# DataFact Canonical URL: https://datafact.ai/ DataFact is an AI analyst for documents and databases, built for answers you can verify. ## Answers you can verify “Answers you can verify.” is DataFact's brand statement. It means DataFact is designed to keep useful supporting material available for review when a workflow returns it. It is not a promise that AI never makes mistakes, that every response contains a citation, or that every claim has automatic provenance. ## AI analyst for documents and databases DataFact helps teams ask business questions that can span structured and unstructured information. Supported documents can provide passages, tables, figures, and business context. Connected databases can provide measurable records and generated analysis. Configured Agents determine which eligible resources and capabilities are available to a workflow. --- # What is a verifiable AI analyst? Canonical URL: https://datafact.ai/verifiable-ai-analyst A verifiable AI analyst is an AI system that helps people analyze business information while keeping the supporting information behind important conclusions inspectable. The work may involve documents, database records, generated queries, calculations, citations, and a sequence of analytical steps. Inspectability gives a reviewer a way to ask where an answer came from, which definitions were applied, and whether the conclusion follows from the available material. “Verifiable” does not mean AI never makes mistakes. It means a person has more of the information needed to check an answer, identify a weak assumption, reproduce a number, or decide that more evidence is required. ## Independently useful facts - A citation shows where supporting material was retrieved; it does not prove that the material was interpreted correctly. - A database query can be inspected for tables, fields, joins, filters, and time ranges. - A document can preserve business context that is absent from a database schema. - Verification depends on access to the relevant source, not only on the wording of an answer. - An inspectable workflow can still produce an incomplete or incorrect conclusion. ## Why it matters Business answers are often used outside the conversation that produced them. A finance, operations, or risk team may need to review the source, repeat the calculation, or explain the result to someone who did not ask the original question. A fluent answer without supporting context makes that review difficult. Keeping useful artifacts visible reduces the amount of guesswork between receiving an answer and deciding whether to act on it. ## How it works 1. Define the question: State the subject, measure, scope, time period, and expected output clearly enough to evaluate the result. 2. Inspect the sources: Review the document passages, database records, or other source material that the workflow made available. 3. Inspect the method: Check queries, filters, definitions, calculations, and analytical steps when they are present. 4. Review the conclusion: Confirm that the answer addresses the question and that its qualifiers match the supporting information. ## Example: Investigating a margin change Suppose an operations leader asks why margin fell on a shipping lane. Database records may establish the change and identify the affected period. A logistics review may explain a temporary surcharge or contract exception. A reviewer should be able to inspect the query that calculated the margin, the filters that selected the lane and dates, and the document passage used to explain the change. If any of those artifacts is missing, the answer may still be useful, but its verification boundary is narrower. ## Limitations - Supporting material may be unavailable because a source is not connected, indexing is incomplete, or access has changed. - A visible source can be stale, ambiguous, or non-authoritative. - A correct query can still answer the wrong business question when the definition or scope is wrong. - Not every DataFact response contains citations, queries, charts, or files; output depends on the Agent, source type, and workflow. ## DataFact's approach DataFact connects supported documents and databases to configured Agents. An answer can include citations, charts, generated files, query context, or execution details when those are produced by the workflow. DataFact also keeps business knowledge and Analysis Model versions separate from source records. Reviewers should still validate consequential conclusions against authoritative sources and approved business definitions. --- # AI answer verification: definition and examples Canonical URL: https://datafact.ai/ai-answer-verification AI answer verification is the process of checking whether an AI-generated conclusion is supported by the underlying sources, queries, business definitions, calculations, or other inspectable evidence. Verification is not one universal score. A document claim may need a source passage and its surrounding context. A database result may need the executed query, the relevant schema, and the approved definition of the metric. A conclusion that combines sources may also need a reviewer to reconcile