AI agent for data analysis
Data analysis through questionsin plain language.
The AI agent for data analysis answers the everyday questions that do not have a ready-made report. It turns the question into an SQL query, runs it only on the data you have access to, and shows you both the result and the logic behind it.
The data exists, but every new question becomes a request to someone else
Periodic reports answer the questions known in advance. But most decisions start from a new question: “Why did sales drop in the western region last month?” By the time it reaches someone who can write the query, the moment has passed.
- Ad hoc questions end up with the same few technical people, who become a bottleneck
- Dashboards show what happened, but rarely allow a follow-up question
- Excel exports create parallel versions of the same figure
- Without an explanation of the calculation, people do not trust the figure they receive
Who it is for
- Managers with specific questions about sales, customers or costs who do not want to wait for an analyst
- Business intelligence teams overloaded with repetitive ad hoc requests
- Companies with good data in databases or data warehouses but few people who know SQL
- Sales and operations departments that want to check a hypothesis before a meeting
What an answer looks like
One question, a whole dashboard.
The agent doesn't answer with a single number: it writes the query, runs it read-only and lays the result out in a dashboard you can explore, with the sources and the reasoning in view.
Illustrative example · demo data, not from a client
The question asked: How did sales by channel change over the last 12 months, and what changed in October?
Sales by channel
- Online
- Distributors
- Stores
↑ 20.1%
RON · on last November
Channel mix
October45%
online
- Online45%418k
- Distributors31%286k
- Stores24%221k
New orders
vs. September+12.4%
Growth came mostly from the online channel
The query it ran
read-onlySELECT month, channel,
SUM(amount) AS sales
FROM orders
WHERE date >= '2025-11-01'
GROUP BY month, channel;The agent's explanation
October's growth comes from online (+21%), after the campaign of 8 October. Stores dip slightly (−3%), in line with last season.
Sources used
3 tables- ordersERP
- customersCRM
- returnsERP
Orders and returns
last 6 months, three years overlaid
- Orders
- Returns
Growth by product
7 days- Flexible tile adhesive C2OnlineRON 48,200+24%
- Aluminium profile 3 mDistributorsRON 31,900+9%
- All-purpose primer 5 LStoresRON 17,400−6%
Orders per day
1,284
this week
The agent only runs read queries, on the data you have access to. The figures above are made up for the example.
What the AI agent for data analysis does
- 01
Understands the business question
Clarifies the period, the segment or the definition of a metric when the question is ambiguous, before calculating anything.
- 02
Generates the query
Writes the SQL query based on the documented schema and the definitions agreed by the team.
- 03
Runs in read-only mode
Executes the query on authorised sources, with a read-only account and volume limits, without changing the data.
- 04
Explains the result
Summarises in words what the figures show, mentions the assumptions made and displays the query used.
- 05
Suggests follow-up questions
Suggests useful breakdowns (by region, product, channel) so the analysis can go further.
- 06
Generates simple charts
Displays the result as a table or chart when it helps understanding, and allows export for verification.
How it works
How it gets from question to answer
The steps marked in yellow are approved by a person.
- Input
- AI
- Action
- Human in the loop
- Result
- 01
Input
A question in plain language
The user writes, for example: “Who are the top 10 customers by margin in Q3?”
- 02
AI
Clarifies and plans
The agent chooses the relevant tables and definitions and asks for details if needed.
- 03
Action
Runs the query
Executes the generated SQL with a read-only account, on the permitted sources.
- 04
AI
Explains the result
Summarises the figures, notes the assumptions and attaches the query.
- 05
Human in the loop
The analyst checks
For important decisions, an analyst confirms the logic before it is used.
- 06
Result
A transparent answer
The figure, the explanation and the query, ready to be checked or saved.
