Answer accuracy on the test question set
Checked before launch and after major content changes, so quality is measured, not assumed.
Protect margin
An internal knowledge assistant built on your own documents
Every established service business has a few people everyone asks. Where is the template for this? What is our policy on that? How did we handle a job like this last year? Those people are often your most senior and most expensive staff, and every interruption costs them focus. When they go on holiday or leave, the knowledge goes with them.
Signs you need this
- Senior staff are interrupted many times a day with questions that are answered somewhere in your files.
- New hires take months to become independent because the know-how is spread across people and folders.
- Staff use different versions of the same template or procedure, and nobody is sure which is current.
- Your shared drive has years of folders with unclear names, and search rarely finds the right document.
- Staff already paste internal documents into public AI tools to get answers, without a policy.
An internal knowledge assistant lets staff ask questions in plain English and get answers drawn from your own procedures, policies, templates and approved past work, with links to the source. It is only as good as the documents behind it. Most of the work is deciding what goes in, cleaning it up and setting permissions, not choosing the AI model.
How it works
Choose the questions to answer
Collect the common questions staff actually ask, from onboarding, operations and client delivery. This list defines the scope and becomes the test set for checking whether the assistant gives good answers.
Gather and clean the source documents
Identify the current procedures, policies and templates that answer those questions. Remove outdated versions, fix obvious errors and give each document an owner and review date.
Set permissions before connecting
Decide who should see what. HR records, salaries, confidential client matters and partner documents need restricting. The assistant must respect the same access rules as your file system.
Connect an assistant that cites sources
Use a tool such as Microsoft 365 Copilot, ChatGPT Business, Claude for business teams, Google's Gemini in Workspace or a custom setup that searches your documents and shows where each answer came from.
Test, launch and keep content current
Check answers against the test questions before launch, then gather feedback on wrong or missing answers. Document owners update sources on a schedule, because stale documents produce stale answers.
Ways to do it
| Approach | Best fit | Notes |
|---|---|---|
| Better structure and search in tools you already own | Firms whose main problem is disorganized files rather than a lack of AI. | A clear folder structure, naming rules and a single procedures hub in SharePoint, Google Drive or Notion may solve most of the problem. It also prepares the ground for an assistant later. |
| AI built into your office suite | Businesses on Microsoft 365 or Google Workspace with documents already stored there. | Microsoft 365 Copilot and Gemini for Workspace can answer from documents the user can access. Permissions matter, because the AI can surface anything that user can see. Check current licensing costs. |
| Business AI workspace with curated documents | Teams that want a focused assistant over a defined set of approved documents. | Business tiers of ChatGPT and Claude support shared projects or custom assistants with uploaded files. Keeping the document set small and curated reduces wrong answers. |
| Custom retrieval build | Large document sets, strict data residency needs or integration with practice management systems. | A custom retrieval-augmented generation setup gives more control but needs ongoing maintenance. Start with off-the-shelf options to learn what staff ask. |
How we deliver this: AI automation for small business or fractional CTO leadership, scoped in writing and run over 90 days. See pricing or the sample scope.
What we would measure
Interruptions to senior staff
A rough weekly count, or a short survey, shows whether the assistant is taking load off key people.
Time to independence for new hires
If onboarding gets faster, the assistant is spreading knowledge as intended.
Share of answers flagged as wrong or unhelpful
Feedback from users highlights gaps in documents and helps prioritize updates.
Risks and how we handle them
| Risk | Mitigation |
|---|---|
| The assistant states confident answers that are wrong or based on outdated documents. | Require cited sources, tell staff to check the source for anything important, remove outdated versions and review flagged answers weekly. |
| Over-broad permissions let staff see confidential HR, financial or client information through the assistant. | Audit sharing permissions before switching on AI search, restrict sensitive folders and test with accounts at different access levels. |
| Client confidential data, such as privileged legal material or health records, handled outside agreed safeguards. | Use business-grade tools with suitable data terms, keep regulated data out unless your compliance lead approves, and follow guidance such as ABA Formal Opinion 512 where it applies. |
When not to do this
If your procedures are not written down, an assistant has nothing to draw on. Document the core processes first. It is also premature if your file permissions are a mess, since AI search can expose what was previously just hard to find. And for a team of five who sit together, asking a colleague may simply work better.
