Business Intelligence

Pilot wave 2 (M7). Impact: high. See the hub. BI may today be a role inside finance or IT rather than a department (assumption — to validate in discovery); the program treats it as a distinct function because most other departments’ AI use cases stand on its output.

Current state (assumptions to validate)

Assumptions — to validate in discovery: no dedicated BI team — reporting is SoftOne standard reports plus person-specific Excel workbooks exported and reworked per question; each department maintains its own version of the truth (sales numbers from sales’ Excel disagree with finance’s); no data warehouse, no documented definitions (“revenue” may include or exclude chart-agency commissions depending on who computes it); master data (items, customers, suppliers) has duplicates and gaps accumulated over years across tens of thousands of SKUs. Pain points: analyst-as-bottleneck for every ad hoc question, definitional disputes in management meetings, data quality blocking any automation built on top.

Target AI-first operating model

By M12 BI operates the company’s semantic layer: a governed set of definitions (revenue, margin, segment, customer, on-time delivery) mapped onto SoftOne tables, exposed three ways — natural-language querying through the assistant (“show navigation-segment margin by month this year” → SQL against the layer → cited answer), self-serve dashboards per department, and scheduled report feeds (the executive briefing, department KPI packs). BI is also the program’s data steward for RAG: it owns which corpora feed which assistants, their freshness, and their quality — including the IMPA↔SKU mapping for sales and the master-data cleanup that everything depends on. Humans own definitions, data-quality decisions, and certification of any number used externally.

AI use cases

Use case	Pain addressed	Data needed	Complexity	Impact	Pilot
Natural-language querying over SoftOne via semantic layer (text-to-SQL with cited results)	Analyst bottleneck, Excel sprawl	Semantic layer, SoftOne replica/extracts	H	H	Y
Semantic-layer definition drafting (metric docs from stakeholder interviews)	Undocumented, disputed definitions	Interviews, existing report logic, SoftOne schema	M	H	Y
Master-data cleanup assistance (duplicate detection, attribute completion across ~10k+ SKUs, customers, suppliers)	Data quality blocks all automation	SoftOne item/customer/supplier masters	M	H	N
Automated dashboard/report commentary (per-department packs)	Numbers without narrative	Semantic layer, KPI targets	M	M	N
Data-quality monitoring narrative (weekly: new anomalies, drift)	Silent decay	Profiling outputs on key tables	M	M	N
RAG corpus health reports (freshness, coverage, retrieval quality per assistant)	Stale corpora degrade every assistant	Corpus metadata, retrieval logs	M	H	N

Process transformation opportunities

• Question answering: from “email the analyst, wait days” to self-serve NL query with certified definitions — the deepest cultural change in the program.
• One truth: retire departmental Excel shadow-reporting; the semantic layer is the single computation of every shared metric.
• Master-data governance: standing ownership and cleanup process for item/customer/supplier masters (with sales, procurement, warehouse), replacing accretion.
• RAG stewardship: formal corpus lifecycle (owner, source, refresh, retirement) for every assistant in the agent registry.

Required data sources

• SoftOne database access — read replica or scheduled extracts (approach per architecture overview and integrations; direct-DB vs API to decide with the SoftOne vendor).
• Existing department Excel reports (to mine definitions and retire).
• E-commerce platform analytics; master-data exports.
• All departments’ corpus sources (stewardship role).

Potential AI agents

• NL Query Agent — translates business questions to semantic-layer queries and returns cited results; results marked “certified” only for approved metrics; BI reviews new query patterns weekly.
• Data Quality Agent — flags anomalies and drafts cleanup proposals; data owner approves every master-data change.
• Report Commentary Agent — drafts narrative on scheduled reports; department owner approves before distribution.

Automation opportunities

• Scheduled extract/refresh pipelines (n8n) from SoftOne to the reporting store.
• Automated KPI pack generation per department per KPI framework.
• Data-profiling jobs with threshold alerts.

Required integrations

• SoftOne data access (the program’s most technically demanding integration — schema documentation from vendor required, see integrations); dashboard tooling; the LLM platform’s RAG services.

KPIs

KPI	Baseline	M12 target
Ad hoc questions answered self-serve (vs analyst)	~0	≥60%
Certified metrics with documented definitions	0	Top 30 metrics
NL query answer accuracy on certified metrics (sampled)	n/a	≥95%
Departments on semantic-layer dashboards	0	All 15
Duplicate rate in item/customer masters	TBD (M3)	−80%

Risks

• Text-to-SQL returns a confidently wrong number that drives a decision → certified-metric whitelist, query-plan display, sampled accuracy audits; uncertified results visibly labeled.
• Semantic-layer politics (whose definition wins) → definitions ratified at the executive level (M3–M4), not negotiated per report.
• SoftOne data access blocked or throttled by vendor constraints → early vendor engagement (M1); fallback to scheduled extracts.
• BI capacity — the steward role is a real job → staffing decision needed by M4 (assumption: role must be named, hire or designate).

Training needs

• BI/analyst staff: semantic-layer tooling, text-to-SQL supervision, corpus stewardship (M4–M6).
• All departments: NL querying basics, reading certified vs uncertified results (M8, with change plan).
• Executives: querying in the weekly rhythm (with executive management, M7–M8).

Deliverables

• Semantic layer v1 (top 30 metrics) over SoftOne; NL query pilot (M7) and production service (M9).
• Department dashboard set; master-data cleanup program with measurable duplicate reduction.
• RAG corpus registry and stewardship routine.

12-month execution milestones

Month	Milestone
M1	SoftOne data-access approach agreed with vendor
M2–M3	Definition interviews; schema mapping; master-data quality audit
M4	Executive ratification of top-30 metric definitions; steward role staffed
M5–M6	Semantic layer v1 built; extract pipelines live; first dashboards (serve wave-1 KPI needs)
M7	Wave-2 pilot: NL querying for a pilot user group (exec + finance + sales mgmt)
M8	Accuracy audit; certified-metric whitelist v1
M9	NL Query Agent in production company-wide; corpus registry operational
M11	Master-data cleanup wave 1 complete (items A-class)
M12	KPI review; year-2 data roadmap