AI · 2026-08-04 · 10 min read · by Juan Carlos Zuloaga
Business intelligence has always had the same problem. The dashboard exists. Nobody looks at it. When someone does, they do not trust the number.
AI changes the interface to that data, and it genuinely shortens the distance between a question and an answer. What it does not do is repair the definitions underneath.
Here is what is actually changing, and where the limits are.
The traditional BI flow: an analyst builds a dashboard, stakeholders request a change, two weeks pass, the moment has gone.
Natural-language querying collapses that loop. A manager asks a question in plain language and gets a chart back.
The practical effect for a mid-size company is that the analyst stops building views and starts curating the semantic layer — defining what "active customer" or "completed job" means so the machine answers consistently. That is a better use of an analyst.
Most reporting time is not spent producing numbers. It is spent explaining them in a document.
Generating that narrative — the summary, the deltas, the anomalies worth attention — is now reliable enough for internal use. It typically saves a day or two per reporting cycle.
Keep a human on the interpretation. The machine describes what moved; it does not know that the drop in week 32 was a public holiday.
Humans are bad at watching stable metrics. Attention drifts and small drifts go unnoticed until they compound.
Statistical anomaly detection over your core metrics — with alerts routed to a person, not a channel nobody reads — catches the slow leaks: a form that broke, a supplier whose costs crept, a location that stopped converting.
This needs less sophistication than people assume and pays back faster than any predictive project.
Forecasts improve with better methods, but the ceiling is set by how predictable your business is.
For a service business with seasonality and a stable client base, forecasting is genuinely useful for capacity planning. For a business whose revenue depends on a handful of large deals, the forecast is a story about a small sample.
Be honest about which one you are. Applying heavy forecasting machinery to a lumpy revenue line produces confident nonsense.
Conflicting definitions. If sales and finance count revenue differently, an AI layer will produce two confident, contradictory answers faster than before.
Missing data. No model recovers events you never recorded.
No decision rights. Better information changes nothing if the person who sees it cannot act.
The uncomfortable rule: AI amplifies the quality of your data governance in both directions.
1. Agree definitions for your ten core metrics. Write them down. One page.
2. Consolidate to one source of truth per entity.
3. Add anomaly alerts on those ten metrics.
4. Add narrative summaries to the recurring reports.
5. Only then open natural-language querying to non-analysts.
Doing step five first is the common failure. It produces distrust, and distrust in BI is very expensive to reverse.
Track two things: time from question to answer, and the number of decisions in your leadership meeting that reference a number.
Both are easy to observe and hard to fake. If neither moves after a quarter, your problem was never the tooling.
We run this diagnostic as part of our AI and digital transformation work: https://inspiralgrowth.com/services/ai-digital-transformation