Ask Your Data. Get answers backed by real maths.
Ask Your Data in DataHub Pro is a tool-use AI: it picks the right pandas operation, runs it on your file, and returns the answer with the call trace attached. No more "the AI said the total was £4.2m, but it didn't actually compute the total". Every answer ships with the rows used, the filters applied, and the function called — so reviewing the AI's work is faster than redoing it.
Updated 7 May 2026
Why most "chat with your data" products miss
The promise of a conversational analyst is genuinely compelling: type a question, get an answer. The reality with most current implementations falls into one of three traps.
Trap 1 · Free-form text generation
The model reads your data as text and predicts plausible answers. Sometimes those answers are right. Sometimes they're confidently wrong. You can't tell which without recomputing — defeating the point of the AI.
Trap 2 · Pre-built dashboards in disguise
Some tools route every "question" to a pre-built tile. Ask anything outside the tile and you get "I can't answer that". It's a glorified search bar, not a chat with your data.
Trap 3 · No audit trail
Even when the AI does compute, you don't see the working. For finance, audit, regulated industries — "trust me" isn't a deliverable.
Ask Your Data is built on the opposite premise: the AI exists only to orchestrate tools. Every number it returns was produced by code on your data, and the code is shown.
How tool-use AI works under the hood
A loop, not a single LLM call. The model proposes a tool to call; the tool runs against your file; the result feeds back; the loop continues until the answer is grounded.
Tool catalogue, not free generation
The AI can only act through 20+ deterministic tools — load file, head, describe, filter by value/date/regex, aggregate by group, sort, correlate, top-N, pivot, forecast, segment. Each tool returns numbers, not prose.
Visible reasoning trace
Each AI turn shows: thought (one sentence on what to do next), tool name, arguments, result. Click any answer to expand the trace. The same question gives the same trace and the same answer every time.
Multi-step workflows
Questions like "top 5 customers by revenue in Q3 minus refunds" require multiple tool calls — load → filter Q3 → aggregate by customer → subtract refunds → top 5. The loop handles the orchestration.
Conversation memory
Follow-ups work naturally: "why did customer X drop in October?", "break that down by product", "forecast for the next quarter". The model carries context across turns and re-uses prior tool outputs where it can.
Charts on demand
Ask for a chart in natural language and the AI picks an appropriate type (bar, line, scatter, histogram) and renders it inline. Charts can be pinned to a custom dashboard or sent into Auto Report.
Suggested follow-ups
After every answer, the platform proposes 2-3 high-leverage follow-up questions based on what's interesting in your data. They're not generic prompts — they're contextual to the file you uploaded.
What this looks like in practice
Three example sessions, each starting from a single CSV upload.
Session 1 · Sales analysis
- You: "What's our biggest customer this year?"
- Trace: filter_by_date(2026-01-01, today) → aggregate_by_field(customer, sum, revenue) → sort(desc) → top_n(1)
- Answer: "Acme Ltd at £487,200 across 18 invoices." — with the underlying rows shown.
- You: "Plot their monthly revenue."
- → chart rendered inline, pinnable.
Session 2 · Variance investigation
- You: "Why was October so far over budget?"
- Trace: filter_by_date(2025-10) → group_by(line_item) → variance(actual, budget) → sort(desc, abs)
- Answer: "Two line items drove 86% of the variance: Marketing £42,100 over (campaign launch), and Cost of Sales £28,400 under (delayed supplier invoices)."
Session 3 · Customer cohort
- You: "Which clients haven't ordered for 90 days?"
- Trace: aggregate_by_field(customer, max, order_date) → filter_by_date(< today - 90)
- Answer: 47 clients, total prior LTV £312,000. CSV download offered.
Each of these would have taken 5–20 minutes in Excel. Each took under 30 seconds in Ask Your Data, with a complete audit trail of what was computed.
