ChatGPT is brilliant. So why can't it do your company's work?

#General-purpose AI has never seen your sales

Ask a general-purpose AI service, "What were sales at each of our branches this month?" and the answer comes back in seconds. The prose is smooth, and it even draws a table. Look at the numbers, though, and none of them can go into a report. It has never seen your company's sales, so it either offers generalities or produces plausible example figures.

The problem isn't the model. It can't read your company's data, the material its answers need, and there is no way to apply your company's access and record-keeping rules to it.

#Materials and rules come before intelligence

AI used for work has to meet four conditions before its intelligence matters. Two concern the materials its answers are built from. The other two concern the rules your company has to keep.

Start with the materials. The AI has to know your company's data: internal policies, product documents and business databases must be what its answers are made of. That data also has to be current. It needs to read today's inventory and orders. The figures in last month's report won't do.

The rules are permissions and records. If someone isn't allowed to see HR data, the AI must not show it to them either. And what the AI did has to leave a trace. Accountability depends on being able to check later which data it queried and what answer it gave.

#General-purpose AI can't see company data, and access and records can't be applied to it

General-purpose AI learned from text published on the internet, so it has never seen the databases and documents inside your company. Upload a file and it will read that file. But someone has to find the material and paste it in every time, and the upload is only a snapshot of that moment.

The rules are harder still. When employees each upload material to an AI service on their own, your company has no way to know who saw what. There is no place to enforce permissions and no record to show an auditor.

#It reads your database directly, within the asker's permissions

When OwlAI gets a question, it picks what it needs from the connected databases, documents and internal tools. Ask about sales and it queries your database directly for today's numbers, then finds the criteria in internal documents and answers with both.

Administrators decide, role by role, what can be seen. A role determines which databases (down to individual tables for databases connected through OwlDB), which document collections and which external tools are available, and the AI works only within the role of the person asking. Every SQL statement is recorded when it runs, so administrators can see in the SQL executions tab who queried what in which database.

The person asking can check too. The steps the AI took pile up under Thinking above the answer, and expanding a step shows which data it queried and what came back. Whether a number goes into the report as is becomes a decision you make from that evidence.

#The questions this series takes on, one by one

"OwlAI Insights" covers the questions companies actually run into when adopting AI for work, one per post. We look in turn at why asking a company database in plain language is hard, what it takes for AI to know your company and your people better the more you use it, how much repetitive work and document writing you can hand over, and how to handle the egress, permission and audit questions your security team will ask.

Each post is paired with an "OwlAI Engineering Notes" post that covers the same topic technically. If IT staff are reviewing adoption, pass the paired post along to them.