--- title: "AI You Can Trust – Answers from Your Definitions" canonical: "https://www.owox.com/solutions/ai-you-can-trust" updated: "2026-10-20" --- AI you can trust # When the AI gets a number wrong, the question lands on you OWOX Data Marts makes the AI answer from the Data Marts you define. It can't pick its own meaning of revenue, a question your data can't answer gets a refusal, and every answer leaves a record you can open. [Start free](https://www.owox.com/app-signup) · or [book a call](/demo) ↓ A real data analyst, mid-conversation. Try it — click anything. ## The problem you inherit ## An AI answer goes wrong in two ways, and neither announces itself Ask the same revenue question three ways and you can get two figures. Or you get one wrong figure every time, because the AI settled on a meaning of revenue that still counts cancelled orders. Either way the number looks right, and the question about it comes to you. Without a written definition ### The AI picks the meaning * Two people ask about last quarter's revenue and get two figures, each correct for what it measured * The AI settles on its own meaning of revenue and repeats it: in one of our tests, $670,408.76 in three runs out of three, where completed orders only give $573,984.92 * A figure that is already in a deck has no calculation anyone can open * A pasted export leaves no record of the calculation outside the chat * When it turns out wrong, it is your name on the number With Data Marts you own ### You pick the meaning * Each figure has one written meaning, in a Data Mart with a named owner * The AI can ask only for published Data Marts, and only for data the person is permitted to query * A question your data can't answer gets a refusal: asked for headcount, our demo project replied "I can't give you a headcount from this connector: it has no employee or HR data" * Run History keeps the request, who sent it, and the SQL that ran * Change a definition once and every report and every answer reads the new one This doesn't make the AI right every time. It makes the meaning yours and a wrong answer findable. [Read what our own tests showed →](https://www.owox.com/blog/articles/can-you-trust-ai-business-numbers) F04 · Definitions you own ## One written meaning for every figure, with a name on it You define each figure once, as SQL, in a Data Mart: what it counts, what it leaves out, who owns it. The AI no longer chooses between honest meanings of revenue. It reads yours. * A Technical Owner and a Business Owner on every Data Mart, so anyone can see who defined the number * Descriptions and relationships between Data Marts live next to the definition, not in a separate document * Plug in the SQL you already have; there is no new modeling language to learn ![](https://cdn.owox.com/www/portfolio-management-data-destination/claude/logo/2.svg/public)Claude · March revenue?$412K ![](https://cdn.owox.com/www/portfolio-management-data-destination/google-sheets/logo/1.svg/public)Google Sheets · Sales in March?$412K ![](https://cdn.owox.com/www/portfolio-management-data-destination/data-studio/logo/1.svg/public)Data Studio · March income?$412K Revenue · defined once · by the data owner F05 · Run History ## Open the calculation behind any answer Every query an AI runs through OWOX Data Marts is listed in Run History: the request the AI sent, who sent it, when, which Data Mart it used, and the SQL that ran. Open one and check it yourself. * The record keeps the request and the executed SQL, never the result rows * It has an entry for each query the AI runs, not for each reply * An answer with no record cannot be shown to come from your data, which is how you spot one that came from the chat's memory Run History · 4 requests · revenue by channel, 30d · Claude · Anna · just now · ad spend by campaign · Claude · Marco · 12m ago · orders by country · Sheets · Lena · 1h ago · refunds, last week · Claude · Marco · 3h ago SQL that ran · SELECT channel, SUM(revenue) AS rev FROM revenue WHERE date >= 30d GROUP BY channel request · sender · SQL · no result rows kept F06 · Scope you control ## "I don't have that" instead of an improvised number You decide which Data Marts the AI may read and which Contexts a person sees. Ask for something nobody has defined and the AI says so, and the request goes to the data owner instead of into a guess. * The AI can ask only for published Data Marts * Contexts group Data Marts and limit what each member sees; only a Project Admin creates or deletes them * A metric nobody defined is a request to add it, not a number the AI makes up ![](https://cdn.owox.com/www/portfolio-management-data-destination/claude/logo/2.svg/public)Claude · How many people work in Sales? Data Marts · Revenue · Ad Spend · Orders · Headcount OWOX · published only · I don't have that. · No published Data Mart defines headcount. · Asked the data owner ## Forward this to whoever asked you whether the AI can be trusted [Get started free →](https://www.owox.com/app-signup) ## How it works ## From "can we trust this?" to answers you can defend, in three steps Start with one number. The long step is agreement on what it means, not the connection. Step 1 ### Write one number down Pick the figure your leadership argues about most. Write what it counts and what it leaves out as SQL in a Data Mart, and get the person who uses it to sign it. Step 2 ### Publish it where people ask Publish the Data Mart so it reaches Google Sheets and the AI chats your team already uses, Claude and ChatGPT. Step 3 ### Check any answer in Run History Open the SQL behind a figure. Then ask one thing your data can't answer and see that the AI declines. ## Straight answers ## What we don't claim No tool makes an AI right every time. These are the limits we found, and they hold whichever tool you use. ### It can pick the wrong number from a correct list A written meaning narrows the choice. Read the label on the first figure every time. ### It can still add up wrongly If the AI does its own sums on correct figures, check the arithmetic against the calculation in Run