Common questions
What is an internal knowledge assistant?
It is an AI tool that answers staff questions using your own documents, such as procedures, policies, templates and approved past work, instead of general internet knowledge. Good ones show which document each answer came from so staff can check. It works best on a curated, current set of documents.
Will it make things up?
It can. AI models sometimes produce confident answers that are wrong, especially when the right document is missing or outdated. Citing sources, curating documents and training staff to check important answers reduce the risk. We would never treat the assistant as the final word on policy or client advice.
Is Microsoft 365 Copilot enough, or do we need something custom?
For many firms on Microsoft 365, Copilot or similar built-in tools are a sensible starting point because they work with existing permissions and files. A custom build makes sense when you need tighter control over sources, specific integrations or behavior the built-in tools cannot provide.
How do we keep the information current?
Give each source document an owner and a review date, remove outdated versions from the assistant's sources, and act on feedback when answers are wrong. Treat it like maintaining a procedures manual, because that is essentially what it is.
Can clients use the same assistant?
We would keep internal and client-facing assistants separate. Internal documents often include information clients should not see, and client-facing answers need stricter review. A public website assistant should use only content approved for publication, with clear limits on what it can answer.
By industry
Internal knowledge assistant for your industry
Accounting and CPA firms
Help staff find review checklists, workpaper standards, client notes and prior-year decisions fast, with cited answers.
See the details →Architecture firms
Help staff find your firm's standards, details, spec sections and lessons learned in seconds, with every answer linked to its source.
See the details →Consulting and professional services firms
Help consultants find past deliverables, frameworks and proposal language across SharePoint and Drive, with confidentiality controls. Talk with InTheory.
See the details →Engineering firms
Let engineers find design criteria, firm standards and past calculation approaches in your own documents, with sources linked for checking.
See the details →IT managed service providers
Let MSP techs search runbooks, client docs and past ticket fixes in plain language, with access controls.
See the details →Law firms
Help associates and staff find the firm's forms, prior work product and procedures fast, with cited answers and matter-level access.
See the details →Small manufacturers
Help operators and staff find work instructions, customer requirements and past job notes fast, with answers citing controlled documents.
See the details →Nonprofit organizations
Help nonprofit staff and volunteers find policies, procedures and grant boilerplate from your own documents. Talk with InTheory about a pilot.
See the details →Guides
Getting staff to adopt new toolsMicrosoft Copilot for small businessCybersecurity basics to fix before adopting AIWriting standard operating procedures with AI, and keeping them trueBuilding an internal knowledge base your team will actually useAI for small business: an honest guide for service firms
Questions
How do you train employees on AI?Is ChatGPT safe for confidential information?Copilot vs ChatGPT: which is better for business?
Glossary
Large language model (LLM)Retrieval-augmented generation (RAG)Prompt engineeringModel Context Protocol (MCP)Custom GPTEmbeddingsVector databaseAI copilot
Other solutions
KPI and pipeline dashboardsSoftware stack auditDocument intake and data extractionInvoicing and collections follow-upJob scheduling and dispatchHiring and staff onboardingAI usage policy and governanceSystems integrationContract and document review supportCompliance documentationTechnology vendor selection
Next step
Bring one expensive workflow. Leave with a clear next step.
In 30 minutes we look at one recurring workflow, what it costs you today and whether there is a credible case to improve it. If there is not, we will say so.
Book a 30-minute workflow review
Not ready to talk? Take the 3-minute check
30 minutes with Phil or Martin. No technical brief, no pitch deck. Phil or Martin replies personally, usually within one business day.