Ask a general purpose AI assistant how your company handles a warranty claim and it will give you a confident, well written, completely made up answer. It has never seen your warranty policy. It is guessing at how warranty policies usually read, in beautiful full sentences. That is not the tool failing. That is the tool doing what it was built to do with the information it was given, which was nothing about you.
The interesting move for a small business is not a better assistant. It is a better supply of information. Point an AI tool at your own procedures, past proposals, product specs, and support answers, and the questions people ask all day start getting answered correctly. We call this building a company brain. The hard part is not the technology. The hard part is your filing.
What A Company Brain Actually Is
The pattern has a name in the industry. Microsoft’s documentation describes retrieval augmented generation as “a pattern that extends LLM capabilities by grounding responses in your proprietary content.” Strip the vocabulary: when someone asks a question, the system searches your documents first, pulls the relevant passages, and hands them to the AI along with the question. The AI answers from the material in front of it instead of from memory.
That matters more than any other detail here. The model is not retrained on your data. Your documents are handed to it at the moment of the question, the way you would hand a new hire the binder before asking them to answer the phone. Change a document and the next answer changes.
For a small business, the useful content is boring and you already have it.
- Standard operating procedures. How to open a ticket, process a return, handle a disputed invoice, and who approves what.
- Past proposals and statements of work. Your scope language, your exclusions, the phrasing your best salesperson uses.
- Product and service specifications. Model numbers, tolerances, lead times. The answers your team gets today by walking over and asking one specific person.
- Resolved support conversations. Often the highest value content in the company, and almost always buried in mailboxes.
- Policies and the handbook. Time off, expenses, travel, security rules. These generate huge volumes of repetitive questions to people whose job is not answering them.
Why It Beats A General Assistant Internally
A general assistant is good at general work: drafting, summarizing, rewriting. It cannot know that your company charges a restocking fee on custom orders but not stock items, because that fact exists in exactly one place on earth and it is your operations folder.
The risk is not that the general tool says “I don’t know.” It is that it does not say that. NIST’s Generative Artificial Intelligence Profile, published in July 2024, defines confabulation as “a phenomenon in which GAI systems generate and confidently present erroneous or false content,” and notes that “humans may over-rely on GAI systems or may unjustifiably perceive GAI content to be of higher quality than that produced by other sources,” which it identifies as automation bias.
Grounding does not eliminate that risk, but it changes its shape. A grounded answer can cite the document it came from. An ungrounded answer offers nothing to check, which is why it is so easy to believe.
The Housekeeping That Decides Whether This Works
This is the part vendors leave off the slide. A company brain is a mirror. If your shared drive holds four versions of the return policy, your AI answers confidently from whichever it finds, and is right about a quarter of the time.
- Decide what is in and out, in writing. Current approved procedures in. Drafts, personal working files, and anything with employee records or customer financial details out, unless you have deliberately planned for it.
- Kill the old versions. Move superseded documents where the system does not look, or delete them. Every stale copy in scope is a wrong answer waiting to be given.
- Put a date and an owner on every document. That turns maintenance into a task, and gives your team a fast way to judge an answer.
- Write for retrieval, not for prose. Clear headings, one topic per document, plain language. If a document is hard for a new hire to skim, it is hard for a retrieval system to use.
- Schedule a review cadence. Quarterly is enough, on someone’s calendar with their name on it, because the whole thing decays quietly.
Most businesses say the cleanup was worth doing on its own, independent of the AI.
Who Gets To Ask What
This is where these projects go sideways, and fast. Index everything, let everyone ask, and you have built the world’s most efficient tool for surfacing the salary spreadsheet somebody saved to the wrong folder in 2019. Search makes people work to find things. A conversational tool hands it over on request.
The AI must respect the permissions you already have. Microsoft’s documentation on document level access control states that with the integration in place, “only users or groups that have access to the source item in SharePoint can retrieve it from search results.”
Two cautions. First, that only helps if your existing permissions are correct, and in most small businesses they are not, because folders accumulate access the way garages accumulate boxes. Audit permissions before you index, not after. Second, permission changes are not always instant. Microsoft notes such changes “are only reflected in search results after that metadata is synchronized to the index.” Confirm how that works with your vendor.
One more thing. Without an approved tool, people paste company documents into whatever free assistant is open in another tab, the dynamic we describe in shadow IT and the hidden risk inside your own organization. A sanctioned company brain is partly a security control.
Start With One Department
Do not index the company. Pick one department with a high volume of repetitive internal questions, a manageable document set, and at least one genuinely enthusiastic person. Support and operations are usually the best candidates. Sales is a close second, because proposal language is so reusable.
Before you turn it on, write down twenty real questions that department gets, with the correct answers, and ask the system all twenty. That is your test set, and it tells you more than any demo will. If it gets sixteen right and cites its sources, you have something useful. If it gets nine, your documents are the problem.
Then teach people to use it like a well read new colleague rather than an oracle. NIST’s Generative AI Profile recommends organizations “review and verify sources and citations in GAI system outputs,” which here means: click the link and confirm the document says what the answer claims. We cover the wider version of that in how to actually prepare your team for AI without the hype.
The Bottom Line
A company brain is not a moonshot. It is a search problem with a conversation on top, and the value comes from the documents, not the model. Harrison likes to say AI is the dumbest it will ever be today, which is the reason to start now. The tools keep improving on their own. Nobody else will clean up your procedures folder.
Start narrow. One department, a defined document set, permissions checked first, twenty test questions, a named owner, a quarterly review. Capabilities and pricing move constantly, so confirm current features with any vendor before you commit.
If you want help scoping this without buying something you do not need, we can review what your current platform already includes, sort out permissions before anything gets indexed, and get one department running on a pilot you can measure. We serve small and mid sized businesses across Denton County and North Texas. Contact us today
Sources:
Comments are closed