Finding knowledge gaps in Amazon Connect

What does Amazon Connect already give you?

Amazon Connect teams usually start with more conversation data than they use. Contact Lens provides the foundation:

• Transcripts for calls and chats, with speaker turns.
• Sentiment across the contact and over its course, so you can see where a conversation turned.
• Rules-based categories that tag contacts containing words, phrases or conditions you define.
• Contact summaries that describe what happened on each contact.

Alongside that, Amazon Connect records the contact itself: queue, handle time, hold, transfers and agent. Together this is a detailed record of every conversation. What it doesn't do on its own is compare those conversations with your knowledge base, which is a separate system with its own content and its own owners.

Why aren't Contact Lens categories enough to find knowledge gaps?

Categories are a rule engine, and a rule engine finds what you tell it to find. That is exactly right for compliance phrases, known complaints and issues you are already tracking. It is the wrong shape for knowledge gaps, for three reasons.

• You need to know the question first. A category for "card declined overseas" requires someone to suspect that question is a problem. Gaps you haven't noticed have no rule.
• Phrases don't capture intent. Customers ask the same thing many ways. "Why did my payment bounce in Bali" and "my card isn't working on holiday" are one question that a phrase list may count twice or miss once.
• A category says nothing about coverage. Knowing 800 contacts mentioned overseas cards doesn't tell you whether an article answered them, or whether two articles answered them differently.

Gap detection approaches it from the other side. It reads each transcript, writes down what the customer wanted in plain language, groups similar intents, and checks each one against the knowledge base. No rule is needed, so new questions surface on their own. Keep your categories for the issues you already watch; they complement each other well.

Doesn't knowledge recommendation solve the problem?

Amazon Connect can surface knowledge content to agents during a contact, retrieved from the knowledge sources you connect to it. That makes good content much faster to reach, and it is worth doing.

Recommendation still depends on the content being there. If no article covers a question, the best match is the nearest article, which may be close enough to look relevant and wrong in the detail that matters. If two articles disagree, the recommendation depends on which one ranks first. Faster retrieval over the same content makes the content easier to reach. It doesn't show what's missing, stale or contradictory. That is the same pattern we describe for search logs in failed searches vs knowledge gaps.

The two work best together. Gap detection fills the knowledge base with answers to what customers actually ask, and recommendation puts those answers in front of agents at the moment they're needed.

How does gap detection connect to Amazon Connect?

Knowledge Quest connects to your Amazon Connect instance using its instance ID, region and credentials created for Knowledge Quest alone. You enter them under Service connections, and the onboarding team completes and verifies the setup with you before analysis starts. Credentials are write-only and held in Azure Key Vault.

From there, Knowledge Quest ingests transcripts and contact data, and you map your queues to knowledge areas, so each team's contacts are checked against the right set of articles. Your flows, routing and agent workspace stay as they are. Nothing is added to the contact itself.

How does an Amazon Connect transcript become a ranked gap?

Contact Lens has already done the transcription, so the work starts from text. Each transcript is summarised and marked resolved, partly resolved or unresolved, with a note of whether a callback was needed. It gets a satisfaction score and is placed in your knowledge taxonomy, so volume and outcomes show by topic rather than by queue. Then every question the customer raised is pulled out as a separate intent: a caller who rang about a declined card and then asked about travel notifications produces two, each checked against your articles on its own.

For each intent, the Knowledge Gap Evaluator retrieves candidate articles, compares them side by side, and decides whether one answers the question. If none does, the intent becomes a gap, recorded with the candidates that were considered and a confidence score. A reviewer can open any gap and read the transcripts behind it.

Gaps are then ranked by impact: volume, weighted by how difficult the conversations were. Volume matters, but friction matters more: a question that produces transfers, long holds, callbacks and unresolved outcomes does far more damage per contact than one agents answer cleanly. Handle time is split into talk, hold, after-call work and connect, so you can see where the extra minutes go. The knowledge gap analysis guide walks through the ranking step by step, and the cost of a knowledge gap shows how to turn it into dollars with your own cost per contact.

How do you catch new questions early?

The most valuable gaps are often the newest. A product launch, a price change, a system outage or a new regulation creates a question your knowledge base has never seen. By the time it is big enough to justify a category rule, agents have already been improvising answers for weeks.

Because gap detection doesn't depend on rules, a new question appears as soon as customers start asking it. Insight alerts watch for the following; emerging topics and drift detection explains the baselines behind them.

• Emerging topics: intents that barely existed last month and are growing quickly.
• Knowledge-gap surges: a sudden rise in questions that no article answers.
• Volume, sentiment and callback spikes on a topic or queue, with your public holidays and quiet periods taken into account.

Daily, weekly and monthly briefs summarise what changed, with every number checked against your contact data before the brief is sent. Kai, the grounded analyst inside Knowledge Quest, answers follow-up questions about a spike and cites the contacts behind each answer.

Where do the fixes go?

Into the knowledge base your agents already use. Amazon Connect teams run a variety of knowledge tools, so Knowledge Quest works with the one you have:

• Live sync with supported knowledge bases, including Genesys Knowledge, Freshservice and Zoho Desk, so the articles you have and the articles you approve stay aligned.
• Import and export for any other tool. Bring existing articles in from PDF, Word, HTML or a URL so gaps are matched against your real content. Send approved articles out as PDF, Word, HTML or structured JSON.

Whichever route you use, the drafting step is the same. Knowledge Copilot writes the new article or update from the transcripts that exposed the gap, so it answers the real question in your customers' words, and links each statement back to its source. Six quality checks run before review: brand alignment, readability, validity, duplicates, contradictions and complexity. Your subject-matter experts approve every article before it reaches an agent or a bot.

Then measure the result. Watch transfers, callbacks and repeat contacts on the topic after the article goes live. If they fall, the gap is closed. If they don't, the evidence shows what the article still misses.

Amazon Connect keeps doing what it does well: routing, recording and transcribing every contact. Your knowledge base keeps serving answers. Gap detection connects the two so your knowledge reflects what customers ask. See how Knowledge Quest finds the gaps.