How to train AI on your business data (and keep it private)

Rolling outDay-Zero Study: running on our own companies now. Opening to new accounts soon.

You bought a tool that promised to “know your business.” The next screen asked you to describe the business in a box, upload a handbook nobody updates, and paste three sample replies. You are now the trainer. The tool is waiting. You had hoped it would be the other way around.

That is the honest shape of most advice on how to train AI on your business data. It is a pile of homework: write a prompt, curate documents, tag articles, hope the next answer uses them. Some of that work is useful. None of it is the same as a new person sitting down with the real history of how you already operate.

There is also a privacy question hidden in the word training. Teaching a private memory that your agents can cite is one thing. Sending your books, tickets, and policies into a public model that retains them, or learns from them, is another. The steps below separate those options. They also say what still will not be true no matter how much you upload.

Prompts: a note on the fridge

A prompt is a standing instruction. Tone, a few policies, topics to avoid, how you sign off. It is worth writing. It is not training.

The ceiling shows up fast. A prompt long enough to cover real life becomes a document no one maintains. The model still has not seen the last two hundred tickets, the vendor you always file under software, or the refund you never give. When two lines in the prompt conflict, you get a fluent compromise. Customers and books do not want a compromise. They want the rule you actually use.

Use a prompt for boundaries that rarely change. Do not use it as the place you try to store the company. If the only “training” a product offers is a bigger text box, you will still be the one answering from memory every time the box falls short.

Uploaded documents and knowledge bases

The next option is a library. Policies, SOPs, product sheets, a help center, a few golden replies. The tool searches that library when a question arrives. This is better than a prompt, and it has its own failure mode.

Most companies do not run on the documents they meant to write. They run on the exceptions in old tickets, the way a bookkeeper recodes a vendor, the paragraph you send when a package stalls. A beautiful handbook from last spring can be wrong today. If the knowledge base is what the tool trusts, it will quote the stale page with confidence.

Someone also has to decide what is in the library, what is outdated, and what should never be uploaded at all. That someone is usually you. Exports of entire inboxes and ledgers, dropped into a folder, are not a system. They are a leak with a progress bar.

If you use a knowledge base, keep it small and owned. Put current policies in. Take dead ones out. Prefer the tool reading live history, with permission, over a dump of files you will forget you shared.

Connecting tools so it reads your history

The third option is the one that matches how people actually learn a company. You connect the help desk, the store, the books. The role reads what you already did. It does not ask you to narrate the business from scratch.

Read comes first, and read-only is the right default. A new connection should not gain the power to send, refund, or recategorize just because it can see. You confirm what each role may read. A Bookkeeper sees the books. Customer Service sees the help desk. Neither gets the other’s private channels unless you say so.

History is the lesson. For support, that means past tickets: tone, sign-off, what you refund, what you never promise. For books, that means up to a couple of years of how you categorized vendors, not a generic chart of accounts. The role can start helping while it learns, as long as helping means a draft or a suggestion you approve.

This is the onboarding analogy that holds. A new hire reads every SOP first, then watches how the work is really done, then asks. You would not hand a stranger your bank feed and your send button on the morning they learn your name. Do not hand those to a model either. The Read-only first piece is that sequence in plain language.

	Good for	Watch for
Prompt	Tone and a few rules	It goes stale
Uploaded docs	Current policies	Quoting an old page
Connected history	How you really work	Conflicting past answers

“Training a model” is a different claim

Be precise about the word train. Fine-tuning or contributing your data to a model that serves other people is not the same as giving your own agents a private memory.

Questions worth asking any vendor, in writing:

• Are AI calls run with no-training and zero-retention settings, and does every provider they use support them?
• Is my data used to train a model, theirs or anyone else’s?
• Can another customer’s workspace see my documents, tickets, or lessons?
• Who at the vendor can open my data, and for what?
• Can I read, change, and download the memory?
• What happens to uploads if I disconnect or leave?

A private company memory is lookup, not a new brain for the world. Agents cite a decision, a document, a policy, or a lesson. They do not absorb your ledger into weights you can never inspect. Allmata’s version of that memory is the Brain: one per company, shared by every agent in that company, and it does not reach the next company. More of the promise and the limits is on security.

Diagram

Dated lessons matter more than a one-time upload. When you correct a draft or a category, that answer should stick. When you change your mind, the newer lesson should win and the older one should stay archived, not vanish. Otherwise you are training the tool every week with the same conversation. That is not learning. That is a queue.

What no amount of training gives you

Reading history will not make a fuzzy policy sharp. If you have answered the same complaint five different ways, the role will see the conflict. You still have to pick the rule. Training is not a substitute for a decision you have avoided.

It will not make tax judgment appear. Books history can teach categorization. It cannot teach what you owe, and it should not try. It will not make a dangerous promise safe. If some sentences must never go to a customer, those sentences belong on a list the role can follow, and a person still approves what is sensitive.

It will not know facts it has not been given. A model that “fills in” a tracking number, a price, or a policy is not trained. It is guessing. Guessing is the failure mode you are trying to remove.

And it will not be finished on the day you connect. The first read can be wrong about what matters. You should see that write-up and correct it before anyone treats it as how the company runs. That correction is part of training. Skipping it means the tool learned a cartoon of you.

How Allmata does it

You should not have to perform the training as a second job. Roles connect read-only, then read the history those tools already hold. Customer Service reads past tickets and drafts from them. The Bookkeeper learns from up to 24 months of your categorization before it asks. Each thing you confirm becomes a dated lesson in the Brain. Point it at a Google Drive folder of SOPs and it reads those too, after showing you the price and waiting for your OK.

The Day-Zero Study is rolling out. Once it’s on for you, before an agent talks to you, it reads your business history and writes how your company runs. Its first message is your Day-Zero briefing, which you confirm or correct. A readiness exam on your own past answers, before the role works live, is rolling out as well. More on the Day-Zero Study page. The path from connect to approve is on how it works.

1. ConnectRead-only, tool by tool.
2. ReadPast tickets. Up to 24 months of categories.
3. BriefingDay-Zero briefing of how the company runs.
4. You decideConfirm it or correct it.
5. RememberDated lessons in your private Brain.

Model calls that work with your business data go through one gateway with no-training and zero-retention settings, where the provider supports them. Your data is never used to train anyone’s model. You can read, change, and download your Brain.

Where to start

Pick one workflow you can check, not the whole company. The last month of uncategorized transactions, or a week of order-status mail. Connect it read-only. See whether the role describes your real practice, cites a source, and asks when it is short of one.

If you want a short map of what can move this week before you connect anything, take the free AI Workforce Blueprint. If you are deciding whether the memory stays private once you do, read can I trust AI with my business? next, then the trial when the answers are good enough. The 14-day trial includes $10 of AI work on us, with no card required.