# AI Knowledge Bases: A Source Checklist for Irish Teams

An AI knowledge base should use approved, current information and respect who is allowed to read it. Digital Bridge in Gorey, Co. Wexford helps Irish teams separate public answers, internal guidance and customer data before connecting an assistant. Begin with a source inventory, document owners, access rules and a clear route for unanswered questions.

Written by Joey Bray, Founder of Digital Bridge. [Author profile](https://digitalbridge.ie/about/joey). Updated: 20 October 2026.

## Which documents belong in the source library?

Start with material that answers a defined business question. Public service descriptions, approved operating instructions and product guidance may be useful. A folder is not suitable simply because it contains a large amount of text. Old drafts and contradictory policies can produce convincing but wrong answers.

For each document, record its owner, intended audience and review date. Identify the version that is approved for use and remove superseded copies from the active source set. Keep a controlled archive where needed, separate from the material used to assemble answers.

## How should public, internal and personal information be separated?

Treat these as different access classes. A public visitor may need to know which service is available. A staff member may need an internal process. A customer record may contain personal information that neither should see through a general assistant.

- Public material: information already approved for website visitors.
- Internal material: instructions restricted to the appropriate team.
- Personal material: information requiring a specific purpose and access assessment.
- Sensitive or special-category material: information requiring additional care and professional advice where appropriate.

A system prompt saying “keep this confidential” is not an access-control design. The software must enforce the user’s permission before it retrieves a document or record. Test the boundary with accounts that should be denied, not just an administrator who can see everything.

## What should you ask about the AI provider?

The Irish Data Protection Commission recommends understanding what personal data an AI system uses, how it uses it and where it goes. Ask whether the provider retains inputs, reuses them, allows deletion and relies on other processors. Check the product and account terms that actually apply to your deployment.

A promise of an EU location does not answer every data-protection question. A privacy assessment also concerns purpose, access, retention and the rights of the people involved. Seek appropriate advice for high-risk uses rather than presenting an assistant as automatically compliant.

## How do you stop unsupported answers?

Require the assistant to use the approved source material and make its limits clear. If the source is absent, conflicting or out of date, it should say that the answer needs review. A confident sentence without support is not a substitute for the missing document.

Test questions should include facts that are present and facts that are deliberately absent. Add a question about a retired policy, an ambiguous service and an instruction asking for another customer’s information. Check both the answer and the source records that were retrieved.

In our published approach to knowledge-base assistants, the source boundary matters more than a larger document collection. A smaller approved set is easier to review and maintain. That is an operating principle, not a claim about a measured improvement or a completed client project.

## What is a useful preparation example?

Imagine a small Irish team preparing an internal onboarding assistant. This is an illustrative scenario. The source set includes current opening procedures and approved staff guidance, but excludes employee files and private customer notes. Each instruction has a named owner responsible for updates.

The assistant answers process questions using that set and points to the relevant document. A request about a customer’s personal circumstances is routed to the appropriate person instead of being answered from the shared library. This shows why the intended audience must be defined before importing documents.

## What should the handover include?

A knowledge base needs continuing ownership. Explain who can add sources, who approves a replacement and what happens when an old policy is withdrawn. Agree a process for corrections from the people using the system.

- A documented list of active sources and their owners.
- Permission tests covering allowed and denied access.
- A retention and deletion process appropriate to the data used.
- A human contact for missing, disputed or sensitive answers.
- A way to pause the assistant if the source set is no longer reliable.

Review logs carefully. They can help investigate an answer but may themselves contain personal information. Collect only what is justified, restrict access and agree a retention period rather than keeping everything indefinitely.

## Where can you get practical help?

Read about our [AI knowledge-base assistant service](/ai-integration/ai-knowledge-base-assistant), discuss the source and access assessment through [AI consultancy](/services/ai-consultancy), or explore [AI integration](/ai-integration). For actions that follow an answer, the [human-approval checklist](/blog/ai-workflow-human-approval-checklist-ireland) explains the review boundary.

