Agile & Tools
How Business Analysts Use AI Tools Like ChatGPT, Safely and Well
Part of our guide to business analyst tools.
Where AI helps a business analyst, and where it does not
AI chat tools such as ChatGPT are language tools. They are very good at producing text that sounds right. They do not know your company, your stakeholders or your systems unless you tell them. Even then, they can fill gaps with confident guesses. That one fact decides where they help a business analyst and where they get in the way.
Tasks where AI usually helps
- First drafts of user stories, acceptance criteria, emails and agendas, which you then rewrite.
- Summaries of long notes or documents you are allowed to share, so you can find the key points faster.
- Test ideas and edge cases: asking what could go wrong with a rule often surfaces boundaries you missed.
- Question lists for interviews and workshops, as a starting point for your own elicitation plan.
- Plain-English rewrites of technical text. This is useful if English is your second language or your audience is non-technical.
- Explaining unfamiliar terms, such as an API concept or a SQL function, before you check them in official documentation.
Tasks where AI does not help, or is risky
- Discovering what stakeholders actually need. Only real conversations do that.
- Confirming business rules, limits and policies. AI will invent a plausible number if the real one is missing.
- Making priority or scope decisions. Those belong to the product owner and the sponsor.
- Interpreting laws or regulations. Rules vary by country and industry, so use official sources and your legal team.
- Sign-off, approvals and anything else that needs a named, accountable person.
A simple way to sort any task is to ask two questions. How much time could AI save here? And how bad would it be if the output were wrong? The matrix below shows where common BA tasks usually land.
Diagram in words
- Good fit (top left): Rewording an email for a non-technical reader; Listing interview questions to adapt; Summarising your own anonymised notes
- Use with strict review (top right): Drafting acceptance criteria; Suggesting test cases and edge cases; Drafting a SQL query to check
- Optional (bottom left): Fixing grammar in a short message; Suggesting headings for a page
- Do it yourself (bottom right): Confirming business rules and limits; Prioritising the backlog; Recommending sign-off
Data privacy rules before you paste anything
This is the most important part of the article. Many organisations have a policy on AI tools, and some block public tools completely. Before you paste even one line, find out what yours allows. This is general guidance, not legal advice. Data protection rules differ between countries and sectors, so check with your company's data protection, security or legal team.
As a default, treat any text you paste into a public AI tool as text you have shared outside the company. Then apply these rules.
- Use only approved tools. If your company provides an enterprise AI tool, use that, not a personal account.
- Never paste personal data: customer names, phone numbers, emails, account numbers, addresses or employee details.
- Never paste secrets: passwords, API keys, tokens, connection strings or internal system addresses.
- Remove client and commercial details: client names, contract terms, unreleased financial figures and pricing.
- Anonymise first. Replace real names with roles (Customer A, Ops lead) and real values with placeholders.
- Check the tool settings for chat history and whether your inputs may be used for training, and follow your policy on them.
- When in doubt, ask. A short message to your manager costs far less than a data incident.
Before you paste: a quick privacy check
- The tool is approved by my organisation for work use
- The text has no customer, employee or patient personal data
- The text has no passwords, keys, tokens or internal system addresses
- Client names, contract terms and unreleased figures are removed or replaced
- Real names are replaced with roles such as Customer A or Ops lead
- I know whether this tool keeps my chat history and how my policy treats that
How to use AI on a BA task, step by step
The pattern below works for most BA tasks, from a meeting summary to a set of acceptance criteria. The key idea is simple: AI writes the draft, you own the result.
- Check the policy and the tool. Confirm the tool is approved and decide what you may share.
- Prepare clean input. Collect the confirmed facts (rules, notes, the user story) and anonymise them.
- Give context, rules and a format. Say who the output is for, what the feature does and what shape you want, for example Given/When/Then.
- Tell it not to guess. Ask it to list missing information as questions instead of filling gaps. It will not always obey, but it helps.
- Ask for a draft, not a decision. Use AI for options and wording. Keep choices about scope and priority with the people who own them.
- Review line by line. Compare every statement with your source. The review section below explains how.
- Confirm with people. Walk the subject matter expert, developer or tester through the corrected version before it goes into the backlog.
TEXT
Role: You are helping a business analyst draft acceptance criteria.
Context: [one or two sentences about the product and its users]
User story: [paste the story]
Confirmed rules: [paste only rules that stakeholders have confirmed]
Task: Write Given/When/Then scenarios. Cover the main path,
each boundary value and each failure case.
Constraints: Use only the rules above. Do not add features.
If information is missing, list it under Open questions
instead of guessing.A template like this does not make the output correct. It makes the output easier to check, because every line should map back to something you supplied. If you are new to the format, read acceptance criteria and Gherkin examples first so you can judge the draft with confidence.
Diagram in words
- 1. Check policy: Confirm the tool is approved and what data you may share.
