How to Train an AI Voice Agent on Your Business Knowledge Base

Your AI Is Only as Good as What You Teach It

Here is the uncomfortable truth about AI voice agents: the technology is not the bottleneck. The knowledge base is.

A perfectly configured voice agent with a poor knowledge base will give wrong prices, misquote availability, and frustrate every caller. A basic voice agent with an excellent knowledge base will handle 85% of calls flawlessly — which is why knowledge preparation is a critical step in any AI automation strategy. Before you start building your knowledge base, it is worth assessing whether your data and processes are ready — our AI readiness framework helps you evaluate this in five minutes.

Most businesses that struggle with voice AI did not have a technology problem. They had a preparation problem. They switched on the system without properly documenting what the AI needed to know. The AI guessed. The AI got things wrong. The business owner decided "AI is not ready."

It was ready. They were not.

This guide fixes that. Whether you are setting up a voice agent for the first time or improving an existing deployment, this is the step-by-step process to train an AI voice agent on your business knowledge so it answers calls like your best employee — on their best day, every day, at every hour.

For the complete voice AI overview, read AI Voice Agents for UK Businesses: The Complete Guide. For the setup process, see Setting Up an AI Voice Agent for Your Business: What to Expect.

What Is a Knowledge Base (And Why Does Your AI Need One)?

A knowledge base is a structured collection of information that the AI references when answering questions. Think of it as the training manual you would give a new receptionist on their first day — except it needs to be more thorough, because the AI cannot improvise.

A human receptionist can figure out a reasonable answer to a question they have never heard before. They use common sense, context, and experience. An AI voice agent cannot. If the information is not in the knowledge base, the AI either gives an incorrect answer (bad), refuses to answer (less bad), or escalates to a human (ideal, but defeats the purpose if it happens too often).

The goal when you train an AI voice agent is to give it enough information to handle 70–85% of calls autonomously, while knowing precisely when to escalate the remaining 15–30%.

What the Knowledge Base Contains

Step 1: Audit Your Current Call Data

Before you write a single FAQ, understand what your customers actually call about. This is not guesswork — it is data.

How to Audit

If you have call recording: Listen to 50–100 recent calls. Categorise each one. You will find that 80% of calls fall into 8–12 recurring categories.

If you do not have call recording: Ask your receptionist or team to keep a tally for two weeks. Every call, tick a box: pricing enquiry, booking, status check, complaint, general question, other. The pattern will emerge quickly.

If you are a solo operator: You already know. You answer the same five questions every day. Write them down.

Common Call Categories by Industry

Your call audit will reveal the exact knowledge gaps your AI needs to fill. Do not skip this step — it is the foundation everything else builds on.

Step 2: Build Your FAQ Document

This is the single most important component of your knowledge base. The FAQ document is where you train an AI voice agent to handle the majority of routine calls.

Structure Each FAQ Entry

For each question, provide:

1. The question (as a customer would ask it — in natural language, not corporate-speak)
2. Alternative phrasings (the same question asked differently)
3. The answer (clear, complete, conversational)
4. Follow-up actions (what should the AI do after answering?)

Example FAQ Entry

Question: "How much does a screen replacement cost?"

Alternative phrasings:

• "What is the price for a new screen?"
• "How much to fix a cracked screen?"
• "Screen repair cost?"
• "My screen is smashed — what will it cost?"

Answer: "Screen replacement costs vary by device model. For example, an iPhone 15 screen replacement is £129, an iPhone 15 Pro is £159, and a Samsung Galaxy S24 is £149. All repairs include a 90-day warranty. Would you like me to check the price for your specific model?"

Follow-up action: If the caller provides a model, look up the price. Then offer to book an appointment.

How Many FAQs Do You Need?

• Minimum viable: 20 FAQs covering your top call types
• Good: 30–50 FAQs with alternative phrasings
• Excellent: 50–100 FAQs with detailed answers, edge cases, and follow-up logic

More is better, but start with the 20 questions that represent 80% of your call volume. You can expand the knowledge base over time as you identify gaps from real calls.

Step 3: Document Your Pricing

Pricing is the number one reason customers call service businesses. Get this wrong, and the AI will actively damage your business.

Pricing Documentation Format

Create a structured pricing document that covers:

Example: Phone Repair Pricing

Handling Price Uncertainty

Not every service has a fixed price. For variable pricing, train the AI to give a range and set expectations:

"Water damage repair starts from £79 for a basic clean and dry. If components need replacing, the cost will depend on which parts are affected. We would need to inspect the device first — shall I book a diagnostic appointment?"

This is honest, helpful, and moves the caller toward a booking. Never let the AI invent a price or give guarantees it cannot back up.

Step 4: Define Your Processes

The AI needs to understand how your business operates — not just what you offer, but how you deliver it.