History. ### Sound definitions don't prove sound rows A written meaning says nothing about whether the data under it is clean. That takes its own checks. ### Your figures go to the AI company To write a reply the AI reads the figures behind it, so they reach Anthropic, OpenAI or your cloud provider. You choose what it may read. ## In the series ## Read the tests behind this page Our own runs, step by step. [Article · Can You Trust AI with Your Business Numbers? · Oct 11, 2026](/blog/articles/can-you-trust-ai-business-numbers) [Article · Trace an AI answer to its SQL: from number to source table · Oct 14, 2026](/blog/articles/ai-analytics-hallucinations-sql-traceability) [Article · Governed AI Analytics on BigQuery: Who Writes the SQL When Your Team Asks in Chat · Oct 9, 2026](/blog/articles/ai-analytics-on-bigquery) [Article · Governed AI Analytics on Snowflake: Who Writes the SQL When Finance Asks · Oct 10, 2026](/blog/articles/ai-analytics-on-snowflake) [Article · AI Data Analyst Tools You Can Trust: An Honest Comparison · Oct 9, 2026](/blog/articles/ai-data-analyst-tools-comparison) [Article · Semantic Layer for AI Agents: Do You Need One, or Do You Already Have It? · Oct 9, 2026](/blog/articles/semantic-layer-for-ai-agents) ## FAQ ## Questions analysts ask Does OWOX stop the AI from hallucinating? It removes the place where most wrong answers start: the AI picking its own meaning of a number. It doesn't make the AI right every time. In our own 25-question test on a demo shop, an AI that wrote SQL from table names and could not run it returned a plausible wrong figure in 7 of 66 runs. Through published Data Marts, one run per question, 20 of 22 answerable questions matched the reference. One model and one demo shop, so it shows what happened, not how often. The full test is in [our guide to trusting AI with business numbers](https://www.owox.com/blog/articles/can-you-trust-ai-business-numbers). Who writes the SQL that produces the number? The SQL comes from your Data Marts and the relationships you defined, not from a language model guessing at table names. The AI sends a request for a published Data Mart, and Run History shows you the request and the SQL that ran. You can read both. Where do our figures go when someone asks? To write a reply, the AI has to read the figures behind it, so they are sent to the company that runs the AI: Anthropic if you use Claude, OpenAI if you use ChatGPT. Our [documentation](https://docs.owox.com/docs/getting-started/setup-guide/mcp/) says the data rows and totals are sent to that provider. Decide which Data Marts the AI may read before anything is connected, and leave out what must not leave the company. We already have definitions in dbt or a BI tool. Do we start over? No. Use them. The test is the same wherever they live: each number has one written meaning that someone owns, the AI may answer only from what is written, and you can open the calculation behind every answer. OWOX Data Marts lets you plug in SQL you already have. Is this a semantic layer? No. OWOX builds a data model: a description of your business, its entities and how they relate, that exists before anyone asks a question. A semantic layer is context assembled to answer one question. [Do you need a semantic layer for AI agents?](https://www.owox.com/blog/articles/semantic-layer-for-ai-agents) compares the options. How long does it take? Connecting is the short part. The long part is getting the people who use a number to agree on what it means, and no tool shortens that. Start with the one number your leadership argues about most. [Still have questions? Talk to us →](/demo) ## Two ways to start RECOMMENDED ### Start with one number Define the figure your leadership argues about most, publish it, and ask for it in your AI chat. * Free to start * Works with Claude and ChatGPT * Every answer has a record you can open [Get started free →](https://www.owox.com/app-signup) NEED HELP? ### Book a demo with our team Walk through Data Marts, Run History and the AI connection with a data strategist, on your own warehouse. * 30-minute walkthrough * See an answer traced to its SQL * Bring the number your team argues about [Book a demo](/demo) ## Book a call You are booked. The invitation is on its way to , with the calendar invite and the meeting link. [Reschedule](#) [Cancel](#) ## Schedule a meeting ## Schedule in person meeting at Tip Top Toronto You are booked. The invitation is on its way to , with the calendar invite and the meeting link. [Reschedule](#) [Cancel](#) ## Links Pages this page links to, on this site and on docs.owox.com. Where the page has a Markdown twin, its address follows the link. - [Start free](https://www.owox.com/app-signup) - [book a call](https://www.owox.com/demo) — /demo.md - [Read what our own tests showed →](https://www.owox.com/blog/articles/can-you-trust-ai-business-numbers) — /blog/articles/can-you-trust-ai-business-numbers.md - [Article · Trace an AI answer to its SQL: from number to source table · Oct 14, 2026](https://www.owox.com/blog/articles/ai-analytics-hallucinations-sql-traceability) — /blog/articles/ai-analytics-hallucinations-sql-traceability.md - [Article · Governed AI Analytics on BigQuery: Who Writes the SQL When Your Team Asks in Chat · Oct 9,…](https://www.owox.com/blog/articles/ai-analytics-on-bigquery) — /blog/articles/ai-analytics-on-bigquery.md - [Article · Governed AI Analytics on Snowflake: Who Writes the SQL When Finance Asks · Oct 10, 2026](https://www.owox.com/blog/articles/ai-analytics-on-snowflake) — /blog/articles/ai-analytics-on-snowflake.md - [Article · AI Data Analyst Tools You Can Trust: An Honest Comparison · Oct 9, 2026](https://www.owox.com/blog/articles/ai-data-analyst-tools-comparison) — /blog/articles/ai-data-analyst-tools-comparison.md - [Article · Semantic Layer for AI Agents: Do You Need One, or Do You Already Have It? · Oct 9, 2026](https://www.owox.com/blog/articles/semantic-layer-for-ai-agents) — /blog/articles/semantic-layer-for-ai-agents.md - [documentation](https://docs.owox.com/docs/getting-started/setup-guide/mcp)