Ask questions about your data in plain language and see where each number comes from
A simple question like “which region sold most in May?” often means a ticket to the analyst or an afternoon in five dashboards. We build AI tools that turn plain-language questions into queries on your own database, warehouse or APIs. The answer comes with numbers, a chart and the query behind them. Recurring reports are written and sent on schedule. The model decides what to look up; the numbers come from your data, not from its memory.
The situations where teams usually call us.
Managers wait days for an analyst to answer questions that take five minutes to look up.
Your data sits in the CRM, the ERP, a warehouse and a few spreadsheets, and nobody outside IT can put it together.
You have dashboards, but every meeting ends with a question none of them answers.
The monthly report takes someone two days of copying numbers, and no two versions match.
Your customers want to explore the data in your product, not just look at fixed charts.
The model never guesses a number. It writes a query, your database answers it, and the model explains the result.
From the projects we’ve built in this area.
Most wrong answers aren’t AI mistakes. They come from three definitions of “revenue” or “active customer” living in different reports. A short glossary that the model reads with every question fixes more errors than a better model does.
1 source
Connecting everything at once makes it hard to tell why an answer is wrong. One well-described source, plus a list of questions your team really asks with answers they trust, gives you a test you can rerun after every change.
Duplicate customers and missing dates stay duplicated and missing. The tool only makes them easier to spot. When two systems disagree, the first weeks often go into deciding which one is right, and every later report benefits.
Numbers you check every day belong on a fixed dashboard: fast, cheap and identical each time. The chat is for the questions nobody built a chart for, and those questions show you which new charts are worth building.
All three answer questions about data. They differ in which questions they can answer and how far you can trust the result.
| BI dashboards (Power BI, Looker) | AI chat with uploaded files | AI analyst on your systems | |
|---|---|---|---|
| New questions without an analyst | No | Yes | Yes |
| Works on live data | Yes | No | Yes |
| Shows how each number was calculated | Yes | Partly | Yes |
| Respects who may see what | Yes | No | Yes |
| Same answer to the same question | Yes | Partly | Partly |
| Time to first use | Weeks | Minutes | Weeks |
A typical timeline. We confirm yours after the project workshop.
Week 1
Collect the real questionsIn the project workshop we list the questions your team asks, find where the answers live and agree what each metric means.
Weeks 2–4
Connect and describe the dataRead-only access to the first sources, table descriptions, the glossary and permissions. You try it in a demo every week.
Weeks 5–6
Test against known answersWe run the question list until the numbers match what your analysts expect, then open the tool to a small group.
Then
Widen and automateMore sources, scheduled reports and alerts, chosen by the questions people actually ask.
Systems we’ve built that are close to this one.
An AI assistant that answers questions about building permits and contractors using current database records.
Read the case study
Datair uses AI to help people analyze data and create reports without technical knowledge.
Read the case studyWe design it so it can’t do that quietly. The model writes a query, your database returns the figures, and every answer shows the query and its source. If a question can’t be matched to your data, the tool says so instead of guessing.
No. We can start on a read-only copy of your main database, a few APIs or well-kept spreadsheets. A warehouse helps once you combine many sources or large volumes, and we’ll tell you if you’ve reached that point.
Databases such as PostgreSQL, MySQL or SQL Server, warehouses such as BigQuery or Snowflake, and CRM, ERP or accounting tools through their APIs. Where a system has no API, a scheduled export usually does the job.
The tool has read-only access and follows the permissions you set: a sales rep sees their own accounts, a regional manager sees the region. Every question and query is logged, so you can check who asked what.
Only what the model needs to write the query and explain the result, usually table descriptions and the rows in the answer. We use providers with EU hosting that don’t train on your data. When data must stay in-house, we run open-weight models such as Llama, Mistral or Qwen on your own servers.
Yes. Charlie, the assistant we built for Shovels, answers their customers’ questions about building permits from live database records. Datair, our own SaaS project, builds dashboards from plain-language requests. Customer-facing tools need stricter limits, so each customer sees only their own data and the cost per question fits your pricing.
A narrow tool on one source can fit our €5–10k range for a single process. Most analyst tools with several sources, permissions and reports fall in the €10–30k range. Running costs are model usage per question plus hosting, and we estimate them from your expected volume before the build.
Our first versions take 6 weeks to 3 months. A tool on one or two sources with clear questions sits at the shorter end. Many sources with conflicting definitions take longer.
Send us ten questions your team asks every month and tell us where the data lives. We’ll tell you which ones a tool can answer reliably and what a first version would cost.
JurajCEO, QikBuild
You leave the call with
Or write to hello@qikbuild.com
