BI Tells You Where. The Wiki Tells You Why.

A live BI anomaly soft-joins to the compiled soft-data world — ranked candidate explanations no metric drill-down can reach.

Picture the war room. Western region sales are down 18% against the prior period, concentrated in enterprise, inside the last ninety days. The filters are clean. The dimensions are clean. Nobody is confused about where the number moved.

And the room is still stuck — because the question that funds the meeting is different. Not “which slice is red?” but “what actually happened in the organisation?”

The dashboard did its job. The organisation’s soft world was never on the join path.

BI’s competence is real — and bounded

Business intelligence earned decades of trust by making hard data joinable. Revenue, region, product, customer, period: once those lived in a common representation, “where did the number move?” became cheap. That is not a small achievement. For many mature metrics, where is honestly solved.

Industry practice has always known dashboards show only part of the picture. Curated metrics can mislead when context is thin; drill-down usually stays inside the structured store that fed the chart.¹ Vendor “root cause” features mostly mean interactive re-slicing of the same analytical model — valuable, and still shaped like a where-question with extra steps.²

So teams drill harder. More hierarchy. More account lists. Sometimes they find a data error. Often they find a cleaner description of the same mystery.

Horoscope answers

The cheap next move is to paste the anomaly into a frontier model: sales down 18%, west, enterprise, ninety days — why? Back comes competitive pressure, pricing, macro, seasonality, churn. Tidy prose. Unearned confidence. The same list you could get for a different company with the same chart shape.

Symptom without organisation is not analysis. It is a horoscope with better grammar.

That is not model stupidity. That is context starvation: the symptom and none of the organisation. There is no CRM migration in the prompt. No unassigned accounts. No regional manager change. No discount-authority memo. The model invents a generic landscape from training priors because you never gave it a company-shaped world. Stuffing more chart tokens does not fix it; attention is finite.³

Three shallow modes:

• BI alone — excellent where; soft causes invisible
• AI alone on the metric — horoscope causes; no org memory
• RAG the shared drive — resemblance to query words, not a join to the live signal

Veterans already do the soft join manually, from hallway graph: who left, which migration hurt, which policy is biting. When they are in the room, the dashboard “works.” When they are not, the same dashboard produces meetings that restate the red number in five dialects.

Joinable surface of the enterprise

Hard data got joinable through ETL and semantic layers. Soft organisational reality — meetings, migrations, ownership breaks, policy changes — left exhaust that was collected and stored but rarely joinable to the metric that would one day need it. Gartner’s language for assets you collect and store but generally fail to use is dark data.⁴ Darkness here is not missing files. It is missing joinability to moments of value.

A corporate wiki expands that surface by compiling significance into claims and edges so meaning is navigable. Relationships baked off-cycle turn retrieval into navigation rather than query-time inference.⁵ Vector search finds similar text; it is weak on how facts connect across hops.⁶ Soft join is stronger than resemblance: a live anomaly collides with a precompiled organisational map.

Data exists ≠ value exists. Soft data is not inactive because it is unstructured. It is inactive because it is poorly connected to the moments when it could create value. Every meaningful edge is a tentacle — an activation surface. Storage without edges is a false comfort.

Traditional ETL makes hard data joinable. The wiki makes soft context joinable. Companion work develops soft-join discipline across corpora;⁷ this essay uses that mechanism as the activation engine against a live BI signal.

Western region, down eighteen

Here is the flagship walkthrough (illustrative scenario, not an industry statistic).

The critical join is not sales.region_id = employee.region_id. It is recognising that “western sales deterioration” may connect to “the CRM migration temporarily left several major western enterprise accounts without clear ownership.” No shared key. Often no shared vocabulary. One is a metric; the other is buried in project notes and staff history.

A useful synthesis ranks candidates. Strongest: account-ownership disruption after Project Horizon, amplified by reduced discount discretion. Competitor X may be real and still rank lower if it cannot explain western concentration. Staffing and pricing stay on the board. Each claim should open into wiki paths and raw evidence. If it cannot open, it is theatre.

Now you are doing BI — over a much larger representation of the organisation.

Ranked candidates, not causality

The marketing sentence “our AI found the root cause” is poison. Soft joins are leads with trails. The commercial product is:

1. BI anomaly (measure, dimensions, window)
2. Soft join against the organisational worldview
3. Ranked candidate explanations
4. Source trails / wiki paths
5. Human investigates

Promote candidates that fit the time window, concentration, multi-page coherence, and openable evidence. Demote eternal macro talk, signals that do not explain concentration, and claims without trails. Measure investigation acceleration — time-to-useful-lead, trail completeness, false-lead rate — not omniscience.

Augment once, cache the edge

The first discovery of Sales Performance ↔ Project Horizon ↔ Account Ownership can be expensive. Drawing three known links later is not. Synthetic augmentation, then cache the expensive result. The AI’s high-value role is noticing “these two things may be related”; the organisation decides whether the edge enters the worldview; future humans and cheap agents navigate it.

Validated edges compound collision probability. Rejected edges become negative knowledge so the same false join does not keep returning with fresh confidence. Known absence is part of a mature surface.

The seller’s wedge

Do not open with “a corporate wiki for AI agents.” Open with the anomaly their BI already shows:

Your BI tells you where the number moved. What if we could join that movement to the soft data of the organisation and help explain why?

Product-line gravity asks only what AI features sit in products already sold. The stronger story notices adjacency: ETL workforces already make messy reality legible; soft context is a softer substrate, not a different planet. Sell activation on a live engagement — ranked candidates, trails, cache loop — before you sell platform ontology.

Package like work they already buy: three recurring anomalies; thin soft compile; shadow soft joins; score against veteran conclusions; cache what earns keep; only then widen.

Run the loop

Same doctrine, other red numbers: margin dips joining to discount exceptions and supplier disputes; churn spikes joining to escalations and lost-deal themes; ops incident volume joining to change tickets and known operational lore. Capture the anomaly cleanly. Soft join. Rank with trails. Human investigates. Persist validated edges. Re-run cheaper next time.

Scope honesty: category and ETL architecture live in BI for Soft Data; natural-key join discipline in The Soft Join; wiki-graph compilation in The Index Is the Data. This piece owns the flagship activation and the portable vocabulary: joinable surface.

Soft data is not inactive because it is unstructured. It is inactive because it is poorly connected to the moments when it could create value. Expand the joinable surface of the enterprise.

BI tells you where. The wiki tells you why.