- 2. Clean the input: Remove personal data and secrets, and replace names with roles.
- 3. Prompt with rules: Give context, confirmed rules, a format and a no-guessing instruction.
- 4. Review the draft: Compare each line with the source and mark invented or vague text.
- 5. Confirm with people: Walk through the corrected version with the expert, developer and tester.
- 6. Record and reuse: Store the final version in your normal tool and log open questions.
Worked example: drafting and correcting acceptance criteria
Here is the method applied to a realistic task. Marigold Grocers is a fictional online grocery company. The team is adding a feature so customers can request a refund for an item missing from their delivery. The user story is: As a customer, I want to request a refund for a missing item so that I do not pay for something I did not receive.
Sunil, the operations lead, has confirmed four rules in his notes:
- A customer can request a refund at most 48 hours after delivery.
- If the total refund value for the order is ₹500 or less, including tax, it is approved automatically.
- If the total refund value for the order is over ₹500, including tax, the request goes to the support agent queue for a person to review.
- After 48 hours, the app does not accept the request and asks the customer to contact support.
On your first attempt you paste only the user story, with none of these rules, and ask for acceptance criteria. The AI returns a tidy draft with two scenarios. It looks finished. It is not. Because it had no rules, it made up its own time window and its own limit.
The first AI draft, with review notes
Marigold Grocers: missing item refund (AI draft)
Scenario: Customer requests a refund
Given a customer has a delivered order
When they report a missing item within 24 hours1
Then the refund is approved quickly2
And the customer receives an SMS confirmation3
Scenario: Large refund
Given the refund value is over ₹1,0004
When the customer submits the refund request
Then the refund request is rejected5
- Invented time window. No window was given, so the AI guessed 24 hours. The confirmed rule is at most 48 hours after delivery.
- Vague outcome. Quickly cannot be tested. The real outcome is that the refund is approved automatically and the order status changes.
- Unrequested scope. Nobody asked for SMS. Adding it would create developer work that nobody has approved.
- Invented limit. No limit was given, so the AI guessed ₹1,000. The confirmed limit is ₹500 per order, including tax.
- Wrong outcome. Refunds over the limit are not rejected. They go to the support agent queue for a person to review.
Two more gaps cannot be pinned, because they are things the draft leaves out. There is no scenario for a refund of exactly ₹500, and no scenario for a request made after 48 hours. So the review finds five pinned problems and two missing scenarios.
You take the draft to a short review with Nadia, the business analyst leading the project, and Sunil.
Reviewing the AI draft
Marigold Grocers, a fictional online grocery company: a fifteen-minute review of AI-drafted refund criteria before they go into the backlog.
- You The BA learner
I drafted these with AI from the user story only, with no rules. I checked first that the story had no customer data in it.
- Nadia Business analyst
Good start on privacy. Now we check each line against Sunil's notes. Sunil, the draft says 24 hours. Is that right?
- Sunil Operations lead
No. Customers get 48 hours after delivery. The ₹1,000 figure is wrong too. The limit is ₹500 per order, including tax.
- Nadia Business analyst
The draft also rejects large refunds. Is that the rule?
- Sunil Operations lead
Never. Anything over ₹500 goes to the support agent queue. ₹500 or less, including exactly ₹500, is approved automatically.
- You The BA learner
It also says quickly and adds an SMS. And it has no scenario for exactly ₹500 or for a late request.
- Sunil Operations lead
SMS is not in this release. After 48 hours the app should ask them to contact support.
- Nadia Business analyst
Log the SMS idea as an open question for Helena. Then rerun the prompt with the four rules and edit the output yourself.
The people and the company in this scene are fictional.
You run a second prompt that includes the four confirmed rules and the instruction not to guess. Then you edit the result by hand. The corrected criteria look like this:
GHERKIN
Feature: Refund for missing items
Scenario: Small refund is approved automatically
Given a delivered order
And the customer reports a missing item 10 hours after delivery
And the total refund value for the order is ₹300 including tax
When the customer submits the refund request
Then the refund is approved automatically
And the order status shows Refund approved
Scenario: Refund of exactly the limit is approved automatically
Given the total refund value for the order is ₹500 including tax
When the customer submits the refund request within 48 hours of delivery
Then the refund is approved automatically
Scenario: Refund above the limit goes to an agent
Given the total refund value for the order is ₹650 including tax
When the customer submits the refund request within 48 hours of delivery
Then the request is added to the support agent queue
And the order status shows Under review
Scenario: Request after the window
Given the customer reports a missing item 50 hours after delivery
When the customer tries to submit a refund request
Then the request is not submitted
And the app shows a message asking the customer to contact supportEvery rule now traces back to Sunil's notes, and every scenario has a Given, a When and a Then. The sample values sit clearly on one side of each rule: 10 hours is inside the window and 50 hours is outside it, ₹300 is under the limit and ₹650 is over it. The ₹500 boundary has its own scenario, the vague word quickly is gone, and each Then step is something a tester can see. A tester may also add a request at exactly 48 hours; because the rule says at most 48 hours, it is accepted.