Who Ask Your Data is built for
Analysts who get interrupted with ad-hoc questions
The classic "just one quick number" from a stakeholder that takes 20 minutes to compute. Hand them a Ask Your Data session on the underlying file and they answer their own questions.
Finance, audit and regulated teams
The audit trail is the unlock. Any answer returned is reproducible from the trace, so internal review and external audit both move faster.
Founders and operators without an analyst
If you don't have a data team, Ask Your Data is the closest thing to having one on demand. Drop in your Stripe export or your CRM CSV and start asking questions.
When Ask Your Data is the wrong tool
If you need true real-time analytics over a transactional database (sub-second latency, freshness measured in seconds), this isn't that. Ask Your Data works on the file you upload at upload time. For live data, use a warehouse-connected BI tool — and bring the snapshot here when you need to write a report.
FAQs
How is this different from ChatGPT's Advanced Data Analysis?
ChatGPT's ADA mode runs Python in a sandboxed container and can do real maths. The differences: (1) ADA generates code in the open ended chat — there's no curated tool catalogue, so a wrong column name silently produces a wrong answer; (2) the audit trail is the chat history, not a structured trace; (3) session timeouts and lost context are common; (4) there's no native export to Word/PowerPoint, white-labelling, or dashboard pinning. Ask Your Data is purpose-built for analytics workflows that end in a deliverable.
Does the AI ever hallucinate numbers?
Numerical answers come from deterministic tools running real pandas on your file. The model's role is to pick the tool, not to invent the number. We've not seen a confirmed hallucinated number in production, but the audit trail makes any error visible — wrong column picked, wrong filter applied — so you'd see it before acting on it.
Can it answer questions across multiple sheets or files?
Within a single workbook (multiple sheets), yes — pick a primary sheet on upload and the AI can join across sheets when asked. Across multiple files, the workflow is to combine them first using Pipelines (the platform's join/transform builder) and then chat over the combined output. We're working on multi-file native support.
Does it remember previous sessions?
Within a single conversation, yes — context, prior answers, and tool outputs are reused for follow-ups. Across sessions, no — each new conversation starts clean for data-isolation reasons. Saved insights and pinned charts persist across sessions in the dashboard.
How long do answers take?
Most answers return in 5–8 seconds on files under 1M rows. Multi-step questions (e.g. "top 5 customers by revenue minus refunds, plotted monthly") can take 15–25 seconds because each tool call is a round-trip. Latency on very large files (10M+) isn't supported on the spreadsheet pathway — those workflows belong on a warehouse.
Can I see what data the AI looked at?
Yes. Each tool call shows the rows it touched and the columns it used. For aggregations, you can expand the trace to see the underlying detail. For filters, the output row count is visible — so "the AI summed 50 rows" is never confused with "the AI summed 5,000 rows".
Is there a free tier?
Yes. The free tier includes Ask Your Data with a daily question limit. Pro removes the limit and adds export, dashboard pinning, scheduled questions and audit logs. See pricing for details.
Can I use Ask Your Data on Shopify, Stripe, or HubSpot exports?
Yes — any of those tools' CSV exports work directly. We also have native Shopify and Google Sheets connectors that pull data on a schedule, so the file in Ask Your Data is always fresh.
Does the chat support charts?
Yes. Ask "plot revenue by month" and the AI picks an appropriate chart type, renders it inline, and offers options to pin it to a dashboard or include it in Auto Report. Chart type can be overridden if you want a specific format.
How is data handled in regulated industries?
UK/EU data residency by default. SOC 2 Type II in progress. Signed DPA available on request. Audit logs of every AI conversation, tool call and export. We're a fit for finance, professional services, healthcare back-office and regulated SMEs; we're not yet certified for HIPAA-PHI workloads.
Have a real conversation with your spreadsheet.
Drop in a CSV. Ask the questions you'd usually ask a junior analyst. Get answers with audit trails attached, in seconds.
See also: · AI spreadsheet analysis · AI Excel analysis · AI data analyst · Automated reports from Excel · · Home