dates, identifiers, and conflicting statements. The goal is not to make an AI response look certain. The goal is to make the basis of the response clear enough that a person can judge what is supported, what is inferred, and what remains unresolved. ## Independently useful facts - Source verification asks whether the cited material actually supports the conclusion. - Query verification checks tables, fields, joins, filters, grouping, and time boundaries. - Definition verification checks that a business term or metric means what the decision-maker expects. - Calculation verification checks arithmetic, units, denominators, aggregation, and rounding. - Cross-source checking looks for conflicts in identity, scope, date, and authority. - Visible SQL alone does not prove an analytical answer is correct. ## Why it matters AI can produce a plausible explanation even when an input is incomplete or a business term is ambiguous. Verification separates the readability of the answer from the quality of the support behind it. The check should be proportional to the decision. A low-impact exploratory question may need a quick review, while a financial, regulatory, or customer-facing conclusion may require an authoritative source and a reproducible calculation. ## How it works 1. Source verification: Open the relevant passage or record and read enough surrounding context to evaluate the claim. 2. Query verification: Check the selected table and fields, join behavior, filters, grouping, date logic, and access rules. 3. Definition verification: Confirm the metric or business term against the definition approved for that decision. 4. Calculation verification: Reproduce the arithmetic and check units, denominators, aggregation, null behavior, and rounding. 5. Cross-source checking: Compare source dates, identifiers, scope, and authority; record conflicts instead of silently choosing one. ## Example: Checking renewal revenue at risk An answer may report the annual recurring revenue associated with accounts renewing this quarter. The SQL can be valid and still use the wrong revenue field, include already-renewed contracts, omit an account-status filter, or apply a different definition of annual recurring revenue. Verification therefore includes the query and the interpretation around it. If account notes are used to explain risk, the reviewer should also check whether the note is current and whether it refers to the same account and renewal period. ## Limitations - Visible SQL may select the wrong table, field, filter, join, date boundary, or business definition. - A citation may be topically related without supporting the specific conclusion. - A calculation can be arithmetically correct while using incomplete or incompatible inputs. - Conflicting sources may require an owner to decide which source is authoritative. - Missing artifacts should be treated as a limit on verification, not silently filled in. ## DataFact's approach DataFact can expose supporting material such as document citations, query context, charts, files, and execution details when a workflow produces them. Datasets and Analysis Models can also carry reviewed data boundaries and reusable business knowledge. These artifacts support review; they do not remove the need for human judgment. Important answers should be checked against authoritative source data and the business definition that applies to the decision. --- # Why AI analysis needs documents and databases Canonical URL: https://datafact.ai/documents-and-databases Documents-and-databases analysis combines structured records with unstructured business material so one question can be evaluated across both kinds of source. Structured data is useful for measurable facts such as revenue, usage, transaction history, dates, and counts. Documents often contain business context that is not represented in a database schema, including contract terms, explanations, decisions, exceptions, and plans. The two source types are complementary. A database may show what changed, while a document helps explain why. Combining them does not automatically resolve ambiguity: the workflow still needs consistent identifiers, compatible time ranges, appropriate permissions, and a review of which source is authoritative. ## Independently useful facts - Structured data is useful for measurable facts such as revenue, usage, and transaction history. - Documents often contain business context that is not represented in database schemas. - A shared account, product, supplier, or project identifier is often needed to connect the two source types. - Source dates matter because a current database record may conflict with an older document. - Access rules still apply when an analysis spans more than one source. ## Why it matters Many operational questions have a quantitative half and a contextual half. Treating either half as the complete answer can hide an important exception or produce an explanation that is not supported by the measured result. This is not primarily a connector problem. The analytical work is deciding which facts belong together, which definitions apply, and how uncertainty or conflict should be represented. ## How it works 1. Identify the measurable part: Use structured records for quantities, dates, status, usage, transactions, and other consistently typed fields. 