Cross-Functional Summary — AI Contribution

This page consolidates what generative AI adds, and does not replace, across the four topics in this section: project management, tools, maturity matrices, Power BI reporting.

Guide≈ 10 min readCheckedReading

In brief

This page consolidates what generative AI adds, and does not replace, across the four topics in this section: project management, tools, maturity matrices, Power BI reporting.
It provides a consolidated table (gain /5, risk /5, maturity /5), a governance framework (privacy, traceability, sign-off), and a four-stage adoption roadmap.
The ratings are contextual expert assessments (team of 5 to 10 people, BI/datamart development, use with review): see RATING. Product features (Copilot, Fabric, Azure DevOps) evolve quickly: any feature mentioned should be rechecked at the source (status 2026-10).

On this page

The principle: speed up without delegating responsibility
By topic: key takeaways
Consolidated table: gain, risk, maturity
Governance of AI use
Confidentiality
Traceability and content signatures
What AI does not replace
Adoption roadmap
See also

Level: 🟢→🔴

The principle: speed up without delegating responsibility#

🟢 Novice. A generative AI assistant (language model) produces text, code, queries, and summaries from a prompt. It is fast, inconsistent, and confident. Three ideas are enough to use it well:

It proposes, you decide. The output is a draft to review, never a decision.
It only knows what it is given (and what it has learned). It does not know your context, conventions, or data — unless you provide them, which raises the privacy question.
It can be confidently wrong (“hallucination”): nonexistent function, plausible but wrong DAX measure, invented reference. The more the output is mechanically verifiable (an executable query, a passing test), the safer AI is to use.

🟡 In practice — the decision rule. Classify a use case with two questions: is the result verifiable at low cost? and is the error reversible? Verifiable and reversible (a test query, a script skeleton): go ahead. Unverifiable or irreversible (a published number, a hiring decision, data deletion): AI assists, humans decide and prove.

🧪 Example

On a SQL Server datamart fed by a sales force application, asking AI for a reconciliation query between a report total and the fact table is a good use case: you run it, then compare. Asking it “what is our quarterly revenue?” from a report summary is a bad one: the number is not verifiable from the answer alone.

By topic: key takeaways#

Project management#

Details in AI and Project Management. AI mainly helps with the writing and analytical work of project steering: rephrasing a backlog into user stories, drafting acceptance criteria, summarizing a meeting, spotting ambiguities in a specification, preparing a retrospective, drafting a status report. It helps little — and can be harmful — for decision and relationship work: arbitrating a priority, estimating instead of the team, sensing conflict. Project Management Fundamentals and the Scrum Framework (Scrum Guide 2020) remind us that the team is self-organizing: AI is neither Scrum Master nor Product Owner.

Tools#

Details in AI and Tools. Assistants are integrated into the editor, repository, ticket tracking, and delivery pipeline: generating commit messages, pull request summaries, ticket drafting, conflict resolution support, script generation. Main safeguard: traceability — a Git In Depth or TFVC In Depth history must make it possible to know what was produced by the assistant and who reviewed it. Project Metrics and Reporting must measure outcomes (escaped defects, change lead time), not volume of generated code.

Maturity matrices#

See self-assessment and the principle of matrices. AI prepares (summarized evidence, computed grid, drafted plan) but does not rate on its own: milestone progress is a team judgment based on reviewed evidence. A model can “find” evidence that does not exist: that is exactly the gap closed by the “reviewed evidence” rule.

Power BI reporting#

Details in the two AI sections of the reporting tracks: OneLake + SaaS Mobile Apps: AI Contribution and SQL Server in SaaS — AI Contribution. Assistants on the Power BI and Fabric side (natural language queries, help writing DAX, report summaries, SQL query generation) and general-purpose assistants are useful to speed up writing; number accuracy remains a matter of acceptance testing and reconciliation (BI matrix — testing). AI features in Power BI and Fabric, and their licensing and capacity prerequisites, change frequently: recheck at the source before any decision (status 2026-10).

Consolidated table: gain, risk, maturity#

Ratings /5: gain = expected value; risk = probability and severity of an undetected error; maturity = current reliability of the practice (5 = mature). Context: team of 5 to 10 people, with systematic human review.

