A bar chart made of gray wooden blocks on a table, the tallest block blue

AI for Data Analysis: Ask Your Company Data in Plain English

Most companies have plenty of data and very few answers. Sales live in the ERP, leads in the CRM, customer requests in email and messengers, ad spend in the ad accounts. To find out why revenue dipped in one region or which products get returned most often, a manager asks an analyst and waits for an export. Sometimes a day, sometimes a week.

AI for data analysis shortens that path: a person asks a question in plain language, and the system finds the right tables, runs the numbers and shows the answer together with how it was produced. Below we look at what AI analytics for business can actually do, how it is built, where it gets things wrong and how to run a pilot without unnecessary risk.

What AI for data analysis can do

"AI analytics" covers several different tasks. It helps to tell them apart, because each has its own way of being checked and its own cost of error.

Natural language questions about your data

The most visible scenario. A manager types: "How many orders did we get from Lyon in October, and how does the average order value compare with September?" The model understands the question, finds the right tables, writes an SQL query, runs it and answers with numbers, a table or a chart.

The key point is that the database does the math, not the model. A language model is good at turning a question into a query but bad at adding up thousands of rows in its head. So the answer is built on the actual query result, not on the model's guess.

Automated reports and summaries

Recurring reports, such as a weekly summary of sales, inventory or support performance, can be assembled automatically. The numbers are calculated by fixed rules, and the model writes a short commentary on top: what went up, what went down, what deserves attention. The summary lands in email or Slack before the workday starts.

Here AI does not replace the report; it saves time reading it. A manager sees the three main changes and opens the full table if needed.

Anomaly detection

A sudden spike in returns for one product, orders with unusual discounts, a conversion drop in one region, duplicate payments. The anomalies themselves are found by standard statistical rules and machine learning models, while the language model helps explain the finding in plain words and pull together context: which orders, which sales rep, since which day.

Making sense of text data

Reviews, support tickets, sales reps' notes in the CRM, survey responses: this is data too, just unstructured. The model sorts it by topic, detects sentiment and highlights recurring complaints. Text turns into a table you can count: how many delivery-related tickets this month, which reasons for dissatisfaction are growing.

Where the data comes from: connecting sources

Data analysis with AI starts where any analytics starts: with access to data. Typical sources in a mid-sized company:

Connecting the model directly to the production ERP or CRM database is usually a bad idea. It is better to export data on a schedule into a separate analytics warehouse: queries do not load the operational system, and data from different sources can be joined. ERP data is the most common and the most labor-intensive case, because every company's setup is different, with its own customizations and reference tables.

Our developers handle ERP integrations end to end, including changes on the ERP side when the data you need is not in the standard objects.

The data description layer: without it, AI guesses

This is the most underrated part of the project. The database has a table called doc_sales with the fields sum, sum_vat, status and type. An outsider would not know which amount to use or which statuses count as a sale. Neither does the model, but unlike a person it will not ask. It will pick something plausible.

A model writes queries exactly as precisely as your data and metrics are described.

That is why a description layer sits between the model and the database, often called a semantic layer. It records:

For frequently used metrics, it is better not to let the model write the query from scratch but to provide tested formulas. Then a question like "what was revenue for the quarter" uses the ready definition, and the number matches what finance sees. One metric, one definition across the company: that pays off even without any AI.

Checking answers and access rights

AI analytics is only useful if you can trust the answer. Trust is built on transparency, not promises: the user must see how the number was produced.

Access rights follow the same rule as any enterprise system: everyone sees only their own data. A branch manager sees their branch, a sales rep sees their clients, and salaries are visible only to those who should see them. The restriction is set at the data level: the user's query automatically gets the right filters, and restricted tables are simply unavailable. You cannot rely on the model "choosing not to answer" an awkward question.

A separate question is where the data goes. If your analytics includes customers' personal data or trade secrets, that affects the choice of model: a cloud service or one deployed on the company's own servers.

Where AI makes mistakes and how to control them

Model errors in analytics are more dangerous than in correspondence: a wrong number looks just as convincing as a correct one. It helps to know the typical failures in advance.

ErrorWhat it looks likeWhat helps
Wrong metricRevenue with VAT instead of without, orders counted together with cancellationsReady metric definitions in the description layer
Wrong table joinTotals double when orders are joined with line itemsDocumented relationships, test questions
Wrong period"Last month" read as the last 30 daysShowing the period explicitly, a clarifying question
Invented conclusionThe model explains a drop without any data on causesSeparating numbers from interpretation, labeling hypotheses
Dirty dataDuplicate customers, empty fields, city names spelled different waysCleanup during export, reference tables

The last row deserves special attention. If one customer is entered in the CRM three times under different names, neither an analyst nor a model will calculate their revenue correctly. AI analytics projects often begin by surfacing problems in the data itself, and that is a useful side effect.

For decisions involving money (budgets, prices, bonuses), the AI's answer remains a starting point. Any number a decision rests on should be checked against a report or verified through the query shown.

How to start a pilot

Building AI analytics for the whole company at once is expensive and slow. A pilot on a narrow task will show within a few weeks whether the approach works on your data.

  1. Pick a department and questions. For example, the sales department and twenty to thirty questions managers regularly ask analysts. The more specific, the better.
  2. Pick one or two sources. Often the ERP and the CRM. Everything else is connected once the foundation works.
  3. Describe the data and metrics. Together with the people who build reports today, pin down definitions and business rules. This is the most important step.
  4. Build a test set. Questions with known correct answers, including tricky and ambiguous ones.
  5. Build a prototype. Export to a warehouse, a description layer, a web or messenger interface, the query shown under every answer, access rights.
  6. Hand it to a small group and review errors. A few weeks of real use, analysis of wrong answers, updates to the descriptions.
  7. Decide what comes next. Expand to other departments and sources, add automated summaries, or stop at what already works.

Technically, a pilot like this is a small web service with integrations, and we build it the same way as other projects: first we dig into processes and data, then we design and develop. As a digital agency, we build web services and implement AI for companies; you can see examples among projects by Sphere, a web development agency.

Who it is for, and who should wait

AI for data analysis delivers the most where there is a lot of data, questions come up often and there are not enough analysts. Signs it is time:

If records are kept only partially, every department calculates metrics its own way and data lives in scattered spreadsheets, start not with a model but with putting your records in order and building a single warehouse. AI can then sit on top of a solid foundation and answer more accurately.

FAQ

Can I just upload an export to ChatGPT?

For a one-off task with a small anonymized table, yes. For regular work, no: a public service does not know your metric definitions, does not respect access rights, and uploading customer data and trade secrets there is often prohibited by security policy. Production AI analytics connects to your sources and calculates by your rules.

Does the model calculate the numbers itself?

It should not. The model turns the question into a database query, and the database does the calculation. The model then formats the result and writes an explanation. That way the numbers are accurate and their origin can be checked through the query shown.

Can AI analyze data from our ERP or accounting system?

Yes, it is one of the most common cases. ERP data is exported on a schedule to a separate analytics warehouse, where it can be joined with the CRM and other sources. If the data you need is not in the system's standard objects, some customization on the ERP side is required.

What if the AI gives a wrong answer?

Flag the answer as wrong. The data owner reviews the query and finds the cause, most often an undocumented metric or a non-obvious business rule, and adds it to the description layer. The question goes into the test set so the error does not come back.

Will AI replace analysts?

No, but it will change their work. Routine questions and recurring summaries move to the system, while the analyst describes data and metrics, checks answer quality and focuses on tasks that need deep research rather than counting.

Let’s discuss your project
Related reading