AI limitations

The AI is useful for getting to a first answer quickly. It is not always right. This page lists where it goes wrong and how to check it.

What is reliable and what is not

Totals, averages, group-bys, trends and period comparisons are calculated by code on your data, not by the language model. The model decides which calculations to run and then describes the results. Most mistakes therefore come from choices and wording, not arithmetic:

• The wrong column or aggregation. With columns such as amount and total_amount, or a question that says "orders" without saying count or value, the model can pick the wrong one.
• The wrong period. "Last quarter" or "recent" can be read differently from what you meant.
• Overreaching explanations. "Revenue rose because of the campaign" is an interpretation, not a proven cause.
• Generic recommendations. The model does not know your market, your team or your constraints.

The number check, and where it applies

In AI dashboards and AI reports, every sentence that quotes a number is compared with the chart data behind it, and sentences with numbers that cannot be derived from that data are removed before the text is saved. Metrics with no data get a factual "no data" caption and no insights. Forward-looking sections, such as goals and next steps, are not checked, because their numbers are targets.

Chat answers in Ask clariBI do not go through this check. The chart in a chat answer comes from the calculation, but the summary text could still misstate a figure. When a number matters, read it from the chart or table.

Data quality

• Missing values change averages, counts and trends.
• Duplicate rows inflate counts and sums.
• Mixed formats (dates in several formats, numbers stored as text) can make a column look like the wrong type.
• Stale data gives stale answers. Check when a source last synced before a critical question.
• Small samples. With a few dozen rows, a "trend" may be noise.

Large data

Calculations run over the full stored data of the sources in scope, not a sample. The model itself sees only column names, a few sample values per column and the calculated results. Large sources with long histories make answers slower and can use more credits; narrowing the question (a period, a filter, one source) helps.

Forecasts are different

Predictions in chat are not forecasts. Real forecasts live on the Forecasting page (Starter and higher) and use statistical methods chosen by backtesting, with prediction bands, and no language model in the calculation. The AI only writes the summary of a finished forecast.

What leaves clariBI

The language model is Google Gemini, listed as a sub-processor in our privacy policy. For each question it receives the column names, a few sample values, your question and the calculated results. When a scanned page needs vision extraction, the image of that page is sent too. If data may not be sent to that provider at all, do not connect it or do not use the AI features on it; Free and Lite plans make no AI calls.

What the AI cannot do

• See data you have not connected. It does not browse the web.
• Change your data. Everything is read-only.
• Replace someone who knows the business.

How to check a result

1. Read the figure from the chart or table, not only from the text.
2. Ask for a number you already know (last month's revenue) and compare.
3. Rephrase with explicit column names and a period if the result looks off.
4. Narrow to one data source to see which one drives the result.
5. Have someone who knows the area review AI reports before they are acted on.

Related

• AI overview
• Writing better questions
• Ask clariBI