An analyst who answers in Telegram,
and checks the SQL before it checks the box
A dashboard answers the question someone already thought to build a chart for. We build an analyst that answers the question you actually have right now, in a chat. It queries your real warehouse, with guards that keep it from ever touching data it should not.
The questions nobody built a dashboard for
An AI business data analyst answers real questions about your numbers: which campaign is actually profitable, which product line is slipping, what changed last week. It queries your own data directly and writes the answer in plain language. It is for a team that has data but not the time or staff to query it properly every time a question comes up. It does not replace a dashboard you check daily. It is for the questions nobody built a dashboard for, because nobody knew to ask them until now.
From raw data to a traceable answer
Connectors pull from whatever you actually run on: ad platforms, a CRM, order data, an ERP. They feed a warehouse where the joins and cost logic get worked out once, instead of recalculated by hand every time. A SQL guard sits between the model and the database. Every generated query passes through it before execution, blocking writes, unscoped scans and anything outside a user’s permitted role. Questions arrive as plain chat messages in Telegram, Slack or a web interface. Answers come back in plain language with the number traceable to its source. A log of every question and every query means a number can always be checked, not just trusted.
Getting the joins right before the chat layer
We start by mapping where your real data already lives and what the correct joins and cost logic actually are. A warehouse with wrong joins produces confidently wrong answers faster than no warehouse at all. The SQL guard gets built and tested against intentionally bad queries before a single real question reaches it, not added as an afterthought once something goes wrong. The first version connects one data source and proves out the chat interface and the guard. We layer in more sources once that foundation is solid. We run a tuning period reading the question log weekly, catching cases where the model misread intent before they become a wrong answer repeated across your team.
The mistake that sounds the most confident
Wrong joins are the real danger here. A warehouse that silently double-counts revenue or misses a cost component gives an answer that sounds authoritative and is simply incorrect. That is why we check the SQL guard and the join logic against numbers your team already knows are correct. This happens before the chat layer goes anywhere near a real question. Role-based access also needs real testing, not just configuration. An analyst answering from data a person should not see is a bigger problem than a wrong number. Expect to revisit the guard rules as new data sources get added later.
Timeline and price
| Option | Price | What it covers | Timeline |
|---|---|---|---|
| MVP | from $2,500 | One data source, SQL guard, Telegram or web chat delivery | 4 to 6 weeks |
| Production | from $6,500 | Multi-source warehouse, role-based access, saved questions, query logging | 7 to 9 weeks |
| Full control (handover-ready) | from $7,500 | Everything in Production, plus a full handover package: architecture docs, test suite, admin access audit, and a walkthrough so your own team or another vendor can run it without us | 9 to 10 weeks |
Running cost on top of the build is usually $25 to $90 a month in warehouse hosting and model calls, depending on data volume and query frequency.
What stays yours
You own the warehouse, the connectors, the guard rules and the full source code, running on infrastructure in your name. Documentation covers the schema and every join’s logic. A future hire, or our team on a maintenance basis, can extend coverage without reverse-engineering what is already there.
Related
Pairs with data warehouse and BI product for the dashboard layer most teams still want alongside a chat interface. See the AI monitoring and alerting product for the watch-and-flag side of the same data. See the analytics service page for audits and warehouse builds beyond the AI layer. Real builds: the analytics hub with AI analyst case study. Also the factory ERP recovery case study, where the recovered data became the warehouse this kind of analyst runs on. Tired of being the only person who can answer a data question? Get in touch and tell us what you query most often by hand.
FAQ
How much does an AI data analyst cost?
From $2,500 when it connects to data you already have in one place, a CRM or an ads platform. A full warehouse joining several sources with a tuned SQL guard runs $6,500 to $7,500.
How long does it take?
Four to six weeks for a single data source. Joining several sources into one warehouse with correct cost and revenue logic takes longer. Typically eight to ten weeks, because getting the joins right matters more than the chat layer on top.
What is the stack?
PostgreSQL for the warehouse. Python and FastAPI for the connectors and the guard layer. Claude or GPT turn a question into a query and the result into plain language, delivered through Telegram or a web chat.
Who owns the warehouse and the analyst?
You. The warehouse, the connectors and the code run on your own infrastructure. We document the schema and the guard rules, so your own analyst, if you hire one, can extend it without starting over.
What stops it from running a dangerous query?
A SQL guard checks every generated query before execution: no writes, no unscoped full-table scans, no access outside a user's role. Anything the guard rejects gets logged and shown to the user as a limitation, not run anyway.