AI in BI today: an honest assessment

Every BI product now has an AI button. Some of them save real time. Some of them produce confident paragraphs about numbers that are not in your data. The difference is rarely the model. It is how the product around the model is built: who does the arithmetic, what gets checked, and what you can see when an answer looks wrong.

This is an assessment of where AI in BI stands in 2026, written for someone who runs a small or mid-size business and has no analyst to double-check the output. We use clariBI as the example where it helps, because we know exactly how it is built, including the parts that are still weak.

What AI does well in BI now

Turning a plain-English question into a calculation

Asking "What were our top 5 products by revenue last quarter?" and getting a correct bar chart back is no longer a novelty. Language models are good at reading a question, matching words to columns and deciding which calculation answers it: a total, a trend, a top-N list, a period comparison.

The design question that matters is whether the model also does the maths. A language model asked to add up 3,000 rows will sometimes get it wrong, and it will not tell you. The sound pattern is to let the model plan and describe, and let ordinary code do the counting. In clariBI a planner model picks the calculations, pandas code runs them over the full stored data, and a second model writes the summary from the results. The model never produces a total by itself.

Picking a sensible chart

Once you know the shape of a result, the chart mostly follows: a value over time is a line, a ranked list is a bar chart, two numeric columns are a scatter plot. This part works reliably because it is closer to rules than to judgement. We explain the logic in how algorithms choose chart types.

Drafting a first version

AI is useful for first drafts: a dashboard proposed from the columns in a source, a report outline with a summary and findings, a list of follow-up questions. Drafts are where the time saving is. The draft still needs a person who knows the business.

Where AI still falls short

Explaining why

"Revenue rose 12% because of the spring campaign" is an interpretation, not a finding. Models are fluent at causal-sounding sentences and have no way to know about the supplier delay, the price change or the competitor that closed. Treat any "because" in AI output as a hypothesis to check.

Numbers in prose

The most damaging failure is a sentence that quotes a figure the data does not contain. It reads exactly like a correct sentence. The fix is mechanical: compare every number in the text with the chart data and drop sentences that cannot be backed. clariBI does this for AI dashboards and AI reports, and a metric with no data gets a plain "no data" caption instead of a description. Chat answers do not go through that check yet, so when a figure matters, read it from the chart or table. Our AI limitations page lists exactly where the check applies.

Forecasting with a language model

A language model asked "what will revenue be next quarter?" will give you a number. It should not. Forecasting is a statistics problem: pick a method by testing it on your own history, then show a range, not a single figure. clariBI's forecasting uses no language model in the calculation. It picks a method per series by walk-forward backtest and shows 50%, 80% and 95% prediction bands. The model only writes the summary of a finished forecast.

Ambiguous words

"Orders" can mean a count or a value. "Last quarter" can mean calendar or fiscal. "Customers" can include refunds or not. Most wrong answers come from these choices, not from arithmetic. The help article on when AI analysis falls short has workarounds, and the short version is: name the column, the aggregation and the period.

Messy data

Duplicate rows inflate counts, dates in three formats break trends, and a source that last synced a week ago gives a week-old answer. AI does not fix any of this. It can hide it behind a well-written paragraph.

Fully automatic reporting

An AI report is a good draft for a person to review before it goes to a client or a board. We cover what to check in AI reports: what to trust and what to verify.

Questions to ask any AI BI vendor

Use these in a demo, with your own data rather than the vendor's sample set.

1. Who does the maths? If the language model computes totals, expect occasional silent errors.
2. Can I see how a number was calculated? Look for a drill-down that shows the rows, the date range and the aggregation behind a chart.
3. What happens to a narrated number that is not in the data? "We have good prompts" is not an answer. A check that removes unbacked figures is.
4. What data leaves my account, and to whom? Column names, sample values and results usually go to a model provider. Find out which one and whether it is listed as a sub-processor.
5. How is AI usage billed? Per seat, per question or by usage. Ask to see the cost of a real question before you buy.
6. Ask something you already know. Last month's revenue, from a source you trust. If it does not match, find out why before going further.
7. Ask something vague. "How are we doing?" shows whether the tool asks for detail, guesses sensibly or invents.

How clariBI handles these trade-offs

• Calculations are code. Totals, averages, group-bys, trends, period comparisons and correlations run as pandas code over your stored data. Questions can span several sources, joined by date or a shared column.
• Numbers are checked in AI dashboards and reports before the text is saved.
• Forecasts are statistics, not a language model, and come with prediction bands. Forecasting starts on Starter.
• Access is read-only. clariBI reads your connected sources and never writes back.
• Credits follow usage. Each answer costs at least 1 credit, and the exact amount is shown under the reply. See how AI credits work.
• You can switch it off. The Free and Lite plans make no AI calls at all and still give you dashboards.

None of this makes the AI right every time. It makes the wrong answers easier to spot.

A test you can run this week

Start the 14-day trial (50 AI credits, no card), connect Stripe, and ask "What was total revenue last month?". Compare the answer with the Stripe dashboard. Then ask "How are we doing?" and see what comes back. Those two questions tell you more about any AI BI tool than a feature list.