsible for judgement.
The broader research programme examines how this separation connects to privacy, evidence retention and the overall user experience. It is the combination of these responsibilities, rather than any single clever prompt, that I have been testing through determinIQ.
Research context
Research programme: Making data analysis understandable, reproducible and accountable
Topic: Evidence and verification
Explore further
Supporting material is listed below.
Further reading is listed below.