Grounding (AI)

Grounding is the practice of tying an AI system's outputs to a specific, trusted source of truth — your own documents, data, and rules — so its answers reflect your business rather than generic internet knowledge. A grounded AI works from your sources; an ungrounded one guesses. For why ungrounded AI makes things up, see Why AI makes things up.

Key characteristics

• Answers are derived from your documents and data, not general training
• Reduces hallucination — the AI works from real sources
• Lets outputs reflect your products, policies, and procedures
• Often paired with sources or citations so claims are checkable
• The difference between a plausible answer and a correct one

Example

A loan officer grounds an AI teammate in their file-requirements checklist. When a borrower's package arrives, the teammate checks it against those exact requirements — flagging the missing bank statement — instead of guessing what a generic loan needs.

How it relates to Kuvai

Grounding is the core of how a Kuvai teammate works. Company Context reads your website and drafts a six-section profile of your business (the basics, products and services, customers and market, competitors, policies and procedures, org and people); every new teammate is grounded in it by default. Files you attach — PDFs, Word, Excel, CSV, slides, images and more — are read (with OCR for images), for your account only. Projects keep the files, chats and instructions for one piece of work together, and you give the teammate the files it needs. Every reply shows its steps, including the web sources it used, so you can check what the answer came from.

Frequently asked questions

What does it mean to ground an AI?

Grounding means tying the AI to your own trusted sources — documents, data, rules — so it answers from your business rather than generic knowledge. A Kuvai teammate is grounded in your Company Context and the files you give it.

Does grounding reduce AI hallucination?

Yes — because a grounded AI works from real, specified sources rather than inventing from general training, it's far less likely to fabricate. It doesn't eliminate errors, which is why Kuvai teammates show their steps so you can verify the output.