On this page
An independent investigation into whether natural-language interaction and automated analytical discovery can make complex data easier to work with while keeping calculations, evidence and decisions under explicit human and system control.
The question behind the work
I began determinIQ with a question that seemed simple enough: could generative AI make useful business analysis more accessible without requiring people to trust an AI model to decide what the data says?
I had been using language models extensively in practical development work. Their usefulness was clear. They could help translate a rough idea into a more structured question, make unfamiliar material easier to approach and accelerate an extraordinary amount of detailed work. But that usefulness brought a difficulty into focus. A fluent answer is not necessarily a correct one. An answer can be largely correct while resting on a single assumption the person asking never made. If that answer concerns money, performance or risk, the ambiguity matters.
Traditional business intelligence has a different problem. It provides substantial analytical power, but often expects people to know the structure of their data, choose the right measures and construct the right analysis before they can learn anything from it. That is a reasonable expectation of an analyst. It is less reasonable for someone whose expertise lies elsewhere and who wants to understand what is happening in their organisation.
The research was therefore not about replacing established analytical methods with a conversational interface. It was about finding out whether some of the complexity could be carried by the system without losing the controls that make a result worth using.
The architectural choice: separate language from analytical authority
The most consequential early decision was to give the analytical engine authority over calculations and leave AI with a narrower role. Language interpretation can help establish what someone intends to ask. Narrative can make an established result easier to understand. Neither should be permitted to invent or alter the underlying calculation.
In the implemented architecture, a question does not travel directly from a text box into an unrestricted answer generator. It is prepared as a governed analytical request, checked for support and validity, and passed to deterministic execution only when the relevant constraints are satisfied. The engine is responsible for calculating the result; the explanatory layer works from governed material that follows.
Research figure 01
AI interprets; the core calculates
A simplified public model of analytical authority
- HUMAN INTENTA person asks a questionOrdinary language expresses what they want to know
- LANGUAGE SUPPORTInterpret and prepareAI may assist; intent is not executable authority
- GOVERNED CONTRACTValidate the questionThe UQC defines supported intent and constraints
- ANALYTICAL AUTHORITYCalculate deterministicallyThe governed engine produces results and evidence
- PRESENTATION & JUDGEMENTExplain the established resultPeople inspect evidence and retain the final decision
Illustrative architecture, not a complete runtime sequence.
That distinction matters even when the calculation is straightforward. Consider a question such as “Which customer bought the most?” It sounds precise until we ask what most means. The highest number of transactions? Greatest total quantity? Largest spend? Over which period? Should returns be included? A confident numerical answer to the wrong interpretation is not a useful answer. The system needs a way to clarify the intent or decline unsupported analysis before calculation begins.
I refer to the governing intermediary as the Universal Question Contract, or UQC. The detailed implementation is more involved, but the public architectural principle is uncomplicated: the question must become a controlled, checkable analytical request rather than being treated as an instruction a language model can interpret however it likes.
This is one example of a broader rule I have followed throughout the work: important responsibilities need clear owners. Presentation should not become calculation. Narrative should not become evidence. Configuration should not silently override the rules that govern results. Keeping those boundaries separate makes faults easier to locate and later extensions easier to reason about.
An answer needs a route back to its evidence
Deterministic execution addresses one important source of uncertainty, but it does not automatically make every answer meaningful or every input correct. Data may be incomplete. A question may be badly framed. A perfectly repeatable calculation may answer the wrong business question.
The next part of the research has concerned what should remain attached to the answer. Instead of relying on a statement that a result is “trustworthy”, determinIQ is designed to retain the basis on which the result was produced. For supported analytical operations, the interface can expose calculation and verification information while the underlying governance model preserves a record of analytical activity within its implemented scope.
This is also a question of presentation. Most people should not have to examine calculation details just to understand a result. The answer needs to be legible first. Greater detail should remain available when the result is disputed, surprising or consequential. A clear answer and an inspectable answer should not be alternatives.
There is another distinction that became important during development: a valid result is not necessarily a significant finding. Large datasets can support many technically correct observations. A useful system should not present all of them as equally deserving attention. determinIQ uses governed qualification and attention processes to distinguish analytical results from findings that may warrant further examination. Sometimes the appropriate outcome is to report that nothing has yet qualified. That is preferable to manufacturing something interesting to fill a screen.
Giving people more than one way into the data
The question of accessibility is partly technical and partly human. A blank conversational text box is useful when someone already knows what to ask. It is much less useful when they do not know the fields, the range of supported analysis or what might matter.
I developed complementary ways to begin. Ask is for a question someone can express in ordinary language. Explore provides routes based on what a dataset can actually support. Summary gives an orientation to the available analytical picture. Insights brings forward findings that pass the relevant qualification process. Change-oriented capabilities help investigate what has moved between comparable states.
These are not meant to become competing calculation engines. They are different ways to reach a governed process. Their value lies in reducing the distance between a person’s business question and an analytical route the system is equipped to execute.
The interface has undergone substantial iteration for the same reason. Adding another button, control or dashboard can make an application more capable while simultaneously making it harder to use. Much of the later design work has concerned removing redundant choices, clarifying state and making transitions predictable. In this research, usability is not decoration applied after the engineering. It is one of the conditions under which the engineering becomes useful.
Privacy belongs at the boundary, not the end
Early data handling decisions affect everything that comes later. If a system unnecessarily discloses identifiable source values during preparation or language processing, a privacy statement written afterwards cannot undo that disclosure.
