Guide · AI for BAs

Business analyst AI tools — an honest 2026 guide

Categories of AI tools that show up in BA workflows, where each one earns its keep, and the questions to ask before you commit.

Every BA gets asked some version of "which AI tool should I use?" at least once a month now. The honest answer depends on what you are trying to achieve — there is no single tool that does all of BA work well, and several categories of tool serve genuinely different parts of the workflow. This is a practical map of the categories, where each shines, and the limitations to watch out for.

1. General-purpose LLMs (ChatGPT, Claude, Gemini)

What they do well: Brainstorming, prose polishing, explaining concepts, workshop-question generation, summarising a meeting transcript, talking through a tricky stakeholder situation. For early-stage thinking and short tasks, these are the right tool.

Where they break: When the deliverable is an artefact that has to survive senior review. A general LLM has no idea what your project is. It invents requirements, names plausible stakeholders, and produces an uncalibrated confidence level. The output reads well in isolation and falls apart in a Senior BA review. Acceptance criteria are fluent and wrong.

When to use: Early thinking, prose polish, one-off summaries, brainstorming.

2. Custom GPTs / instruction-tuned prompts

What they do well: A custom GPT (or a saved prompt template) consistently produces output in a specific style. For repetitive tasks where you control the input — formatting a transcript, generating workshop questions for a specific industry — they save real time.

Where they break: They do not solve the two hardest problems: grounding (the model still cannot index your discovery files) and structured output (you get prose, not editable structured data). The output is still a chat reply, not a workspace artefact.

When to use: Repeated formatting tasks. Personal productivity. Not for deliverables.

3. Generic AI documentation tools (Notion AI, Coda AI, etc.)

What they do well: Drafting wiki content, polishing internal notes, suggesting improvements inside a document. They are document-aware in the sense that they can see the page you are on.

Where they break: They have no opinion on what a BRD or business case should look like, no retrieval against your full project corpus (only the current page), and no structured artefact schemas. You can write a BRD in Notion AI — but the BRD is up to you, and the tool will happily produce a BRD with no acceptance criteria if you do not prompt for them.

When to use: If your team already lives in Notion and you want light AI assist on general content.

4. Specialised BA / PM AI workspaces (Reqify)

What they do well: Opinionated artefact schemas (BRD, FRD, business case, user stories, RAID, charter, more), retrieval-grounded generation against your discovery corpus, structured outputs you edit as data, integrations with Jira and Confluence, real document editor, version history, audit trails, data residency. The output is a deliverable, not a chat reply.

Where they break: They are not the right tool for casual brainstorming or general content. If you want to draft a Slack message or a marketing one-pager, use a general LLM. A BA workspace is for BA artefacts.

When to use: When the deliverable is a BA / PM artefact that has to land — BRD, FRD, business case, user-story backlog, RAID log, status report. See Reqify for BAs.

5. Workflow-automation AI (Zapier AI, n8n)

What they do well: Automating handoffs between tools. "When a Jira ticket is updated, summarise it and post to the steering Slack channel." Useful operational glue.

Where they break: They are not authoring tools. They cannot produce a BRD.

When to use: Operational automation, not artefact creation.

How to evaluate any BA AI tool

The honest evaluation questions:

Does it ground in my actual project files? Or does it produce confident-sounding output from a prompt alone? Demand to see a citation per requirement.
Is the output structured or just prose? Can you filter requirements, mark owners, push to Jira? Or is the output a Word file with no underlying data?
Does it flag gaps honestly? Senior reviewers notice when a model invents to avoid saying "I don't know". The tool should mark uncertainty inline.
Does it round-trip with delivery tools? Can it push to Jira, export to .docx with formatting, integrate with your existing PMO templates?
Is the data handling acceptable for your client / regulator? Data residency, no-training guarantees, audit logging, SSO.

The realistic stack for a senior BA in 2026

A general LLM (ChatGPT, Claude) for brainstorming, prose polish, workshop prep, stakeholder-conversation rehearsal.
A specialised BA workspace (Reqify) for the artefacts that have to land — BRD, FRD, business case, user stories, RAID, charter, status reports.
Workflow automation for the operational glue between systems.
Your delivery tools — Jira, Confluence, Excel, Word — integrated with the workspace so artefacts round-trip cleanly.

The mistake is trying to make one tool do all of it. The mistake on the other end is treating every artefact as a hand-written first draft when grounded generation can take the first 70% off your plate.