Topic	Use case	Gain /5	Risk /5	Maturity /5	Prerequisites	Main safeguard
Project management	Drafting acceptance criteria and reports	4/5	2/5	4/5	Context provided without confidential data	Review by the deliverable owner
Project management	Detecting ambiguities in a specification	4/5	2/5	3/5	Complete specification	Alerts are questions, not verdicts
Project management	Estimating instead of the team	1/5	4/5	2/5	—	Prohibited: estimation belongs to the team
Project management	Risk analysis (brainstorming)	3/5	2/5	3/5	Existing register	Human owner for each selected risk
Tools	Generating commit messages and merge summaries	3/5	2/5	4/5	Repository accessible to the assistant	Review before merge
Tools	Script generation (PowerShell, T-SQL, pipeline YAML)	4/5	3/5	4/5	Test environment	Run outside production first, review script
Tools	Automatic merge conflict resolution	2/5	4/5	2/5	Strong automated tests	Never without a passing test suite
Tools	Migrating code from one tool to another	3/5	3/5	3/5	Comparison datasets	Before/after sample comparison
Matrices	Evidence summary before workshop	4/5	2/5	4/5	Non-confidential evidence or approved tool	Each piece of evidence opened in its source
Matrices	Milestone rating instead of the team	2/5	5/5	2/5	—	Prohibited: team judgment
Matrices	Drafting the improvement plan	4/5	2/5	4/5	Validated scores	Review, human arbitration
Power BI reporting	Help writing DAX and SQL queries	4/5	3/5	4/5	Described model, test dataset	Reconciliation tests, measure review
Power BI reporting	Automatic report summary or natural language question	3/5	4/5	3/5	Clean, documented semantic model	Never publish a number from an answer without verification
Power BI reporting	Metadata translation (labels, descriptions)	4/5	3/5	3/5	Business glossary	Review by a speaker; control of fixed terms
Power BI reporting	Model documentation (table and column descriptions)	4/5	2/5	4/5	Metadata access	Business review

💡 Expert tip

The best ratio is almost always in producing mechanically verifiable artifacts (tested scripts, reconciliation queries, test skeletons) and supporting documentation (reports, descriptions). The worst is in what cannot be verified (estimates, ratings, decisions).

Governance of AI use#

🟡 In practice. A team adopting AI without a framework ends up with individual use cases, data sent anywhere, and unattributed content. A minimal framework fits on one page, signed by the team lead and reviewed every six months.

Element	Question to settle	Example rule
Approved tools	Which assistants are we allowed to use?	Explicit list; any other tool is prohibited for non-public data
Authorized data	What can we send?	Data classes (public, internal, confidential, personal); only the first two classes may go to an unapproved external service
Review	Who reviews?	Any output intended for a third party or for production is reviewed by a named person
Attribution	How do we mark content produced by AI?	Systematic signature or mention (see below)
Usage register	What do we keep?	List of use cases, tool, version, date, owner
Vendor assessment	What guarantees do we require?	Data location, training use, retention period, contract
Incidents	What do we do if a leak or error is discovered?	Notification and content withdrawal procedure
Training	Who knows what?	Half-day session on limits, hallucinations, privacy

✅ Good practice

Make governance a maturity milestone: a milestone “AI uses are inventoried, bounded, and reviewed” belongs in a governance, security, or quality theme of your matrix (for example PD — Security, Privacy, and Compliance or DL — Security).

Confidentiality#

🔴 Expert. The main risk is not that AI gets it wrong, but that data leaks out. In a sales force datamart context, typical high-risk extracts include: customer names, sales volumes, prices, sales hierarchies, rep identifiers, discount definitions.

