Turn policy and process truth into traceable requirements, with AI drafting only from approved sources.
The business analyst (BA) turns policy, process and domain truth into requirements a team can build and test. AI (artificial intelligence) can draft stories, compare documents and search code in minutes. That speed is the risk: generated text can look like a requirement without being one. In this seat the rule is simple: every requirement traces to an approved source and a human decision.
This page covers the analyst's decisions, ladder, evidence and hand-offs. The full due-date chain lives on the worked example. Counts and thresholds here are illustrative.
1. At a glance
Item
For the business analyst
Mandate
Turn obligations, process and domain rules into traceable requirements, decision tables, negative paths and testable acceptance criteria
Primary output
Requirements the product owner can accept and the QA engineer (quality assurance) can prove
First safe win
Stage 1: ask an approved assistant to explain an API (application programming interface) document, then write the questions for its owner
Biggest failure mode
Laundering: pasting generated text into a requirement without a source, so nobody can tell what is true and who decided it
Line: 1st line · Family: Shape · Variants: business analyst (process and policy), systems analyst (interfaces and data)
Variant
What differs
Business analyst
Works from policy, contracts and the current process. Owns decision tables, process maps and the exception list
Systems analyst
Works from interfaces, data definitions and service maps. Owns field definitions and the consumer list in impact analysis
2. Your journey: first steps to fully adopted
The analyst moves from asking the assistant to explain documents to running shared, sourced drafting; each stage adds controls and evidence, never the right for AI to decide what a requirement says.
Start here: your first steps
List the approved sources for your area and confirm each one with its owner.
Ask the approved assistant to explain one API document, then write the questions for its owner.
Write the team glossary rule into the story prompt: the due date is date-only; the cutoff is an instant.
Run the reviewed story prompt on one real ask, and log every rejected line with its reason.
Fully adopted: how the work runs
Moment in the work
Before AI
Fully adopted (Stage 4–5)
Decision that stays human
Requirement
Stories typed by hand from long documents
An agent drafts stories, criteria and decision tables from the brief and the contract; each claim cites its source
What the requirement says
Source conflict
Conflicts surface late, in test or in production
An agent scans the sources and lists each contradiction with both citations
Which source wins: the source owner decides, never the agent
Negative paths
Edge cases depend on memory
A shared negative-path library and evaluation set; the agent proposes cases, you add the missing ones
Which exceptions are in scope (the product owner accepts)
Impact analysis
Manual search across services
A read-only agent drafts the impact analysis, every row linked to evidence; the architect confirms
That each finding is true before it enters the record
Policy change after release
The change reaches the team by email, if at all
You are notified; an agent lists the affected criteria with citations
What the updated requirement says
What changes for the organization
Before
After
Requirements without sources; nobody knows who decided
Every accepted criterion traces to an approved source and a named role
Missing edge cases escape to production
Negative paths are reused across teams and proven by tests
Refinement time goes to typing
Refinement time goes to contradictions, sources and decisions
In any organization
Who covers it: in a small or unregulated organization, the product owner, a lead developer or a tester writes requirements; the source check still applies.
What scales: formal three amigos sessions and evaluation sets can shrink to a second reader and a list of past mistakes. What never drops: AI never settles a contradiction, no customer data goes into unapproved tools, and every requirement has a source and a human decider.
3. Your decisions
Decision relationships on this site: owner or approver, independent validator or challenger, consulted control partner, evidence recipient, notified.
Rejection rate rises; a missing negative path escapes to test
4 · Agentic
An agent drafts the impact analysis from catalog and code search, inside the front-loaded pack
Read-only access; every finding links its evidence
Impact analysis and consumer list
Owner of draft; architect confirms
A finding without evidence; a service the map does not know
5 · Scale
Maintains the shared story prompt and a negative-path library other teams reuse
Material-change review when prompt or model changes
Catalog entries with owner and limits
Contributor; catalog owner approves
Periodic review finds drift in draft quality
Stays human at Stage 5: resolving a contradiction between two sources, and deciding what the requirement says. An agent can find the conflict; the source owner and you settle it.
For the Stage 4 shape, where one owner front-loads the artifacts before the first meeting, see ask-to-release.
