How to Train AI on Your Business (No Fine-Tuning)

You don't need to fine-tune a model to train AI on your business. You need to give it four things it can read every time it works: a company knowledge document, your key documents, a voice guide, and a short list of rules. That's how most businesses get an AI that answers in their voice with their facts, and it takes an afternoon, not a data science team.

This guide explains why context beats fine-tuning for a small business, what goes in each of the four pieces, and gives you a copy-ready template for a "company brain."

Why not fine-tuning?

Fine-tuning means further training a model on your own examples so its behavior changes. It's a real technique with real uses, but it's usually the wrong tool for "know my business."

The AI companies say so themselves. OpenAI's guide to optimizing LLM accuracy recommends starting with prompt engineering, then adding retrieval (feeding the model your information) when the problem is missing knowledge, and fine-tuning when the problem is inconsistent behavior, such as format or tone. Anthropic's prompt engineering overview makes a similar case: prompting is much faster to iterate on, needs far less data, and keeps the model's general abilities intact.

For a small business, your knowledge changes constantly: new services, seasonal hours, a price increase. You want something you can edit in a minute. That's context.

Piece 1: The company knowledge document

This is the core of it. It's one document, usually two to five pages, that answers every question a new employee would ask in their first week.

What to include:

• What you do, in one paragraph. Plain words, as you'd say it to a neighbor.
• Who you serve. Your best customers, and who you're not for.
• Services or products, with what's included and what isn't.
• Prices, or price ranges, with the date you last checked them.
• Location, hours and service area.
• Policies: booking, cancellation, refunds, warranties, payment.
• Common questions and your real answers.
• What makes you different, in specifics ("we send a photo of every finished job"), not adjectives ("we're passionate").
• Things that often get misunderstood.
• Contacts the AI may share, such as the public phone and email, and which to use for complaints.

What to leave out: customer names and personal details, passwords, bank details, staff home information, and anything you wouldn't want quoted. If the AI doesn't need it to do the job, don't give it.

Write facts, not marketing

The most common mistake is pasting in your homepage copy. "Award-winning, customer-first solutions" tells an AI nothing. "Same-day callouts for leaks in the 94110 to 94117 zip codes, $125 diagnostic fee credited against the repair" tells it everything.

Piece 2: Your documents

Next, gather the documents that already contain your knowledge:

• Your website pages, especially services, pricing, FAQ and about
• A price list or menu
• Service descriptions or spec sheets
• Policy documents (terms, refund policy, warranty)
• Onboarding guides or checklists you give staff
• Past articles or newsletters

Clean before you upload. Delete outdated price lists rather than hoping the AI picks the newest one. If two documents disagree, the AI may pick either. One current document beats three historical ones.

Prefer plain formats. Text, Markdown, simple PDFs and CSVs are read reliably. Scanned images of documents often aren't.

Piece 3: The voice guide

Facts make the AI correct. A voice guide makes it sound like you. The fastest way to build one is from your own writing, not from adjectives.

1. Collect 5 to 10 samples you're proud of. Emails to customers, posts, a good review reply, your about page.
2. Describe the pattern. Sentence length, formality, humor, how you open and sign off.
3. List words you use and words you never use.
4. Add a before/after pair. Show a generic sentence and your version.

Example voice guide:

Voice: friendly, direct, a little dry. Like a good local contractor talking to a neighbor.
Sentences: short. One idea each. Contractions are fine.
We say: "we'll", "quick fix", "here's what we'd do", "no charge for the callout if we do the job".
We never say: "solutions", "valued customer", "don't hesitate to reach out", "we're passionate about".
Exclamation marks: almost never.
Sign-off: "— The Bayline Plumbing team"

Generic: "We're committed to delivering exceptional plumbing solutions for all your needs!"
Us: "Most leaks we fix in one visit. If we can't, we'll tell you why and what it costs before we start."

Piece 4: Rules

Rules are short, absolute instructions. They matter most for the things that would embarrass you or cost you money.

Good rules are specific and testable:

• Never quote a price that isn't in the knowledge document. Say "we'll confirm after a look."
• Never promise a date or time. Offer to check availability.
• Never mention competitors by name.
• Never share or confirm any customer's details.
• If you're unsure of a fact, say so, or flag for the owner. Don't guess.
• Keep replies under 120 words unless asked.
• Always use US spelling.

Weak rules are vague: "Be professional." "Be accurate." "Use good judgment." The model already tries to do these. Rules should cover what it wouldn't otherwise know.

Add rules as you go. Every time you correct the AI's output the same way twice, turn the correction into a rule. After a month, your rules list will be the most valuable part of the setup.

Copy-ready template: your company brain

Copy this into a document, fill in the brackets, and delete what doesn't apply.

# [Business name] — company brain
Last updated: [date]

## Who we are
[One paragraph: what you do, where, for whom.]

## Who we serve
- Best customers: [describe]
- Not a fit: [describe, so the AI can politely redirect]

## Services / products
| Service | What's included | Not included | Price (checked [date]) |
|---|---|---|---|
| [Service 1] | [ ] | [ ] | [ ] |
| [Service 2] | [ ] | [ ] | [ ] |

## Hours, location, service area
- Hours: [ ]
- Address / area: [ ]
- Booking: [how customers book]

Policies

• Cancellation: [ ]
• Refunds: [ ]
• Warranty / guarantee: [exact wording, or "we don't offer one"]
• Payment: [ ]

Questions customers ask (and our answers)

1. Q: [ ] — A: [ ]
2. Q: [ ] — A: [ ]
3. Q: [ ] — A: [ ]

What makes us different (specifics only)

• [ ]
• [ ]

Often misunderstood

• [e.g. "We don't do new installs, only repairs."]

Contacts the AI may share

• General: [phone / email]
• Complaints: [name, phone / email]

Voice

[Paste your voice guide here.]

Rules (never break these)

• [ ]
• [ ]

Do not include or mention

• [Topics, people or details that are off-limits.]

## How to test that it worked

Before you let the AI do real work, quiz it. Ask ten questions you know the answers to, including a few trick ones:

1. "How much is [service]?" Does it match your price list?
2. "Are you open on [holiday]?" Does it guess, or say it doesn't know?
3. "Do you do [thing you don't do]?" Does it politely say no?
4. "Can you guarantee it'll be done by Friday?" Does it avoid promising?
5. "Write a two-line reply to a customer who's upset about a late arrival." Does it sound like you?

Every wrong answer points to a missing fact or rule. Add it and test again. That loop of test, find the gap and fill it is what "training" means in practice.

## Keep a person in the loop

Even a well-trained AI will occasionally get something wrong, usually because a fact changed and nobody updated the document. That's why anything a customer sees should pass an approval step first. Here's [why every AI agent needs an approval queue](/blog/ai-agent-approval-workflow) and how to set approval levels that loosen as trust builds.

## Doing this in Full Circle

[Full Circle](/agents) builds most of this for you. When you sign up, it reads your website and writes your company brain, which every agent you hire shares. You can upload documents (PDF, TXT, MD and CSV) to add to it, and it builds a voice guide from writing samples you provide. Each agent also gets its own name, persona, job, rules and knowledge, and when you correct one in chat, it saves the correction as a lasting rule. All finished work goes to an approval inbox before anything is sent.

Once the brain is in place, the same knowledge powers every job, from [writing and publishing articles to WordPress](/blog/ai-agent-publish-to-wordpress) to [drafting replies to Google reviews](/blog/respond-to-google-reviews-with-ai). If you'd rather not assemble it all by hand, [see the plans](/#pricing).

Either way, start with the template above. An hour spent writing down what you know will improve every AI tool you use, not just one.