Reducing Repeat Contacts with Better Knowledge Bases

Repeat contacts — when a customer comes back about the same issue — are often a knowledge base problem in disguise. The AI agent didn't have the answer the first time, so it gave a partial response, escalated, or punted. The customer comes back. Investing in knowledge base quality is the single highest-leverage move for reducing this churn.

TL;DR

• Repeat-contact rate is a leading indicator of agent quality.
• Most repeats trace to knowledge gaps, not AI bugs.
• A small clean knowledge base outperforms a large messy one.
• The KB needs continuous tending — it's not a "set it and forget it" artifact.

What counts as a repeat contact

Definition: the same customer (matched on phone or account) contacts you again about the same issue within a defined window (typically 7-14 days).

If they come back about a different issue, it's a new contact. If they come back about the same issue, you didn't fully resolve the first one.

Why repeats happen

Common causes:

1. Knowledge gap. AI didn't know the answer; gave a partial / wrong / vague response.

2. Resolution gap. AI thought it solved the issue; actually didn't.

3. Promise gap. AI said "we'll get back to you"; nobody did.

4. System gap. AI made the change but it didn't propagate to the customer-visible system.

5. Confusion gap. AI explained correctly but the customer didn't understand.

The first is by far the most common. Knowledge base quality fixes the most repeat contacts per dollar spent.

What a good knowledge base looks like

For an AI to use it effectively:

Clean prose. No navigation chrome, no "related articles" sidebars, no headers from a CMS template. Just the content.

Specific. "Our return window is 30 days from delivery" beats "We accept returns within a reasonable timeframe."

Up to date. Outdated answers are worse than no answers.

Cross-referenced. Internal links connect related topics so retrieval can find adjacent context.

Chunked appropriately. Articles broken into 200-500 token chunks for embedding.

Tagged with metadata. Category, last-updated date, source authority.

Common KB problems

Stale content. Policy changed in March; KB still has the old version. Agent confidently gives wrong info.

Conflicting articles. Two articles with different answers to the same question. Agent picks one.

Vague answers. "It depends" doesn't help an AI.

Buried specifics. The answer is in paragraph 8 of a 12-paragraph article.

Missing topics. The 50 most-asked questions don't all have articles.

How to find the gaps

Audit calls where the AI failed:

1. Pull 50 calls where the AI escalated due to "I don't know."
2. For each, identify what info would have answered the customer.
3. Check if that info is in the KB.
4. If not: write an article. If yes: figure out why retrieval missed it.

This audit is the single highest-leverage hour you can spend on a voice agent monthly.

Writing for AI consumption

Subtle differences from writing for humans:

Front-load the answer. AI retrieves chunks; the relevant chunk should contain the answer, not lead up to it.

Include the question phrasing. "How long is the return window? Our return window is 30 days from delivery." The question phrasing helps retrieval match.

Avoid jargon without definition. AI doesn't infer.

Use specific numbers, not ranges where possible. "$45 fee" beats "around $40-$50."

State exceptions clearly. "30-day return window EXCEPT for personalized items, which are non-returnable." Don't bury exceptions.

The chunking decision

How you chunk affects retrieval quality:

By heading. Each H2 section becomes a chunk. Natural boundaries; sometimes too long.

By paragraph. Smaller chunks; more precise but can lose context.

By topic. Each chunk covers one topic, regardless of source structure.

Sliding window. Overlapping chunks (e.g., 400 tokens with 100-token overlap). Catches answers that span boundaries.

For voice agents, paragraph or sliding window typically works best.

For more on the retrieval pattern, see retrieval-augmented generation for voice agents.

KB maintenance cadence

Three habits:

Weekly: Review last week's escalated calls. Patch KB gaps.

Monthly: Audit popular articles for accuracy. Update if anything changed.

Quarterly: Full KB review. Retire obsolete articles, merge duplicates.

Total effort: 4-6 hours/month for a typical mid-sized KB.

Tracking KB performance

Useful metrics:

• KB hit rate. Of AI calls that needed an answer, what percentage found it in the KB?
• KB confidence. Distribution of retrieval similarity scores.
• Articles that triggered escalations. Articles that retrieved but didn't actually answer.
• Articles with low usage. Possibly outdated / unfindable.

These metrics guide the maintenance work.

When KB isn't the right tool

Some questions shouldn't go through KB at all:

Real-time data. Order status, account balance, today's availability. Use function calls.

Customer-specific. Anything per-account. KB is for static content.

Action-oriented. "Cancel my subscription." Function call, not KB lookup.

If you find yourself adding "FAQ" articles for these, you have a function-calling gap.

Translating KB for multilingual

If you support multiple languages:

• Translate the KB; don't rely on the LLM to translate at query time.
• Keep article structure parallel across languages.
• Chunk and embed each language separately.
• Test retrieval per language.

Native-language KB significantly improves multilingual agent quality.

Related reading

• The Definitive Guide to AI Customer Support in 2026
• Building a Tier-1 AI Support Agent Step by Step
• Why "Human-in-the-Loop" Beats "Fully Autonomous" for Most Teams
• How to Calculate ROI for AI Customer Support
• How AI Support Agents Should Handle Account Verification

FAQ

How big should my KB be? Start with 50-100 articles covering top intents. Expand based on what actually gets asked.

Can I auto-generate KB articles from call transcripts? Risky — generated content can hallucinate. Use as drafts; humans verify.

What about a separate KB for the AI vs for human-facing help center? Some teams do this. The AI-facing KB is more terse and structured. Trade-off: more maintenance.

Should the KB include policies? Yes — and tag them as authoritative so the agent prioritizes them.

How do I prevent the KB from going stale? Tie updates to product / policy changes. If a policy changes, the change-management process should include KB update.