ChatGPT for Business Analysis: What Works and What Breaks

ChatGPT is useful for business analysis when the work is language-heavy or the data is small and already in your hands: drafting requirements, structuring a problem, explaining a method or sketching an analysis plan. It struggles when the question needs live, complete or cross-tool data, because it only knows what you paste and cannot check its sums against your systems.

What it does well

Framing the problem. Ask it to turn “sales are down” into a list of testable questions: which segment, which period, which channel, versus what baseline. This is a strong use because you can judge the output immediately.

Explaining methods. Cohort analysis, funnel conversion, attribution models, confidence intervals: it explains them clearly and can tailor the explanation to a non-technical audience.

Drafting artefacts. User stories, survey questions, stakeholder emails, meeting summaries, process descriptions.

Writing formulas and queries. It can produce spreadsheet formulas or SQL from a description. Treat the result as a draft and read every line.

Brainstorming hypotheses. Given a symptom, it can suggest causes to investigate. Treat each as a hypothesis, not a finding.

Key takeaway: use it for thinking and writing, where you can judge the output; be careful when it produces numbers you cannot trace.

Where it breaks

Weakness	What happens	Mitigation
No live data	You paste a snapshot; it forgets next session	Use a tool that connects to sources
Partial context	Definitions, history and caveats are missing	Paste definitions with the data
Arithmetic and aggregation errors	Plausible but wrong totals on larger tables	Recompute key figures yourself
Confident narrative	Explains causes it cannot know	Ask “what evidence would confirm this?”
Data handling	Sensitive data leaves your systems	Follow policy; remove personal data
No cross-tool joins	Cannot match visitors to CRM to payments	Use a unified, connected setup

The cross-tool problem is the big one. A typical question such as “which channel brings customers who stay” needs analytics, CRM and payment data resolved to the same person. BusinessMCP’s guide to unifying analytics, CRM and ads describes why stitching those together is the hard part, and why dashboards often disagree.

Pasting data: a short checklist

1. Remove personal data you do not need. Replace names and emails with IDs.
2. State the definitions: what counts as a session, a lead, a customer.
3. Give the date range and time zone.
4. Ask the model to restate what it received before analysing, so you catch truncation.
5. Ask for the calculation steps, then recompute one total in a spreadsheet.
6. Keep a copy of the prompt and the answer, so you can reproduce it later.

The connected alternative

The alternative to pasting is an assistant that reads your sources directly. The BusinessMCP AI business analyst page frames the contrast directly: instead of pasting CSVs into a chatbot, you ask a question and the assistant, already connected to analytics, CRM, revenue, ads and search data, answers with the metric, timeframe and source. It also says it keeps a business memory distilled from conversations, which is the opposite of a blank slate each session.

The same page notes the assistant is model-agnostic: you can bring your own key for several providers or use the included free tier. The page does not state paid pricing, so confirm on the pricing page if that matters to you.

That does not make one tool “better” in all cases. The comparison is by task.

Task	ChatGPT-style chat	Connected analyst
Write a requirements document	Strong	Not the point
Explain a statistical concept	Strong	Not the point
Analyse a small CSV you hold	Reasonable, with checks	Possible, but overkill
“Why did conversions drop last week?”	Weak without data	Designed for it
Weekly metric summary across tools	Manual and repetitive	Natural fit
Draft outreach to a segment	Drafts text only	Can draft with approval gate

Prompts that improve results

• “List the assumptions you are making and which of them I should verify.”
• “Show the calculation for each total before giving the conclusion.”
• “Give three alternative explanations and the evidence that would distinguish them.”
• “What data would you need to be confident about this?”
• “Summarise in two sentences for a non-technical executive, then list caveats.”

These prompts push the model toward showing its working, which makes errors visible.

A safe division of labour

Use chat for the thinking and writing layer, a spreadsheet or BI tool for numbers you must audit, and a connected analyst when the question needs several live sources. Read AI tools for business analysts for a map of the options, and how far to trust AI data analysis for verification. If you are weighing dashboards against assistants, see AI business analyst vs BI dashboard.

One last point: settings and plans change. Confirm your provider’s current data-use terms yourself before putting company data into any tool.

A worked prompt, start to finish

Suppose you have a small export of last month’s orders and want a first look at what drives refunds. A careful prompt might read: “Here is a table of 300 orders with columns for product, channel, order value and refund flag. Definitions: refund flag is 1 when any part of the order was refunded. First, tell me how many rows you received and the overall refund rate. Then compare refund rate by channel and by product, show counts next to each percentage, and list two possible explanations for any large difference along with the data I would need to test them.”

Notice what this does. It makes the model confirm what it read, it forces counts alongside rates, and it asks for hypotheses rather than conclusions. You then recompute the overall rate in a spreadsheet. If it matches, you can trust the model’s slicing a little more; if not, you have found a problem cheaply. The same discipline applies whichever tool you use, which is why verification habits matter more than the choice of product.