Key Processes to Document

Booking process:

• How far in advance can customers book?
• Do you require a deposit?
• What information do you need to create a booking? (Name, phone, email, service type, device model)
• Can customers book same-day?
• What is the cancellation policy?

Service delivery process:

• What happens when the customer arrives?
• How long does a typical service take?
• Do you offer wait-while-you-repair or drop-off?
• How is the customer notified when the work is complete?

Payment process:

• What payment methods do you accept?
• Do you take payment upfront or on completion?
• Do you offer payment plans?

Complaint handling process:

• What should the AI say to a dissatisfied customer?
• When should it escalate to a manager?
• What goodwill gestures (if any) is the AI authorised to offer?

Document these as clear decision trees. The AI follows logic, not intuition.

Step 5: Set Escalation Rules

This is where you train an AI voice agent to know its own limits — arguably the most important part of the entire knowledge base.

Escalation Triggers

Define specific scenarios that should always be transferred to a human:

Escalation Quality

When the AI transfers a call, it should pass along:

• Caller's name and contact number
• Summary of what was discussed
• The specific reason for escalation
• Any relevant account or job numbers

This is what makes AI escalation superior to IVR transfers. The human receiving the call has full context. The caller does not repeat themselves.

Step 6: Define Boundaries and Tone

What the AI Should Never Do

• Never invent information. If the answer is not in the knowledge base, the AI should say so and offer to find out.
• Never guarantee outcomes. "We will definitely fix it" should be "we will do everything we can."
• Never discuss competitors. The AI should not comment on other businesses, positively or negatively.
• Never provide medical, legal, or financial advice. General information only; qualified advice requires a human professional.
• Never share internal pricing logic. "Our markup is..." is never appropriate.
• Never make promises about timelines it cannot verify. "Your repair will be done by 3pm" should be "repairs of this type typically take 45 minutes to an hour."

Tone and Voice

To train an AI voice agent to match your brand, document your preferred tone:

• Formal or conversational? Most UK service businesses do best with "professionally friendly" — clear and polite without being stiff.
• First name basis? Should the AI use "Mr Porter" or "James"?
• Industry-specific language? Should the AI use technical terms or plain English?
• Humour? Generally best avoided unless your brand is explicitly casual.

Provide example phrases the AI should use:

• Instead of "I don't know," say: "That is a great question — let me find out for you."
• Instead of "We can't do that," say: "That is something I would need to check with the team. Can I arrange a callback?"
• Instead of "Your call is important to us," say nothing of the sort. Just help them.

Step 7: Test Relentlessly

A knowledge base that has not been tested is a knowledge base that will fail in production. Here is how to test properly.

The 50-Call Test

Before going live, run 50 test calls covering:

• 20 calls for your top 5 FAQ categories (4 each) — does the AI answer correctly?
• 10 calls with unusual phrasing — "I need that thingy where you replace the glass on me phone" — does the AI understand?
• 10 calls designed to reach the AI's limits — questions not in the knowledge base — does it escalate gracefully?
• 5 calls with emotional callers — frustration, urgency, confusion — does the AI remain composed and helpful?
• 5 edge cases — extremely specific questions, multiple issues in one call, rapid topic switching

What to Look For

Iteration Cycle

After each round of testing:

1. Identify gaps (questions the AI handled poorly)
2. Add new FAQ entries or refine existing ones
3. Re-test the specific scenarios that failed
4. Repeat until the 50-call test has a 90%+ success rate

This process typically takes 3–5 days. It is the most important phase of deployment.

Step 8: Maintain and Improve Over Time

A knowledge base is not a set-and-forget asset. Your business evolves. Prices change. Services are added. Policies are updated. The knowledge base must keep pace.

Weekly Maintenance (15 minutes)

• Review the AI's escalation log — what questions caused transfers? Can any be added to the knowledge base?
• Check for pricing changes and update the database
• Review call transcripts for common phrases the AI misunderstood

Monthly Maintenance (1 hour)

• Analyse call category distribution — are new patterns emerging?
• Add FAQs for questions that have appeared multiple times
• Update seasonal information (holiday hours, seasonal promotions)
• Review upsell performance — which offers convert and which do not?

Quarterly Maintenance (2 hours)

• Full knowledge base audit — is everything still accurate?
• Review competitor offerings — has the market shifted?
• Update brand messaging if business positioning has changed
• Expand the knowledge base into areas that previously required escalation

Keeping on top of these maintenance cycles is far easier when you have a single view of performance data. AmpliDash surfaces escalation logs, call category trends, and knowledge base gap analysis in real time, so you always know exactly where your AI voice agent needs improvement.

Using Company Cortex to Power Your Knowledge Base

Building and maintaining a knowledge base manually is effective but labour-intensive. Company Cortex automates the process.

