How a Mind answers

A Mind is DataMind's AI analyst over your company's data: it asks when your question leaves room for guessing, works through the data step by step, and shapes its answer to the response mode you chose. This page covers that behaviour, so you know what to expect while an answer is written and why a Mind sometimes stops to ask.

Choose a response mode

The chat offers four response modes: Short, Analyzer (the default), Consultant and Research. You pick the mode before you send your question; it sets how deep the answer goes and how it is shaped:

• Short keeps the answer to a few sentences.
• Analyzer produces a structured brief: the main findings first, then what was analysed, the key numbers, trends and segments, and the next steps worth taking.
• Consultant answers as a consultant would: an assessment of the current situation, strategic recommendations, an implementation roadmap, risks and mitigation, expected outcomes and success metrics.
• Research runs a broader investigation: the objective and method, the data sources and scope, findings with hypotheses tested and correlations, the limits of the evidence and how confident it is, detailed recommendations and further research ideas.

A mode changes depth and answer shape, never how the work is executed. Whether the Mind splits a question across parallel agents is decided from the question itself, in every mode the same way.

Your own wording also wins over the mode: ask a "quick question" in Research and the answer still comes back short, and a request for more detail goes deeper in Short.

Whichever mode you use, an answer can carry charts, KPI cards, diagrams and references to the queries and documents behind it. Those blocks have their own rules, on Charts, KPI cards and diagrams in answers.

Why a Mind asks before it answers

When a request leaves room for guessing, the Mind asks instead of guessing: a vague period such as "recent" or a quarter with no year, an ambiguous metric, an undefined scope, a term with several meanings, or a comparison with no baseline (a "better" that does not say better than what).

• The questions arrive as one set above the message box, and a question's answers come as clickable choices: between 2 and 9 of them. Related questions come together, so one pass clears every ambiguity at once.
• When a date could mean two different years, the choices spell out the readings, so you pick the one you meant.
• The Mind waits for your answers before it starts. How to answer, skip or move between questions is on Questions from a Mind.
• A question whose answer is an open value, such as a name you must type, is asked as plain text instead.
• No response mode skips this step: even Short asks when the request is ambiguous.
• If nobody is there to answer, as in a scheduled run, the Mind states the assumptions it will work from and continues instead of waiting.

Follow the plan of a bigger task

• When a request needs several steps, or you ask for a plan, the Mind publishes one: a small progress strip lists the steps and tracks them while the answer is written. A one-step request gets no plan.
• A plan holds between 1 and 15 steps, and exactly one step is in progress at a time.
• A step is marked completed only when its result is in; nothing is ticked off on a guess.
• If a step fails, the Mind re-plans: it marks the failed step skipped with the reason and carries on with the fix, instead of restarting the task.

Why a large result comes back summarised

• Query results are read in pages of at most 50 rows, and a page always carries the total, so a partial result is never mistaken for the whole.
• When a few more pages are needed for the analysis, it reads them.
• At 200 rows or more it stops reading row by row: it pages through selectively, summarises, and tells you the full total.
• At 500 rows or more it prefers aggregates and samples over reading every row.
• If you truly need every row, the Mind explains the limit and offers an aggregation query that answers the question instead of listing the data. When an answer rests on a query, it references that query, and you can open it and export the full result from there.

When the Mind stops and says why

• If it cannot support an analysis, it declines instead of guessing: the data may be missing, the answer may rest on assumptions it cannot stand behind, the data quality may make the numbers unreliable, or the question may go beyond what it can see. A refusal opens with "I cannot provide this analysis because" and the specific reason, and it says what data would be needed, what it can offer instead, and what would make the analysis possible.
• If results look wrong, it checks before reporting. Values outside any sensible range, sudden swings that defy the business, impossible percentages, or metrics that contradict each other make it re-check the calculation, then flag the finding instead of passing it on as fact. The flag opens with ⚠️ ANOMALY DETECTED, lists the checks that were run, sets out possible explanations (a data-quality problem, a calculation issue, or a real change in the business) and gives a recommendation.