What data and tools the analysis agent uses
- The schema of the database or data warehouse, documented with table and column descriptions
- An agreed glossary of metrics (what “margin”, “active customer” and “completed order” mean)
- A database account with read-only rights and time and volume limits
- Role-based access rules: what data each user can see
- Examples of validated questions and queries, used as a reference
Possible integrations
- PostgreSQL
- Microsoft SQL Server
- MySQL / MariaDB
- Google BigQuery
- Snowflake
- Power BI
- Looker Studio
- Microsoft Teams or Slack as the question interface
- PostgreSQL
- Microsoft SQL Server
- MySQL / MariaDB
- Google BigQuery
- Snowflake
- Power BI
- Looker Studio
- Microsoft Teams or Slack as the question interface
Each integration's compatibility is checked during the analysis phase, for the systems and versions you use.
Control and transparency
What a person validates
- Figures used in financial decisions, budgets or external communications
- The metric definitions in the glossary and any change to them
- Access rights to sensitive data (salaries, customers' personal data)
- New queries that become recurring reports or go into an official dashboard
Limits and how we manage them
- It works best on documented data: a clear schema and a glossary of metrics reduce incorrect queries
- A query can be syntactically correct and still answer a different question; that is why we always display it
- The agent does not establish causes; it shows correlations and breakdowns, and interpretation remains with people
- It does not replace official periodic reports or the existing BI system
- Very complex questions or large volumes may require an analyst or a dedicated data model
How we measure the accuracy of the answers
We agree the indicators together before launch and track them afterwards. The values depend on the process, the data and the volume.
- Query accuracy
- We run a set of test questions with answers calculated by analysts and compare the results
- Rate of correct clarifications
- We check that the agent asks for details precisely on ambiguous questions and does not guess
- Respect for permissions
- We test with users in different roles that each receives only the data they are allowed to see
- User trust
- We collect “useful / wrong” marks on each answer and analyse the cases marked wrong
How we implement the data analysis agent
- 01
Inventory of sources and questions
We choose the data sources and collect the real questions that analysts receive.
- 02
Documenting the data
We complete the table descriptions and the glossary of metrics together with the people who know the data.
- 03
Test set and calibration
We build a set of questions with verified answers and adjust the agent until the results are consistent.
- 04
Launch for a small group
We start with a few users and an analyst who reviews the answers, then widen access.
Use cases
The examples below are conceptual scenarios, not delivered implementations.
- Multi-store retail
A drop in sales in one region
The sales director asks why a region has declined, and the agent breaks it down by store and category and shows where the difference is concentrated.
Conceptual example - B2B distribution
Customers who have reduced their orders
A sales manager asks for the list of customers who ordered less than in the previous quarter, with the value of the difference and the query used.
Conceptual example - Logistics
Delays by route
The operations team asks which routes had the most late deliveries and receives a week-by-week table, ready to discuss in a meeting.
Conceptual example
Frequently asked questions
How is it different from a periodic report or a dashboard?
Reports answer fixed questions at fixed intervals. The agent is for new, specific questions and for the follow-up questions that come up when you look at a report.
Can the agent change or delete data?
No. We connect it with an account that has read-only rights, with time and volume limits per query. Every action is logged.
What happens if the agent is not sure what the question means?
It asks for clarification (the period, the metric definition, the segment) instead of assuming. If it still cannot answer reliably, it says so and suggests involving an analyst.
Does our data reach the AI model provider?
It depends on the architecture chosen. We can send the model only the schema and aggregated results, use providers with processing in the EU, or use separately hosted models. We choose the option based on how sensitive the data is.
Do we need a data warehouse to get started?
Not necessarily. We can start from an operational database, with care for performance, or from a dedicated copy for analysis. The recommendation comes after we analyse the sources.
Do we need to know SQL to use the agent?
No. The question is asked in ordinary words, and the agent writes the query. It always displays it, so an analyst can check it whenever the figure matters for a decision.
Can’t find the answer? Write to us and we’ll reply directly.
Want answers from your data without waiting for a report?
Send us a few questions you ask often and tell us where the data is. We will show you how an agent would answer and what would need to be prepared first.
Related solutions
- AI agent for financial reportingPrepares periodic reports from your financial and operational data and flags variances, which the finance team then analyses.
- LLM integrationLanguage-model features added to the applications and processes you already use, with the model chosen for cost and confidentiality.
- AI evaluation and monitoringTest sets, checks at every change, human spot-check reviews and cost tracking, so you know how your AI performs over time.