Company Cortex is Ampliflow's RAG (Retrieval-Augmented Generation) knowledge base system. It ingests your existing business documents — website content, price lists, policy documents, training manuals, email templates — and structures them into a queryable knowledge base that your AI voice agent draws from in real time.

How Company Cortex Differs from Manual Knowledge Bases

When your voice agent is powered by Company Cortex, the same knowledge that answers phone calls also powers your website chatbot, email auto-responder, and WhatsApp replies through Amplio. One truth. Every channel.

Common Mistakes When Building a Knowledge Base

Mistake 1: Writing for Humans, Not AI

Your knowledge base is not a brochure. It is training data. Be explicit, structured, and unambiguous. "Prices vary" is useless. "iPhone 15 screen replacement: £129. iPhone 15 Pro: £159. iPhone 15 Pro Max: £189" is useful.

Mistake 2: Forgetting Alternative Phrasings

Customers do not ask questions the way you write them. Include multiple phrasings for each FAQ. "What are your hours?" and "When do you close?" and "Are you open on Saturdays?" and "What time can I come in?" are all the same question.

Mistake 3: Not Defining Boundaries

If you do not explicitly tell the AI what it should not say, it will improvise. And AI improvisation with business-critical information is dangerous. Define boundaries clearly.

Mistake 4: Skipping the Testing Phase

Every hour you spend testing saves ten hours of fixing problems after launch. Test with real scenarios, unusual phrasings, and deliberate edge cases. If your mother-in-law can confuse the AI, your customers will too.

Mistake 5: Never Updating After Launch

The initial knowledge base handles 70% of calls well. After one month of real-world data, it should handle 80%. After three months, 85%+. But only if you actually review the escalation logs and add the missing information. Set a recurring calendar reminder.

The Knowledge Base Checklist

Use this checklist to ensure your knowledge base is complete before going live:

• [ ] Business identity (name, address, hours, contact details)
• [ ] All services listed with descriptions
• [ ] Pricing for every service (or clear "from" prices with caveats)
• [ ] Top 20 FAQs with alternative phrasings
• [ ] Booking process documented (steps, information required)
• [ ] Cancellation and refund policy
• [ ] Payment methods accepted
• [ ] Escalation triggers defined
• [ ] Escalation contacts and routing rules
• [ ] Boundary rules (what the AI must never say)
• [ ] Tone and voice guidelines
• [ ] Upsell and cross-sell offers configured
• [ ] After-hours behaviour defined
• [ ] Emergency protocols documented
• [ ] 50-call test completed with 90%+ pass rate

Key Takeaways

• The knowledge base determines AI quality. A brilliant AI with poor training data gives poor answers. Invest the preparation time.
• Start with your top 20 FAQs — they cover 80% of call volume. Expand from there.
• Document pricing explicitly with device/service, price, inclusions, caveats, and upsell opportunities.
• Define escalation rules clearly — the AI should know exactly when to hand off and how to do so with full context.
• Test with 50 calls before launch, covering FAQs, edge cases, unusual phrasing, and emotional callers.
• Maintain weekly (15 minutes), monthly (1 hour), and quarterly (2 hours) to keep the knowledge base accurate and expanding.
• [Company Cortex](/services/company-cortex) automates the process — ingesting existing documents and keeping all channels synchronised from a single source.

Frequently Asked Questions

How long does it take to build a knowledge base from scratch?

For a typical UK service business, the initial knowledge base takes 2–3 days of focused work. This includes auditing your call data, writing 20–50 FAQs, documenting pricing, and defining processes and escalation rules. With Company Cortex, the process is faster because it ingests your existing documents rather than requiring everything to be written from scratch.

Do I need technical skills to train an AI voice agent?

No. Building a knowledge base is a content task, not a technical task. If you can write a staff training manual, you can build a knowledge base. The technical configuration (connecting the AI to your phone system, CRM, and booking platform) is handled during the setup process. See Setting Up an AI Voice Agent for Your Business for the full walkthrough.

What happens if the AI gives a wrong answer?

If the AI provides incorrect information (wrong price, wrong opening hours), it is almost always because the knowledge base contains incorrect or outdated data. The fix is updating the knowledge base entry — which takes effect immediately. Call transcripts are logged, so you can identify incorrect answers quickly through regular review.

Can the AI learn from calls without manual updates?

Partially. Modern voice AI systems identify patterns in escalated calls and can suggest new FAQ entries for you to approve. They do not autonomously add information to the knowledge base — that would risk introducing errors. The human-in-the-loop approach (AI suggests, human approves) gives you the best of both: continuous improvement with quality control.

How often should I update the knowledge base?

Weekly reviews of the escalation log (15 minutes) are the minimum. Monthly deep reviews (1 hour) catch emerging patterns. Pricing and availability changes should be updated immediately — the AI will quote whatever is in the database, so stale prices cause real problems. With Company Cortex, updates to source documents propagate automatically, reducing manual maintenance significantly.