The determinIQ architecture therefore examines sensitive information before it is made available for ordinary analytical use. Supported protection routes can replace identity-bearing values with governed tokens or pseudonymous representations. Analytical operations can continue using the permitted representation, while identity recovery for presentation is treated as a separate, controlled responsibility.
Research figure 02
Protect identity before ingestion
A supported protected-identity data path
- BROWSERReview incoming dataIdentify fields and values needing protection
- PRIVACY BOUNDARYProtect before uploadReplace protected values; keep mappings locally
- GOVERNED SERVERAnalyse permitted valuesExecution works from the protected representation
- GOVERNED RESULTReturn safe analytical outputNarrative assistance uses only permitted context
- BROWSER PRESENTATIONReveal locally, if authorisedThe original identity does not become server truth
The diagram describes supported protected-value paths, not every possible dataset or export.
The distinction between protection and anonymisation is important. A value that can be restored through an authorised mapping has not become irreversibly anonymous. Equally, keeping a value away from the AI language boundary does not mean that no information whatsoever is exchanged with an AI service: the system may still provide appropriately governed schema or narrative context for the specific language task.
The research question is narrower and more practical: what does each component genuinely need to know? Reducing information exposure by design is preferable to sharing everything and hoping later controls will compensate.
This area continues to require careful testing, especially wherever protected values are revealed locally for a particular user. A visual label, export or screenshot can itself become a privacy boundary. Those details are part of the engineering work, not merely interface polish.
From an investigation to a working application
Pursuing the research question led beyond an isolated analysis engine. A real analytical journey involves introducing data, understanding its structure, asking or discovering questions, checking the basis of results and communicating findings. The system also needs administrative, support, operational and reporting capabilities if the research is to become a usable product.
The current determinIQ codebase includes governed analytical execution, question preparation, evidence-related components, discovery and presentation paths, reporting, a historical Ledger, privacy controls and an operational layer. These elements can be inspected in the implementation. The October 2026 R&D paper documented the underlying choices and the transition from a technical experiment towards a coherent product. Development after that paper has continued, including refinements to privacy-aware presentation and analytical interaction.
Research figure 03
From research into a working product
Connected responsibilities, not a feature inventory
- RESEARCH QUESTIONCan useful analysis be more accessible?Without letting AI decide what the data says
- GOVERNED COREContracts, execution and evidenceOne accountable analytical authority
- WAYS INTO THE DATASummary · Ask · Explore · InsightsIncluding change-focused investigation
- INSPECTION & COMMUNICATIONVerification · Ledger · ReportsPreserve reasoning and communicate findings
- OPERATIONAL PRODUCTPrivacy, support and administrationProvisioning, controlled updates and recovery
Conceptual responsibilities; not all features share identical maturity or readiness.
It would be misleading to treat the existence of these components as proof that every possible business question is supported, that every customer environment has been certified, or that every path is ready for unrestricted commercial use. The implementation is working research, with internal verification in defined areas and a separate set of questions about deployment, operations, usability and external assurance.
That difference between a demonstrated component and a production promise is worth keeping visible. It tells us not only what has been achieved, but where further work will provide meaningful confidence.
What the research has established so far
My central observation is that natural-language interaction does not require surrendering analytical authority to a language model. A practical architecture can separate a person’s words from a checkable analytical request, run supported calculations through a deterministic engine and preserve a route to the evidence.
The working application also demonstrates that analytical rigour and approachable interaction are not necessarily opposing objectives. Different entry points can reduce the burden of choosing a method, while progressive disclosure can keep technical detail available without placing it in everyone’s way.
These are observations from design, implementation and internal development work. They are not a declaration that source data is always reliable or that the system has received independent security, performance or accessibility certification. In particular, internal tests conducted on specific datasets and environments should be interpreted within those conditions.
The work has also reinforced a less glamorous lesson: a collection of individually functioning features is not yet a coherent product. Relationships between features, responsibility boundaries, careful error behaviour, revision paths, operational recovery and clear wording matter as much as adding one more capability.
Open questions and next investigations
Several questions remain deliberately open. How should the system communicate uncertainty arising from missing or weak data even when the calculation is correct? How consistently can users understand the difference between a valid result and a finding worth attention? Which analytical families and comparison routes provide useful generality without becoming too permissive? How can protected identity be revealed for a legitimate local purpose without allowing it to leak through an unexpected output? What additional independent tests are necessary before stronger commercial and compliance claims can responsibly be made?
There is also a research question about the development process itself. I have used AI extensively as an engineering assistant, while retaining human responsibility for the architecture, product decisions, implementation acceptance and verification. That experience has informed both the way determinIQ is built and the principles embodied in the application. AI can increase the work that is possible; it does not remove the need to check the result.
Research record and related studies
I published determinIQ: Research and development case study in October 2026 as a more extensive account of the original investigation, the architectural principles and the human-led development process. It remains the primary published historical document for that stage of the research.
Read the October 2026 determinIQ research and development case study (PDF). This is an author-published account of the research, not an independent technical audit.
The connected capability studies examine particular questions in more detail: why analytical authority is separated from language, how identity is protected across the ingestion and presentation boundaries, and why evidence should remain attached to results. A related research journal records the initial concern that prompted the work.
The continuing aim is not to produce more impressive answers. It is to make useful analysis easier to reach, easier to question and easier to explain, without obscuring where the result came from or who is responsible for the decision made with it.
Research context
Topic: Evidence and verification
Explore further
Supporting material is listed below.
Further reading is listed below.