What a switched-off tool means

Some chats run with fewer tools than usual. A switched-off tool is never used, mentioned or promised: the Mind answers with what it can reach, and says so when something falls outside it.

• With web search and page reading off, it answers from your company's data and its knowledge base, and it tells you when a question would need outside information that it cannot fetch.
• With knowledge-base search off, it works from your data alone and does not promise to look up internal documents.
• Planning is always available.

What changes in a long chat

• A long chat compacts itself: when the conversation approaches 70% of what the chat can hold, the older turns are replaced by a summary, while your first message and the newest turns stay as they are.
• The summary keeps the goals, the key findings, the queries and documents that were made, the decisions and anything still unresolved, so a follow-up can still reach the work that came before.
• You do not need to start a new chat after that; the conversation continues from the summary.

Personal data in an answer

• Columns that hold personal data reach the AI masked: where a value would appear, the Mind sees a code such as {{EMAIL_a1b2c3d4e5}} instead, and it never reconstructs what the code hides.
• Counting and grouping still work: the Mind can count the people behind a code and group rows by it. It never joins on these columns or searches inside the codes.
• The values you see in the product stay the real ones.
• The full rules, including what happens when your company blocks personal data from AI requests, are on Personal data in AI work.

Questions, answered

Does the response mode change how the Mind works?

No. The mode sets the depth and shape of the answer, never the execution: the Mind plans, queries and verifies the numbers the same way in every mode. Your own wording in the question still wins over the mode.

Why did the Mind ask me something instead of answering?

Your request left more than one reading open: a vague period, an ambiguous metric, an undefined scope, or a comparison with no baseline. The Mind asks instead of guessing, and it waits for your answer before it starts.

What happens if nobody answers its question?

In a scheduled or automated run there is no one to answer, so the Mind states the assumptions it will work from and continues. In a chat, the question waits until you answer or skip it.

Why is this answer shorter than the one before?

The mode, your own wording, or both. Short keeps answers to a few sentences, and a phrasing like a quick question wins over the selected mode.

Why does a big result come back summarised?

Results are read 50 rows at a time. At 200 rows or more the Mind pages selectively and summarises instead of reading everything, and at 500 or more it prefers aggregates or samples. If you need every row, it explains the limit and offers an aggregation query.

Why did a plan step get skipped?

That step failed, and the Mind re-planned around it: the step is marked skipped with the reason, and the fix is added as new work instead of restarting the task.

What are the parallel agents behind a broad answer?

For a question that splits into independent parts, the Mind hands the parts to sub-agents that work in parallel, and their results arrive together to be checked and combined. Whether to split the question, and how deep to go, is the Mind's own decision, taken from the task itself; it never asks you. While a wave runs, a short status line on the answer shows what is being worked on.

Why doesn't it search the web for my question?

Web search is switched off in this chat, so the Mind answers from your company's data and its knowledge base, and it says when a question would need outside information it cannot fetch. It never promises a search it cannot run.

What happens to older messages in a long chat?

When the chat approaches 70% of what it can hold, older turns are compacted into a summary; your first message and the newest turns stay as they are. Follow-ups keep working, because the summary keeps the goals, findings, named queries and documents, decisions and open questions.

Why does an answer flag an anomaly?

The numbers looked wrong: values outside any sensible range, sudden swings, contradictions between metrics. The Mind re-checked the calculation and flagged the finding with the checks it ran, the possible explanations and a recommendation, instead of reporting the numbers as fact.

Why did the Mind say it cannot provide an analysis?

It declines rather than guess: the data may be missing, the answer may rest on assumptions it cannot stand behind, the data quality may make the numbers unreliable, or the question may go beyond what it can see. The refusal says what is missing, what it can offer instead, and what would make the analysis possible.

How does the Mind work with personal data?

Values from personal-data columns reach the AI masked as codes; the Mind counts and groups with them, but it never joins on them, and you keep seeing the real values in the product. Personal data in AI work has the full rules.