## At a glance: Four tiers of source material

Public material sits at the base, then internal, personal and sensitive material, with access narrowing at each level. Source: Digital Bridge, 20 October 2026.

| Tier | Who it is for |
|---|---|
| Public | Already approved for website visitors |
| Internal | Restricted to the appropriate team |
| Personal | Needs a specific purpose and access assessment |
| Sensitive | Additional care and professional advice |


Diagram alt text: Public material sits at the base, then internal, personal and sensitive material, with access narrowing at each level.

## Is this guide right for your business?

**It fits when:**

- Staff or customers ask the same questions that your documents already answer.
- Someone can own each source and keep it current.

**It is not the right first step when:**

- Your documents conflict or are out of date and nobody owns them.
- The main use would be answering from personal or sensitive records without an assessment.

**Have these answers ready before you call:**

- Which documents are approved and current?
- Which tier does each source belong to?
- Who handles missing or disputed answers?

**Next action:** call Joey on 085 224 1848 or read about our [AI knowledge-base assistant](/ai-integration/ai-knowledge-base-assistant). Digital Bridge confirms scope in writing before any work starts.

What do people ask about AI knowledge bases?

### What are examples of AI knowledge bases?

Typical examples are a staff assistant that answers questions from internal procedures, a website assistant that answers customers from approved service pages, and a support tool that suggests replies from a help centre. Each one only works as well as the documents it is allowed to read.

### What is the best AI knowledge base tool for a business?

The best tool is the one that respects your access rules, lets you control which documents it reads, shows where each answer came from and fits your data-protection requirements. Compare tools against your own source checklist rather than a general ranking.

### Can I build an AI knowledge base for free?

Many tools offer free tiers or trials, which are useful for testing with public documents. Before adding internal or personal material, check the provider's terms, where data is stored, how long it is kept and whether you can export or delete it later.

### Can an AI knowledge base use customer records?

Only after assessing the purpose, access permissions, provider terms and applicable data-protection requirements. Do not add customer records to a public or general staff assistant by default.

### What should the assistant do if a source is missing?

It should explain that the answer is not supported and route the question to a person. It should not invent a policy or customer detail.

### Who keeps the documents up to date?

Each source needs a named business owner and a review process. Replacing or withdrawing a document should update the active source set and be tested.
### How many documents does an AI knowledge base need?

Fewer than most teams expect. Start with the approved documents that answer your most common questions, such as service details, policies and procedures. A small, current and owned source set produces more reliable answers than a large folder of drafts, duplicates and outdated files.

### Do we need a data protection impact assessment?

Possibly. Ireland's Data Protection Commission says an assessment is required where processing is likely to result in a high risk to individuals. If the assistant will use personal or sensitive information, review the DPC guidance and take advice before connecting those sources.

### What happens when a customer asks for their data to be erased?

The erasure must reach every place the data is held, including any source the assistant can read. Under GDPR Article 17, individuals can request erasure in defined circumstances, so your handover should include a tested process for removing a person's records from the active source set.

## Which sources support this guide?

- [DPC: AI, LLMs and data protection](https://www.dataprotection.ie/en/dpc-guidance/blogs/AI-LLMs-and-Data-Protection)
- [DPC: Data protection impact assessments](https://www.dataprotection.ie/en/organisations/know-your-obligations/data-protection-impact-assessments)
- [DPC: Right to erasure (Articles 17–19 GDPR)](https://www.dataprotection.ie/en/individuals/know-your-rights/right-erasure-articles-17-19-gdpr)
## What happens next?

Discuss the task with [Digital Bridge](/contact) and confirm the deliverables, exclusions, operating responsibilities and any ongoing software costs in writing. Work starts after the scope is agreed. Read about real published work in our [portfolio](/our-work), including the Printhouse custom web-to-print platform. No result or funding approval is promised.

[Get an instant estimate](https://digitalbridge.ie/#ai-estimator) · [Call 085 224 1848](tel:+353852241848) · [Email info@digitalbridge.ie](mailto:info@digitalbridge.ie)

Canonical: https://digitalbridge.ie/blog/ai-knowledge-base-source-checklist-ireland