The SMS idea is logged as an open question for Helena, the sponsor, instead of slipping into scope. The AI saved typing. The analysis came from you and Sunil. From here, turning each scenario into tests is straightforward: see writing test cases from acceptance criteria.
How to review AI output before anyone else sees it
Reviewing AI output is a skill in itself. The draft is fluent, so mistakes hide well. Read it as if a new colleague wrote it, someone who has never met your stakeholders and has never seen your systems.
- Trace every fact. Each rule, number, field name and status must come from a source you can point to: notes, an approved document or a stakeholder. If you cannot trace it, delete it or turn it into a question.
- Hunt for invented detail. Look for features, systems, roles or messages nobody mentioned. AI often adds sensible-sounding extras.
- Test the boundaries. For each limit, check the value just below, exactly at and just above it. Missing boundaries are a very common gap.
- Look for what is missing. Check that failures, late actions, missing data and permission problems are covered, not only the lines that are there.
- Remove vague words. Words like quickly, easily, user-friendly and appropriate cannot be tested. Replace them with observable results.
- Check terms and consistency. Names should match your glossary and the rest of the backlog, and no rule should contradict another.
- Check the format. For stories, apply INVEST, as explained in how to write user stories. For criteria, each scenario needs a clear action and each Then step should be visible to a tester.
- Get a second pair of eyes. A developer and a tester will spot problems you cannot, especially in technical details.
If the review takes longer than writing it yourself would have, that is useful information. Some tasks are simply faster without AI, and that is fine.
Common mistakes when business analysts use AI
Most problems with AI at work come from habits, not from the tool itself. Watch for these.
- Pasting first, thinking later. Real customer data or client documents go into a public tool before anyone checks the policy. Fix: run the privacy check every time.
- Treating fluent text as correct text. Polished wording hides wrong numbers and rules. Fix: trace every fact to a source.
- Asking without rules. A prompt with only a user story invites the AI to invent limits. Fix: supply confirmed rules and ask for open questions.
- Only reviewing what is written. Missing scenarios do not stand out on the page. Fix: list the boundaries and failure cases yourself, then check each one is covered.
- Letting AI set scope. Extra features slip in because they sound sensible. Fix: anything new becomes a question for the product owner or sponsor.
- Using AI instead of talking to people. A list of likely requirements is not elicitation. Fix: use AI to prepare questions, then ask real stakeholders.
- Copying technical answers without checking. SQL, API details and formulas can look right and be wrong. Fix: test on sample data and check official documentation.
- Hiding AI use when your team expects openness. Fix: follow your team's norms on disclosure and be ready to explain how you checked the output.
Where AI fits in your wider BA toolkit
AI does not replace the core tools and skills of the job. You still need to run interviews, model processes, write clear requirements, read data with SQL and manage work in tools like Jira. AI sits on top of those skills as a drafting assistant, and it is only as good as your ability to judge what it produces. If you are deciding what to learn first, our guide to business analyst tools puts AI in that wider picture.
The best way to build that judgement is to practise reviewing drafts against real rules. In the BA Lab you work through elicitation, requirements and UAT stages on a fictional project. That gives you realistic material to check AI output against, without any risk to real data.
Frequently asked questions
Is it safe to use ChatGPT for business analysis work?
It can be, if your organisation allows it and you follow its policy. Use approved tools, never paste personal data, secrets or confidential client details, and anonymise text first. Data protection rules vary by country and sector, so check with your security or legal team if you are unsure.
Will AI tools replace business analysts?
AI can speed up writing and summarising, but the core of the role is understanding people, problems and trade-offs. Someone still has to ask the right questions, confirm rules, resolve conflicts and take responsibility for what goes into the backlog. Analysts who use AI well and review it carefully tend to find it a helpful assistant.
Can I use AI to prepare for a business analysis certification exam?
AI can explain concepts and quiz you, but it can be wrong or out of date. Use the official materials from the certifying body as your main source, and check its current exam rules directly, including what you may and may not use during the exam.
Should I tell my team that I used AI to draft a document?
Follow your team's norms. Many teams accept AI drafts as long as the author has reviewed them and can explain every line. Being open is usually the safer choice, and it helps reviewers know which parts to read closely.
Which AI tool is best for business analysts?
The best tool is the one your organisation has approved for work data. Many workplace suites now include AI assistants inside documents, chat and ticketing tools. The privacy and review habits in this article matter more than the brand you use.
This article was written by an AI system and published automatically after automated accuracy, safety and quality checks. All companies, people and numbers in the examples are fictional. BA Mentorship does not issue professional certifications and cannot guarantee any job or interview outcome. Spot an error? Email our support team and we will check and correct it.