2. Identify the contextual part: Use contracts, reports, notes, plans, and other documents for terms, decisions, exceptions, and explanations. 3. Join by business identity: Relate sources through a stable account, supplier, project, product, or other business identifier. 4. Align scope and time: Check that the records and documents refer to the same entity, period, definition, and decision context. 5. Keep each source inspectable: Preserve the query or records behind measured facts and the passage behind document-based context. ## Example: Which customers are at risk of renewal, and why? A database can identify accounts with an upcoming renewal, quantify annual recurring revenue, summarize product usage, and count support activity. Those fields establish who is in scope and what measurable signals changed. Contracts and account documents can add renewal terms, QBR notes, customer reports, and account plans. They may explain an exception or concern that is not encoded as a database field. A complete investigation connects the sources by account and period, then keeps the supporting record and passage available for review. ## Limitations - Names and identifiers may not match across systems. - Documents may be stale, duplicated, scanned poorly, or missing important context. - Database fields may use a different business definition from the document author. - A source may be connected but unavailable to a particular Agent or user because of access rules. - Conflicting facts should be surfaced for review rather than blended into false certainty. ## DataFact's approach DataFact provides separate product paths for document intelligence and database analysis, then lets configured Agents use eligible resources in a workflow. Documents can retain locators used by citations; Datasets define a usable SQL boundary and can carry data-specific knowledge. The availability of a source, query, citation, chart, or generated file depends on the Agent and the work performed. Reviewers should inspect the artifacts that are present and treat missing support as a limitation. --- # Evidence Canonical URL: https://datafact.ai/docs/concepts/evidence-and-citations A citation links an answer to source content returned by retrieval. It helps a reviewer inspect the quoted or highlighted context, but it does not automatically validate the interpretation. Database analysis, model-only answers, tool results, unsupported sources, or changed access can limit which citations and supporting artifacts are available. # Business Context Canonical URL: https://datafact.ai/docs/concepts/knowledge-and-accuracy Business entities, relationships, general knowledge, business knowledge, and optimization examples are distinct DataFact objects that can give an Agent relevant terminology and decision context. They are not merged into one backend object. Incorrect, conflicting, stale, or overly broad knowledge can reduce answer quality. # Deep Analysis Canonical URL: https://datafact.ai/product/agents DataFact Agents combine configuration, instructions, enabled capabilities, and optional linked resources. A Session contains the messages for one Agent. Multi-step work can retrieve, query, analyze, and produce supporting outputs, but the exact steps and artifacts depend on the workflow. # Evaluation Canonical URL: https://datafact.ai/docs/concepts/evaluation-lifecycle DataFact's current evaluation workflow combines user feedback, review, Test Sets, repeatable question runs, task records, and downloadable results. The current Test Set interface does not define expected answers, score individual results, or compare runs automatically. Evaluation conclusions therefore require human review or an external method. --- # DataFact vs ChatGPT: When to Use Each Canonical URL: https://datafact.ai/compare/chatgpt DataFact and ChatGPT can both support analysis, but they start from different product scopes. ChatGPT is a general-purpose AI workspace with use cases across writing, research, data analysis, and business work. DataFact is an AI analyst centered on business questions across connected documents and databases. The right choice depends on the task, the sources that must be used, the controls around those sources, and what a reviewer needs to inspect. Feature availability can also depend on plan, workspace configuration, connectors, permissions, and product changes. DataFact design intent: DataFact is designed for recurring business analysis over configured documents and databases, with business knowledge, access boundaries, and supporting artifacts when the workflow produces them. ChatGPT design intent: OpenAI presents ChatGPT as a broad workspace for tasks that include analyzing datasets, working with files, preparing business reviews, researching, writing, and building workflows. ## Comparison by area ### Documents DataFact: Document ingestion, parsing, retrieval, and citations are part of the product's business-analysis workflow. ChatGPT: Official ChatGPT learning material