i

Description

Vue aérienne d’un grand lac aux rives boisées, traversé par une chaussée.

Source

Mx. Granger · original page

Licence

CC0 Public domain

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

OneLake + SaaS Mobile Apps: AI Contribution

You will be able to position the AI offering in Fabric and Power BI (Copilot, data agents, AI Functions, notebooks) and its maturity status in October 2026.

In brief

• You will be able to position the AI offering in Fabric and Power BI (Copilot, data agents, AI Functions, notebooks) and its maturity status in October 2026.
• You have a use-case table with gain, risk, and maturity out of /5, prerequisites and guardrails, applied to field application data.
• You will be able to write useful prompts and define what AI does not replace: judgment, responsibility, confidentiality.
• Context: overview, architecture, pitfalls.

On this page

1. 1. Overview of the AI offering
2. 2. Rated use cases
3. 3. Copilot and data agents in detail
   1. 3.1 What makes a data agent successful
   2. 3.2 What Copilot Power BI can and cannot do
   3. 3.3 Notebooks and Spark
4. 4. Data Quality with AI
   1. 4.1 Field customer deduplication
   2. 4.2 Anomaly detection
   3. 4.3 Enrichment
5. 5. Sample Prompts
6. 6. Guardrails
7. 7. What AI Does Not Replace
8. 8. See Also

Level: 🟢→🔴

1. Overview of the AI offering#

🟢 Novice.

Vocabulary to define (Glossary):

Availability: according to Microsoft announcements, Copilot and AI capabilities have been open to paid SKUs starting at F2 since April 30, 2025 (status 2026-10). Consumption is charged against the capacity's CU (there is no per-request billing); a dedicated “Copilot” capacity can be created to isolate consumption. Recheck before making a decision: conditions, regions, and admin settings are changing quickly.

2. Rated use cases#

🟡 Practitioner.

Ratings /5: gain (expected value), risk (5 = very risky), maturity (5 = proven). Expert assessments, contextual — see RATING: Reading Ratings, Scores, and the Legend. Context: BI team of 3 to 6 people, field SFA data.

3. Copilot and data agents in detail#

🟡 Practitioner.

3.1 What makes a data agent successful#

An agent translates a question into a query (SQL, DAX, KQL) on the sources you expose to it. Its quality depends on what you provide:

1. Narrow scope: a few gold tables (visit/order facts, customer/product/date dimensions), not the entire lake.
2. Readable schema: explicit names, column descriptions, clear keys and relationships.
3. Instructions: business definitions (“completed visit = status X and duration > Y”), period conventions, currency, language.
4. Validated query examples (few-shot), for frequent questions and ambiguities.
5. Test set: 30 to 50 questions with expected answers, replayed with every model or agent change.
6. Security: the agent must run with the user's rights; the security layer (OneLake security, model RLS) applies with no bypass. Agents with service principal are in preview (status 2026-10): do not use them to bypass security (RLS Security and Access).

3.2 What Copilot Power BI can and cannot do#

• Can: propose a report page, a DAX measure, summarize a visual, answer a simple question on a well-described model.
• Cannot: know your unwritten business rules, guess the difference between “visit date” and “sync date,” fix a poorly designed model.
• Risk: plausible but wrong answer (inappropriate join, missing filter, RLS context).

3.3 Notebooks and Spark#

Copilot in a notebook: PySpark code generation, explanation, error fixing, inline completion (preview, status 2026-10). Useful for standard patterns (MERGE, nested JSON reading, joins). Risks: misused Delta APIs, no idempotency handling, secrets in code. Produced code must go through Git, code review, and testing like any other code (CI/CD for data: foundations, datamarts, pipelines, and reports).

4. Data Quality with AI#

🔴 Expert.

4.1 Field customer deduplication#

Funnel approach, from least costly to most costly:

1. Deterministic rules: normalization (case, accents, punctuation, address abbreviations), exact key (ERP code, SIRET if legally available).
2. Similarity: edit distance, n-grams, phonetics, geographic distance, block grouping.
3. Model (embeddings or LLM) on remaining cases: “do these two records describe the same point of sale?” with score and reasons.
4. Human for mid-range scores; the decision feeds the matching table (and improves the rules).

Why not hand everything over to an LLM? CU cost, variability, non-reproducibility, auditing difficulty. An LLM is a good edge-case judge, not a large-scale deduplication engine.

# Sketch: only ambiguous pairs (mid-range score) are sent to the model
ambigues = paires.filter((F.col("score") >= 0.70) & (F.col("score") < 0.92))
# then human review or LLM with structured output {decision, confidence, reason}

4.2 Anomaly detection#

Before ML, use readable rules: negative visit duration or above 12 h, number of visits per day above a plausible threshold, zero coordinates, price outside ± 5 standard deviations from historical values by product. ML (isolation forest, time series) then complements for non-trivial patterns. In all cases: flag back to field teams, do not correct silently.

4.3 Enrichment#

Categorization of free-text comments (reasons for no-sale, disputes), entity extraction from notes, summary of reports. Precautions: mask or pseudonymize personal data before calling an external model; track the model and prompt version used for each enriched row (reproducibility); measure accuracy on a labeled sample; do not overwrite source data, create derived columns.

5. Sample Prompts#

🟡 Practitioner.

Generating an idempotent Delta MERGE

Model documentation

DAX review

Data agent instruction (excerpt)

6. Guardrails#

🔴 Expert.

7. What AI Does Not Replace#

🔴 Expert.

• Business judgment: deciding that a visit “counts” or that a customer is “active” depends on business rules AI cannot discover on its own.
• Accountability: the published figure is the author's responsibility; “Copilot generated it” is not a defense.
• Acceptance testing: AI proposes, testing and the business validate (Acceptance testing and business validation).
• Modeling: a bad model produces bad answers that sound more convincing.
• Architecture and cost trade-offs: where to copy, which capacity, which storage mode (architecture).
• Change management: adoption by sales reps depends on input simplicity and the feedback they get from it.

For the AI dimension of project steering: AI for Project Management (General Then Data); for AI in tooling: AI in Planning and Version Control Tools; summary: Cross-Functional Summary — AI Contribution. For the relational variant: SQL Server in SaaS — AI Contribution.

8. See Also#

• Series: OneLake + SaaS mobile apps: overview · OneLake + SaaS Mobile Apps: Architecture · OneLake + SaaS Mobile Apps: Difficulties and Pitfalls · Comparison — OneLake / Fabric vs Managed SQL Server for Power BI Reporting · Cross-Cutting Power BI Good Practices · Power BI Reporting Project Checklists
• Matrices: BI Development · Data Governance and Organization · Metadata, Lineage, and Documentation
• Terms: Glossary

✻ Written by Claude Sonnet 5.5