AI is not a shortcut around BI

The AI sales pitch landing on mid-sized companies’ desks has a tempting subtext. You can skip the data cleanup. You can skip the slow work of building a reporting foundation. Point the model at your existing systems and it will sort the rest out.

This sells because getting the numbers in order genuinely is slow, and anyone who has run a BI project knows it. The shortcut just does not exist. The model runs on whatever data is there, and it does not fill in what is missing.

Nagle, Redman and Sammon examined the state of that data for the Harvard Business Review in 2017. They had 75 executives each check about 100 recently created records from their own area. 47% of the records contained at least one critical error, and only 3% of the assessments reached a level the authors rated acceptable.¹

For the AI phase there are now numbers. In February 2025 Gartner predicted that through 2026 organizations will abandon 60% of AI projects not supported by AI-ready data. In a Gartner survey of data management leaders from July 2024, 63% either lacked the data management practices AI needs or did not know whether they had them.² In November 2025 the IBM Institute for Business Value published a survey of 1,700 chief data officers. 81% prioritize investments that accelerate AI. 26% are confident their data can support new AI-enabled revenue.³

Three patterns we see in these projects again and again.

1. Natural language queries on undefined metrics

The model answers the question “What was revenue last quarter?” in two seconds with €18.4 million. No dashboard is needed.

The problem starts upstream of the model. Revenue means different things in different departments. Sales counts order intake, finance counts invoiced amounts net of returns, controlling applies the accounting standard. The €18.4 million is correct under one of these three definitions. A good tool shows which query is behind the number. Whether it is the right definition can only be checked if the company has laid down somewhere what revenue means.

How often a model gets such questions wrong can be measured. BIRD, the standard academic benchmark for translating questions into database queries, supplies the business context for each question as a hint, for example how a metric is calculated. With the hint, GPT-4 answered 55% of the questions correctly when the benchmark launched in 2023, and 35% without it. Human data experts reached 93% and 72%. The best single model stood at 80% in June 2026, with the hint.⁴ A company only has that hint if its metrics were defined beforehand.

2. AI summaries of dashboards nobody designed

A company has dozens of dashboards, none of which visibly drives a decision, and puts an AI layer on top that turns them into summaries in full sentences. “Revenue in the South region declined 14%, driven mainly by a slowdown in mid-market customer acquisition. Recommendation: review pricing strategy.”

The summary reads well and would pass without comment in an advisory board meeting. But it comes from dashboards that were never built backwards from a specific decision. Nobody settled which decision they were supposed to support, and a layer of text on top changes nothing about that.

The test from the dashboard article applies here unchanged. Can anyone name a decision that was made because the dashboard showed something? If not, the AI is summarizing a dashboard that had nothing to say.

3. Forecasts without a baseline

A company deploys a model to forecast demand, predict churn or recommend prices. The model delivers numbers, and nobody can say whether they are good, because nobody settled beforehand what good would mean.

That is a BI problem. A forecast is only useful if you know how the last one performed, what variance is normal for the business, and from what level of accuracy anyone would act on it. This measurement work often stays undone in mid-sized companies because it does not feel like progress.

Without a baseline, the AI forecast is a guess like the spreadsheet before it, only more elaborate and harder to challenge, because it comes from a model.

What belongs to the foundation

Before a model is introduced, three things should be in place, and in most conversations about AI they get skipped.

The metrics that matter to the business are defined, in writing, with named owners. The dashboards in use are built backwards from specific decisions, and those decisions are documented where a new manager would find them. The historical performance of the business is known well enough that a forecast can be checked against something real.

That is the work the sales pitch wants to spare companies, and in our experience it is also what separates the companies that get measurable value out of AI from the rest. Do it, and a natural language query really does return an answer in two seconds, one where you know which definition sits behind it. Skip it, and it gets done later anyway, by then with a model that has been running on the wrong numbers in the meantime.