— AI for small business / Private AI

An assistant that has read your manuals, and shows its sources.

Your business already wrote down most of the answers. They are in policies, price books, job files, and the shared drive nobody can search. We build a private assistant that answers from those documents, links to the exact page it used, and only shows each person what they are allowed to open.

Tell us what you need All AI services

What changes in a normal week.

Today With it in place
A new hire interrupts the same senior person twenty times a day. They ask the assistant first and get the answer with a link to the procedure it came from.
Finding last year’s quote for a similar job means searching three folders and someone’s inbox. Ask “what did we charge for a job like this?” and get the two closest past jobs with links.
The policy changed in March, and half the team is still following the old one. The assistant reads the current version, so the answer changes the day the document does.

What it does.

  • Answers with citations

    Every answer links to the document and section it came from. If the documents do not cover the question, it says so instead of guessing.

  • Respects permissions

    It only retrieves what the person asking is allowed to open in SharePoint, Google Drive, or wherever the files live. The bookkeeper sees payroll policy; the apprentice does not.

  • Stays current on its own

    New and changed documents are picked up on a schedule. Retired ones drop out. Nobody has to retrain anything.

  • Lives where people already work

    In Microsoft Teams, Slack, a page on your intranet, or a browser tab. One place, signed in with work accounts.

  • Shows you what people ask

    A monthly summary of the most common questions, and the ones it could not answer, which is a list of the documentation you are missing.

  • Customer-facing, if you want it

    The same approach can answer customers from your public documentation, product manuals, or FAQs. That version only ever sees public material.

How we build it.

Start with one team and one set of documents, prove it answers well, then widen.

  1. Pick the questions

    We collect fifty real questions your team asks, with the right answers. That list becomes the test the assistant has to pass.

  2. Gather the documents

    Connect the folders that hold the answers, and flag what is out of date. Cleaning up a few key documents usually helps more than any model choice.

  3. Build and test

    Retrieval tuned to your documents, then scored against the fifty questions. We show you the score and every miss.

  4. Pilot with one team

    Real use for a few weeks, with a thumbs-down button that sends bad answers straight to us.

  5. Widen and maintain

    More teams and folders once the scores hold. The test runs again after every change, so a fix never quietly breaks something else.

Private means private.

The assistant is only as safe as the access around it. These are defaults, not options.

  • Runs in your own cloud account or a dedicated one, not a shared multi-customer chatbot.
  • Uses model providers on business terms that exclude your data from training.
  • Retrieval follows your existing file permissions. It cannot show someone a document they could not open.
  • Canadian hosting for the document index and storage where you need it, and a written note on what is processed elsewhere.
  • Answers cite sources, and it declines to answer when the documents do not support one.
  • Every question and answer is logged for a retention period you choose, so you can audit it.

Is it for you?

A good fit when

  • You have real documentation: procedures, policies, manuals, past proposals, or a knowledge base.
  • Staff lose time looking for answers, or asking the same few people.
  • You are onboarding new staff regularly.
  • Your files live somewhere with an API: Microsoft 365, Google Drive, Dropbox, Notion, Confluence, or a website.

Probably not yet when

  • The knowledge is mostly in people’s heads. We can help write it down first, but the assistant cannot read minds.
  • You have a few dozen documents. A well-organized folder and a search box will do.
  • You need it to make decisions, not answer questions.

What it costs.

A fixed-price build for the first team and document set, then a monthly running cost that is mostly hosting and model usage, both of which scale with how much your team asks. We quote both after a short brief, and the monthly figure is based on your headcount and document volume.

What moves the number

  • How many documents, and in how many places.
  • Whether permissions have to be enforced per person, or everyone can see everything.
  • Where it lives: Teams, Slack, intranet, or a website.
  • Canadian residency requirements for the index and storage.
  • Internal only, or also answering customers.

— Questions

Before you write.

Is this the same as Microsoft Copilot?

They overlap. Copilot answers across your Microsoft 365 files and works well when your permissions are tidy. A dedicated assistant makes sense when your documents are spread across several systems, when you want tested answers with citations on a defined set of documents, or when customers need to use it. We will tell you if Copilot alone would do.

Will it make things up?

It is built to answer only from retrieved documents and to say when it cannot. We test it against your real questions before launch and after every change, and every answer shows its source so people can check. No assistant is perfect, which is why the citations matter.

Is our data used to train AI models?

No. We use providers on business terms that exclude training on your data, and the document index lives in an account you control. You can delete it at any time.

Can it stay in Canada?

The document store and search index can be hosted in Canada. Some model processing may happen outside Canada depending on the model you choose, and we document exactly which parts run where so you can decide.

Do you have a working example?

Yes, this site. Ask SetKernel answers questions from our own published writing and cites the pages it used. Try it, then imagine it reading your procedures instead.

The other three AI services.

Further reading Try Ask SetKernel, our own versionThe engineering view: RAG developmentHow we test that answers are correct

— Engage

Where do your answers live today?

Tell us roughly how many documents you have, where they live, and who would use the assistant. We reply in writing within one business day with whether it is a fit and what the first version would cover.

Tell us what you need Ask a question first
Esc