Classify before sending. Never paste into an external assistant: personal data (rep names, contact details), confidential business data, secrets (connection strings, tokens, passwords, keys), proprietary code outside contractual agreement.
Minimize and pseudonymize. To ask for help with a query, provide the schema (column names and types) and dummy data, not a production extract. For an anomaly, describe it without the values.
Verify the contract. An approved service should in principle provide commitments on retention and non-use of data for training; this must be confirmed in the contract and vendor documentation, not assumed (to be checked for each tool, status as of 2026-10).
AI-specific attack surface. Assistants that read external content can be manipulated by instructions hidden in that content (prompt injection). An agent with write access (repository, database, ticket) must run with least privileges, in a test environment, with human validation before any irreversible action.
GDPR. Personal data sent to a third party is processing: legal basis, data subject information, processor, transfer outside the Union depending on the case. This is not a detail: validate it with the organization's data protection officer.

⚠️ Pitfall

Symptom: a user pastes an extract of a customer sales table into a public service “to go faster.” Cause: no rules or approved tool, urgency. Remedy: provide an approved tool that can be used immediately, and a simple rule — “schema and dummy data, never production” — rather than a vague prohibition.

Traceability and content signatures#

🟡 Practical. Any published editorial content intended for an organization must make it possible to answer three questions: who wrote it, with which tool, who reviewed it. Without a signature, text drafted by an assistant at someone's request appears to have been written by that person: that's a problem of honesty and accountability.

Signature of published editorial content (documentation, wiki pages, meeting notes, long tickets): one explicit line in the footer, for example ✻ Drafted by Claude <model> — YYYY-MM-DD. This entire Academy follows this convention.
Do not sign code, comments, commit messages, and technical properties: tracking happens elsewhere (commit Co-Authored-By trailer, ticket label, usage log). Signing everywhere turns the mention into noise.
Labels: a tag (for example “AI”) on tickets, pages, or merge requests whose content was produced or transformed by an assistant, to enable auditing.
Registry of significant prompts for high-stakes uses (generating a migration script, drafting a policy): the instruction and model version must be kept with the output; otherwise, you can't reproduce or explain it.
Accountability: the person who publishes the content is responsible for it, whatever the original draft. A signature is not a waiver of responsibility.

💡 Expert tip

Have traceability carried by your existing tools: field or label in ticket tracking (Azure DevOps / TFS), commit trailer, “Origin” section in a page template. A rule that requires extra manual entry disappears at the first emergency.

What AI does not replace#

Judgment and accountability: deciding a priority, accepting a risk, validating acceptance testing, signing off on production deployment.
Knowledge of context: unwritten business rules, a customer's history, constraints of a version delivered as a kit.
Proof: a passing test, a reconciliation that closes, a signed acceptance testing report.
Relationship: negotiating scope, listening to a user, facilitating a retrospective.
Hidden debt: high-volume generated code without design or review creates debt that volume metrics hide; this is a maintainability risk (DL matrix — technical debt).
Data quality: a language model does not fix a source that changes without warning (PD matrix — sources and contracts).

Adoption roadmap#

🔴 Expert. Adopting AI is a maturity effort like any other: progress happens in stages, each with its own evidence. Indicative timeline for a team of 5 to 10 people; durations are expert orders of magnitude.

Stage	Objective	Actions	Evidence of completion	Indicative duration
0 — Frame	Know what is allowed	One-page usage charter, list of approved tools, data classes, AI lead	Signed charter, published list	2 to 4 weeks
1 — Experiment	Find the 3 to 5 highest-ratio use cases	4- to 6-week pilot on verifiable cases (scripts, descriptions, meeting notes), measurement of time saved and identified defects	Pilot report with observed gains and errors	4 to 8 weeks
2 — Integrate	Make usage reliable and traceable	Shared instruction templates, traceability labels, systematic review, training, usage registry	Up-to-date registry, review observed on a sample	2 to 3 months
3 — Measure	Decide with numbers	Indicators: escaped defects, change lead time, review time per merge request, AI-related incidents; quarterly review of use cases	Quarterly dashboard	3 to 6 months
4 — Scale	Scale up without losing control	New use cases under the same framework, least-privilege agents, continuous tool evaluation	Use cases added or removed by traceable decision	continuous

Preferred indicators (outcomes, not volumes): escaped defect rate, delay between commit and production deployment, review time per merge request, number of confidentiality incidents, team satisfaction. Indicators to avoid: number of lines generated, number of requests sent to the assistant (Goodhart's law: see self-assessment).