includes working with reports, files, and research. Exact source handling depends on the selected workflow and connected tools. ### Databases DataFact: Connected databases are exposed through Datasets, access rules, schema context, and generated analysis. ChatGPT: ChatGPT can analyze datasets and use integrations, but the exact database path and governance depend on the user's workspace and integration setup. ### Analysis workflow DataFact: Configured Agents can use linked resources for questions, multi-step analysis, and report workflows. ChatGPT: ChatGPT supports a wider range of conversational and task workflows beyond data analysis. ### Verification and evidence DataFact: Answers can include citations, query context, charts, files, or execution details when the workflow produces them; these artifacts still require review. ChatGPT: Source visibility varies by ChatGPT workflow. Review the sources, files, analysis, or integrations actually used for the response. ### Context and governance DataFact: Datasets, Knowledge, Analysis Models, Agents, and workspace permissions organize business context and access for analytical work. ChatGPT: Context, retention, access, and connector behavior depend on the ChatGPT product, plan, workspace settings, and integrations in use. ## Official sources checked - OpenAI Learn — ChatGPT use cases: https://learn.chatgpt.com/use-cases — Official overview used to characterize ChatGPT's broad task and data-analysis scope. - DataFact documentation — What is DataFact?: /docs/getting-started/what-is-datafact — Current DataFact product definition and limitations. --- # DataFact vs Hex: When to Use Each Canonical URL: https://datafact.ai/compare/hex DataFact and Hex both address AI-assisted analysis, but their published workflows emphasize different starting points. DataFact centers business questions that can span documents and databases. Hex describes itself as an AI analytics platform for exploring data, collaborating with agents, and building shareable apps. This comparison uses Hex's current official documentation and DataFact's current public documentation. It avoids treating a missing documentation statement as proof that a capability does not exist. DataFact design intent: DataFact is designed for configured analysis across document sources and databases, with support for citations, governed SQL, reusable business knowledge, and multi-step Agent workflows. Hex design intent: Hex is designed around analytics work that can include warehouse connections, SQL and Python notebooks, semantic models, AI agents, visualizations, dashboards, reports, and published data apps. ## Comparison by area ### Documents DataFact: Document intelligence is a first-class product area for PDF, Office, tabular, Markdown, and scanned material, with locators used for citations where available. Hex: Hex's official overview and notebook documentation emphasize warehouse data, uploaded CSV files, APIs, notebooks, and analytics artifacts. Teams with long-form document analysis needs should verify the exact Hex workflow they require. ### Databases DataFact: DataFact turns plain-language questions into governed database analysis through connected Datasets and applicable access rules. Hex: Hex documents direct data connections, warehouse queries, schema browsing, SQL/Python work, and AI analysis over accessible connections. ### Analysis workflow DataFact: DataFact Agents package instructions, resources, and capabilities for questions, deep analysis, and report workflows. Hex: Hex projects use notebook-like cells and can turn analysis into interactive apps, dashboards, and reports; its Notebook Agent assists within projects. ### Verification and evidence DataFact: A workflow can return citations, query context, charts, files, or execution details, with explicit limits when support is absent. Hex: Hex states that teams can retain visibility into how answers are produced and can inspect AI-generated SQL and Python in its analytics workflows. ### Context and governance DataFact: Knowledge and versioned Analysis Models organize reusable context alongside Dataset and Agent access boundaries. Hex: Hex documents native semantic authoring, sync from supported external semantic models, workspace rules, connection permissions, and governed shared context. ## Official sources checked - Hex documentation — What is Hex?: https://learn.hex.tech/docs — Official product overview used for Hex's design intent and published workflows. - Hex documentation — Data connections: https://learn.hex.tech/docs/connect-to-data/data-connections/data-connections-introduction — Official source for database connection and query behavior. - Hex documentation — Semantic modeling: https://learn.hex.tech/docs/connect-to-data/semantic-models/intro-to-semantic-models — Official source for native and synchronized semantic models. - Hex documentation — Notebook Agent: https://learn.hex.tech/docs/explore-data/notebook-view/notebook-agent — Official source for Agent-assisted notebook analysis. - DataFact documentation — What is DataFact?: /docs/getting-started/what-is-datafact — Current DataFact product definition and limitations.