An agent cites the dataset behind each answer
An LLM can answer almost anything. In a business, only the accurate answers matter. Here's how described datasets turn guessing into citing.
There's a demo everyone has seen by now: point a chatbot at a database, ask it a question, watch it write SQL. It impresses right up until you ask something that matters and quietly get a wrong number.
An agent guesses when it gets bare tables and column types. It has no way to know that revenue is net of returns, that one company's chart of accounts differs from another's, or that intercompany transactions need eliminating. So it guesses, fluently and confidently, and sometimes wrong.
In a business, a plausible wrong answer ends up in a board deck as a number nobody can defend.
Context stops the guessing
Better context is the fix. When an agent asks Gatlas a question, Gatlas hands it datasets with definitions, the context an analyst would give you and notes on their columns. "Gross margin" means one thing, computed one way, the same for every company. The agent reads the business from those write-ups.
And every answer carries its receipt: the agent names the dataset every number came from, so you can trust it or challenge it.
Governed, because it has to be
Trust also means boundaries. An agent sees exactly what its connection allows. Each agent connection is scoped to the companies and datasets you grant, Read-only or Read/write. Every request an agent makes is logged.
That's what makes it safe to put an agent in front of real financial data: a layer underneath it that is documented and scoped.
Ask from Claude over MCP, or in SQL, and have the answers delivered to Slack. The trusted answer is the same on every surface.
This is an early draft; we'll keep sharpening the argument as the product and the field evolve.
Written by
