BA Kit: from manual BA work to BAs directing AI

The problem

The Dev team was already using AI and working very fast. The BA team still did everything by hand: taking notes on what the client said in meetings, distilling them afterwards, analysing, then creating tasks. Analysis and testing became the bottleneck: Dev waited for requirements, the client waited for answers.

Before BA KitThe bottleneck sits in the middle

1. 1Client commentsin the meeting
2. 2BA works by handnotes, distilling, analysis, tasks
3. 3Dev uses AIworks very fast, then waits

Context and role

• A BA team of three, serving several projects in parallel.
• Project knowledge was scattered across documents, chat messages and people’s heads.
• There was no shared standard for using AI. Everyone prompted in their own way, so the results were hard to trust.
• I proposed, designed, built, trialled and trained, alongside my BA work on the project.
• Constraints: AI may only read the dev environment; every AI output must be verified by a person.

Step 1: Doing each step by hand and assessing it

Before automating, I did each step of the process manually and rated which steps took the most time and which were repeated.

Process assessmentWhich steps AI should draft

• AI drafts, BA reviews
• AI only gathers data, BA decides

	Time cost	Repetition	Judgement needed	AI's part
Taking meeting notes	High	High	Low	Draft
Distilling client comments into requirements	High	High	Medium	Draft
Looking through the codebase and database to assess impact	High	High	High	Data only
Writing user stories and acceptance criteria	Medium	High	Medium	Draft
Creating Jira tickets, assigning epics and labels	Medium	Very high	Low	Draft
Drawing interfaces	High	Medium	Medium	Draft
Drawing flow diagrams	Medium	High	Medium	Draft
Preparing UAT documents	Medium	High	Low	Draft

Step 2: Building the project knowledge vault

AI is only useful with the right context. Each project’s knowledge lives in an Obsidian vault with wikilinks that agents can read: meetings are recorded, transcribed and linked both ways to the requirements that came out of them.

Vault structureOne vault per project, everything points to requirements

Step 3: Building and refining BA Kit

The processes are packaged into skills for IDE agents (Claude Code, Codex, Antigravity), connected to Jira, the database and the vault through MCP. The kit is improved after each real use and now has 30 skills.

BA KitRepresentative skill groups

	Draft produced by AI	Verified by
Extract requirements from a meeting	Requirements, open questions, action items	BA
Impact assessment	Affected components, risks	BA, re-reading is mandatory
Write stories	User stories, acceptance criteria	BA
Draft Jira tickets	Ticket in the right template, epic, labels	BA clicks create
Draw interfaces	Mockup	BA
Draw diagrams	Flow diagram, system diagram	BA
Prepare UAT	UAT checklist by role	BA
Daily work pack	Pre-meeting summary, end-of-day report	BA

RulesApplied to every skill

• Read only the dev or staging database, never production.
• Do not create tickets autonomously. AI produces a draft; a person clicks to confirm.
• Every conclusion must cite its source: file, transcript, timestamp.
• Do not guess business rules. If one is missing, generate a question for the client.

Step 4: One transcript going through BA Kit

A simulated example: what AI does and what the BA has to fix.

CPClient PM

“…and for the report, sales just wants a list of who’s about to churn, maybe we email it to them every morning. Oh and it should exclude trial accounts, those don’t count.”

REQ-014AI draft

Churn-risk list for Sales

• Daily email to the Sales team every morning
• Exclude trial accounts
• Open questions: What defines “about to churn”? Which timezone is “morning”? Email only, or also visible in CRM?

Source

Weekly client meeting

Suggested links

BR-07, US-04

BA verificationThree points in the draft

• AI suggested reusing BR-07.Correct, kept as is.
• AI missed a stakeholder mentioned in an earlier meeting.Added a question for the client.
• AI treated email as the main channel.Wrong priority, since a CRM sync already exists. Corrected.

Step 5: Training and standardising the new process

I improved the process together with the BAs in the team, then trained them to use it.

New process

• BA
• AI
• Dev

1. 1BAMeetingrecorded
2. 2AITranscriptinto the vault
3. 3AIExtract requirements
4. 4BAReview, ask the client
5. 5AIDraft stories, interfaces, diagrams
6. ?BAApprove
7. 7DevReceive ticket

Step 6: From BA Kit to a kit for the whole IT department

BA Kit then became the core of Software KIT: one kit per role in the IT department, sharing one specification format and one set of rules. Figures counted from the kit repository, 07–09/2026.

Software KIT40 self-written skills, by role

• BA
• Tester
• PM
• DevOps

• 40 skillsplus 18 subagents and 17 shared rules

  30721

One specification format for BA, Dev, Tester and client. Each requirement is a Requirement + Scenario block with an ID (FR-, BR-, NFR-); changes are written as deltas (added, modified, removed). An automatic check catches duplicate IDs, requirements without scenarios and leftover placeholders.

Changes driven by measurementTwo numbers that changed the way we work

• Still had to ask
• No need to ask

• Tasks from bulk AI analysis, first roundDev came back with questions on about 90 of 100

  9010

• 78 open questions, after one sweep23 closed from existing knowledge

  5523

Preview gateEvery external write

1. 1AI preparestickets, client reports
2. 2Preview
3. ?A person confirms
4. 4Then it runs

Results

By internal measurement on the processes already adopted:

• Time spent by BAs on repetitive tasks went down by about 70%.
• BA team productivity went up about 3 times.
• Analysis is no longer the bottleneck relative to the speed of the Dev team.
• The BA team moved from manual work to directing AI: letting AI read the codebase and dev database, applying project rules, drawing interfaces and diagrams, recording meetings and extracting transcripts.

What I learned

• Do it by hand before automating. Otherwise you automate the wrong step.
• Context matters more than the prompt. Without a knowledge vault, every skill has to guess.
• Control rules do not slow things down. They are what makes the team willing to trust and use AI every day.