i

Description

Machine mécanique à prédire les marées, faite de poulies, de câbles et de cadrans en laiton.

Source

Steven Fine · original page

Licence

CC0 Public domain

Original photo, unretouched. The background is a version of it stylised by the site.

AI in Planning and Version Control Tools

You will know where AI truly delivers value in tracking tools (writing and breaking down stories, summaries, deduplication, estimation, risk detection) and in version control…

In Brief

• You will know where AI truly delivers value in tracking tools (writing and breaking down stories, summaries, deduplication, estimation, risk detection) and in version control (code review, commit messages, conflict resolution).
• You will have a gain / risk / maturity /5 table for each use case.
• You will know how to set guardrails (confidentiality, hallucinations, accountability, hidden debt) and write useful prompts.
• AI features evolve very quickly: any dated element “state 2026-10” must be rechecked before making a decision.

Context: team of 5 to 15 people, data/BI project (SQL Server datamart, Power BI reports, TFVC/Azure DevOps Server). Ratings /5 = contextual assessments (RATING). For project management in general, see AI in Service of Project Management (General Then Data); for the cross-functional summary, Cross-Functional Summary — AI Contribution.

On this page

1. 🟢 Novice — What “AI” means here
2. 🟢 Novice — Where to find AI today (state 2026-10)
3. 🟡 Practical — Use cases in planning tools
   1. Writing and improving stories
   2. Work breakdown
   3. Ticket and thread summaries
   4. Deduplication and semantic search
   5. Estimation
   6. Risk and anomaly detection
   7. Triage and classification
   8. Writing reports, release notes, and meeting notes
4. 🟡 Practical — AI and version control
   1. SQL and BI: specifics
5. 🔴 Expert — Summary table: gain / risk / maturity
6. 🔴 Expert — Guardrails
   1. Confidentiality and compliance
   2. Hallucinations and verification
   3. Responsibility and skills
   4. Progressive rollout
7. 🔴 Expert — What AI does not replace
8. See also

Level: 🟢→🔴

🟢 Novice — What “AI” means here#

In planning tools, “AI” almost always refers to a language model (LLM) connected to the tool's data (tickets, comments, wiki, code). It can rephrase, summarize, classify, propose a draft, search by meaning. It can neither decide, nor guarantee accuracy, nor know your team's unwritten context. An AI draft that is reviewed is a gain; an AI draft that is not reviewed is debt.

Three levels of autonomy:

🟢 Novice — Where to find AI today (state 2026-10)#

🟡 Practical — Use cases in planning tools#

Writing and improving stories#

AI turns an informal request (“the North region numbers are wrong”) into a structured story (context, value, acceptance criteria). Real gain on form, risk on substance: it fills gaps with plausible assumptions.

Typical prompt:

You are a BI analyst. Here is a user's raw request: “{text}”.
Context: SQL Server datamart, Power BI reports, measures in euros, offset fiscal calendar.
Produce: (1) a story in the format “As a… I want… so that…”, (2) 5 acceptance criteria
that are verifiable with numeric thresholds, (3) the list of QUESTIONS to ask the requester before starting,
(4) the assumptions you made. Do not invent any business rule: mark it as to be confirmed.

Work breakdown#

AI proposes breaking an epic into stories or a story into tasks. It is good at listing, bad at sorting by value and at respecting real constraints (source dependencies, delivery windows). Apply the INVEST test to each proposal (see Scrum in Practice); reject “layer-based” breakdowns (one story for the database, one for the report) that do not deliver standalone value.

Ticket and thread summaries#

Summarizing 80 comments into 5 lines, decision changes, people involved. High gain, moderate risk: the summary omits what seems secondary. Use it to orient yourself, not to decide; keep the link to the original.

Deduplication and semantic search#

Meaning-based search finds similar tickets despite different wording. Deduplication is suggested, never applied on its own: two similar tickets may have different conditions.

Estimation#

AI can compare a ticket to history (“6 similar tickets took between 2 and 5 days”). It does not replace team estimation (planning poker is mainly used to surface gaps in understanding). Use AI to provide references, not a number.

Risk and anomaly detection#

Flagging aging tickets, unresolved dependencies, overcommitted sprints, incident keywords. Useful as a safety net, with explainable thresholds; opaque alerts are ignored after a week.

Triage and classification#

Labeling, routing, assessing the criticality of a support request. Strong gain at high volume; check a weekly sample and track the error rate.

Writing reports, release notes, and meeting notes#

From tickets linked to a release package, AI writes readable release notes. Verify facts (numbers, dates, scope): this is formatting, not analysis.

🟡 Practical — AI and version control#

SQL and BI: specifics#

• T-SQL / DAX generated: syntactically correct, sometimes semantically wrong (join that duplicates, misunderstood filter context). Always replay on a reference figures dataset.
• SQL script review: explicitly ask to verify idempotence, transactions, missing indexes, NULL in joins, and implicit conversions.
• Sensitive data: never paste real customer data into an unapproved service; anonymize or work on the schema only.

Example review prompt:

Review the T-SQL script below (SQL Server, compatibility level 130).
Check: idempotence (re-runnable without side effects), error handling, joins that
duplicate rows, behavior with NULL, potential locks, qualified object names.
Respond as a table: line | issue | severity | proposed fix.
Do not rewrite the entire script.

🔴 Expert — Summary table: gain / risk / maturity#

Scales: gain (time or quality improvements), risk (impact of an undetected error), maturity (current reliability of the function). Context: team of 5–15, data/BI. Expert assessments, not measured.

🔴 Expert — Guardrails#

Confidentiality and compliance#

• Classify data: what can be shared (generic ticket titles) and what cannot (customer names, personal data, secrets, queries containing connection strings).
• Choose the contractual scope: does the tool use your data to train its models? Where is it processed (region)? Retention? Check this in the provider's terms and validate with the DPO; GDPR: legal basis, informing individuals, minimization.
• Log: who used AI to produce what (especially for agents).

Hallucinations and verification#

Responsibility and skills#

• Responsibility remains human: the name in the changeset or the acceptance testing sign-off commits the person, not the tool.
• Hidden debt: code or text produced quickly but not understood becomes a cost later; enforce “I can explain what I accepted”.
• Junior skills: AI can short-circuit learning. Plan exercises without assistance and reviews focused on “why.”
• Dependency: keep the ability to work without the tool (outages, provider changes, costs).

Progressive rollout#

1. Frame: use cases, authorized data, owner.
2. Experiment on a limited scope, with quantified success criteria (time saved, error rate, satisfaction).
3. Measure over 4 to 6 weeks; compare with a period without AI.
4. Decide: expand, adjust, or stop.
5. Maintain: quarterly review (functions, cost, incidents, compliance).

🔴 Expert — What AI does not replace#

• Priority judgment: arbitrating between two competing values entails product responsibility.
• The relationship with stakeholders: understanding what a user says (and does not say).
• Business acceptance testing: deciding that a number is correct requires business knowledge; see acceptance testing and validation (matrix).
• Architecture decisions: AI suggests options; the trade-off depends on context, accepted debt, and organization.
• Team trust: a retrospective written by a machine is not a retrospective.

See also#

• Overview · Azure DevOps Server (ex-TFS) in Depth · Jira In Depth and Alternatives · Comparison of Planning Tools
• AI in Service of Project Management (General Then Data) · Pitfalls and good practices in project management (cross-functional summary) · Cross-Functional Summary — AI Contribution
• VCS and SQL/BI data · Git in Depth · Glossary

✻ Written by Claude Sonnet 5.5