Niklas Andersson12 Nov 2026 · 3 min readWork since 2023
Who can build AI agents that answer questions about your data in Slack or email?
A ready-made tool, your own team or a specialist. Salesforce went from dozens of hours a week on custom queries to answers in minutes. How to choose, what usually goes wrong and how to test it.
Who can build an AI agent that answers questions about your data?
An AI or data developer can build it, a specialist firm can build it for you, and ready-made tools exist. Slack has its own interfaces for AI agents, so the technology isn't the obstacle. The choice depends on how clean your data is and whether you have data engineers of your own. We are one of the specialists ourselves.
| Route | Examples | Fits when | Keep in mind |
|---|---|---|---|
| Ready-made tool | Basedash and similar | The data is clean and you want to test quickly | A running fee, and your data passes through one more supplier |
| Your own team | Salesforce, Bitrise | You have data engineers who can look after it | The agent becomes a system someone has to own |
| Specialist | Promtech and others | You have no data engineers of your own and one bounded question | Require that the code and the access are yours |
Why in Slack or email instead of dashboards?
Because a dashboard only helps the people who log in, and most people who decide on the numbers don't. Bitrise's data team chose Slack over email and a new web app for the same reason: people should reach the analyst where they already work. Salesforce writes that dozens of hours a week on custom queries became answers in minutes.
| Channel | Good for | Less good for |
|---|---|---|
| Slack or Teams | Follow-up questions in a thread, answers the whole team sees | Answers not everyone in the channel may see |
| The summary before Monday's meeting, managers without Slack | Quick follow-up questions | |
| Telegram or WhatsApp | Answers on the phone | The company's control over accounts and logs |
What usually goes wrong with AI agents on company data?
The agent answers wrongly with confidence. The team behind Toqan at Prosus found that language models don't write SQL accurately enough on real, complex data, and so built an agent layer around the model. It's rarely the model that decides, but what surrounds it:
- Definitions. The agent has to know what revenue means at your company, not guess.
- Permissions. It should only be able to read, and only what the person asking may see.
- Follow-up questions. Better a question back than a silent, wrong assumption.
- Date and source. The answer should show what data it rests on and how fresh it is.
- Log. Every question is saved, so errors can be found.
The agent is never better than the data beneath it. If the flows aren't in place, that is where to start: how to choose someone to build them.
How do you test whether an AI agent answers correctly?
Write twenty questions you already know the answers to, and ask them before anyone else gets access. Include questions that need two systems joined, a question the agent shouldn't be able to answer, and a question from someone who may not see the data. Repeat the test every time the data or the model changes.
Don't buy on a demo. A demo shows questions the supplier already knows will work.
What has Promtech built?
Agents that turn plain-language questions into SQL since 2023, and an assistant in daily operation since February 2026.
| Years | System | What it does |
|---|---|---|
| 2023–2025 | AudienceScan | A schema-aware agent lets the client's team ask the warehouse without writing SQL |
| Since February 2026 | Analytics in Telegram | Plain-language answers, follow-up questions, CSV and a daily summary. Each customer's questions only touch their own data. |
We haven't yet delivered an agent in Slack, or one that answers by email. The channel is the smallest part of the build: the query layer, the permissions and the tests are the same, and only the door changes. But it's better you hear that from us than find out later.
Frequently asked questions
- Can an AI agent answer questions about our data in Slack?
- Yes. Salesforce and Bitrise both describe running such agents internally. The chat is the simple part. Correct answers depend on your definitions, permissions and tests.
- Do we need a data warehouse first?
- Usually, if the question needs more than one system. An agent answers from data it can read, and data spread across systems has to be joined and tested first.
- Is it safe to let an AI agent read our data?
- It depends on how it is built. Limit access to a short list of approved people, let the agent read but never write, keep it inside a sandbox with only permitted data, and log every question.
- What does it cost to have one built?
- With us, SEK 2,200 an hour excluding VAT, and a first delivery usually takes about four weeks. The AI model and the cloud are billed on top, by use.