6. Priority scenarios
Scenario
Value
Risk tier
Stage
First step
Do not use AI when…
Explain an interface or policy
Faster domain learning
Low to Medium
1
Ask, then confirm with the owner
The document is not in an approved source
Current-state process map
Finds the real steps and exceptions
Medium
1 to 2
Draft from walk notes; walk it again with the team
The ask is not approved or the glossary is missing
Decision table from policy
Fewer missed cases
Medium
3
Agent drafts rows with citations; you check each row
The policy is ambiguous; ask the policy owner first
Contradiction scan across sources
Catches conflicts early
Medium
3
Agent lists conflicts with both citations
You want it to choose the answer: it must not
Impact analysis
Finds every consumer
Medium
4
Agent searches catalog and code; architect confirms
It has write access, or findings have no evidence
7. The due-date ask from your seat
The analyst holds the hardest part of this ask: the domain. Full narrative below; every artifact is filled on the worked example.
Step
Input received
Your work
Output and evidence
Recipient
Sources
The ask; the contract clause
Name payment-schedule-service as the source of due date and cutoff; profile-service for the zone
Source list, confirmed by the architect
Architect (confirms)
Domain truth
Assistant explanation of date-only value versus instant
Write the rule in the team glossary: the due date is never converted; the cutoff is an instant shown in the customer's IANA (Internet Assigned Numbers Authority) time zone
Glossary entry; story prompt context
Team; product owner
Current state
Code search results for ten date implementations
Record which three treat the due date as an instant
Current-state table
Developer, architect
Stories
Agent drafts of US-1 to US-5
Reject a draft of US-3 that used a fixed offset; add the case to the evaluation set
Corrected stories, rejection noted
Product owner (accepts)
Negative paths
Edge-case list from QA
Add daylight saving days, missing zone with the fallback label, issued statements
Criteria US-3, US-4, US-5
QA engineer (evidence recipient)
Impact
Agent-drafted impact analysis
Check each row links code or catalog evidence; mark six other consumers as "not in this change"
Impact analysis
Architect (confirms)
Decision table · cutoff display (the analyst's own artifact)
Profile zone
State of residence
Shown cutoff
Label
Present
Any
The cutoff instant in the profile zone
Zone abbreviation
Missing
Present
The cutoff instant in the state's zone
"based on your state of residence"
Missing
Missing
Open question to the product owner
Not shown until decided
The last row is the point. When the sources do not answer, the analyst writes an open question. The agent does not fill the gap.
Agents draft more than you can read. Limit drafts per refinement to what you can check line by line
Owner time
Faster drafts mean more questions to source owners. Agree their response time with the product owner
Role shift
Less typing, more judgment: source choice, contradiction routing, negative paths. Plan learning time with the enablement lead
13. Anti-patterns and stop conditions
Anti-pattern or trigger
Why it hurts
What you do
Laundered text
Requirement with no source and no decider
Delete it or cite it; record who decided
AI picks the winner
A contradiction resolved by a model, not an owner
Route both citations to the owner
Automating before mapping
Faster broken process
Walk the current state first
Unapproved source
Requirement built on a draft or a chat thread
Replace with the approved version or raise an open question
Material change (prompt, model, glossary)
Drafts change quality silently
Re-run the evaluation set; step down to Stage 2 until it passes
Customer data in a prompt
A policy breach
Stop; report the near miss; use synthetic data
14. Action kit and 30/60/90
## Business analyst · requirement source check · v1
For each criterion:
1. Source: document, version, section (approved source only).
2. Rule type: date-only value, instant, zone, business calendar, or other.
3. Negative paths: daylight saving day, missing data, issued record, holiday.
4. Contradiction: none, or both citations sent to the owner on <date>.
5. Decided by: name of the role, not the tool.
6. Drafted by AI: yes or no; trace link if yes.
7. Test link: the QA case that proves it.
Stop if any criterion has no source, or a contradiction was settled by the draft.
Questions to ask your enablement lead:
Which documents count as approved sources for my domain, and where is that list kept?
Who owns the team glossary, and how is a change to it reviewed?
How do I add a rejected AI draft to the evaluation set so the next run improves?
Window
Goals
Done when
First 30 days
Stage 0 and 1: source list for your area; assistant for explanations only
Acceptable use attested; source list confirmed by owners
Days 31 to 60
Stage 2: run the reviewed story prompt on two real asks
Drafts accepted after edits; rejected lines logged with reasons
Days 61 to 90
Stage 3: delegated drafting of decision tables with citations
Evaluation set has your negative cases; traceability sample passes