Klyro
Back to blog
November 17, 2026

Reduce AI transfers without locking up the customer

Explains the reasons why AI transfers too much (content gaps, poorly calibrated limits, out-of-scope question, lack of customer trust), guides sample analysis of transfers by category, review of rules of
A dedicated athlete competes in a marathon using a racing wheelchair on city streets.
Photo by Yusuf Miah on Pexels

Why AI transfers conversations too much and how to fix it

When an AI agent transfers half of the conversations to human attendants, three problems appear at the same time: the volume doesn't drop, the resolution time increases, and the customer has to repeat what they've already explained. The promise of deflection becomes operational frustration.

Excessive transfer is generally not a technology failure. It is a symptom of an inadequate configuration: a knowledge base with gaps, generically calibrated action limits, transshipment rules that do not reflect the actual operation, or answers that passed the test but failed in practice.

Reducing transfers without locking up the customer requires reviewing what the AI is receiving as input, what it can respond autonomously, and when it should pass the conversation - with full context - to a human. This guide presents the method for identifying transfer causes, adjusting limits, and measuring the actual impact on the operation.

Four causes of frequent transference

Knowledge Base Gaps

AI only answers what is documented. When the knowledge base doesn't cover all the variations of a recurring question, the transfer takes place even on simple demands.

Example: The customer asks “how long does it take to process chargebacks?” and the database only documents “refund period is 5 to 7 business days.” AI doesn't recognize “process” and “chargeback” as synonymous with “refund” and transfers.

Review transferred conversations from the last two weeks and group by reason. Questions that appear five or more times and did not generate a satisfactory response indicate a content gap. Update the base before adjusting transfer rules.

Rigidly calibrated confidence limits

AI agents generally use a trust limit to decide whether to respond or transfer. When that limit is too high, the AI transfers even in questions that it could answer.

This adjustment needs analysis: review a sample of 50 transfers marked as “low confidence” and validate that the answer was available in the database. If it was, the limit can be adjusted to allow for a response in similar cases.

The balance is between autonomy and precision. Lowering the limit increases deflection but may generate inaccurate answers; increasing reduces error but increases unnecessary transfers.

Question outside the defined scope

When the customer requests something that is not within the scope of AI's action - contract cancellation, change of registration data, analysis of a specific case - the transfer is correct. The problem arises when the scope has been defined in a generic way and the AI transfers before trying to respond to the part that is within its reach.

Example: the customer says “I want to cancel and also know if the refund has already been processed.” The IA can respond about a refund and then transfer only the cancellation request. If you transfer everything right from the start, you waste the opportunity for partial deflection.

Configure the agent to resolve what's in scope and transfer only the part that requires human intervention. This reduces transfer volume and provides useful context to the attendant.

Customer explicitly asks for human service

This is the most legitimate cause of transfer and should not be blocked. When the customer writes “I want to speak with an agent” or shows frustration with automation, forcing continuity with AI impairs the experience.

The mistake lies in not capturing that signal before the conversation is extended. Configure triggers to detect transfer requests in the first or second message and pass on the conversation with preserved context.

How to analyze transfers by category

Separating transfers by reason allows you to identify where to adjust first. Group transferred conversations into four categories:

**Transfer due to content gap: ** the AI did not find an answer in the database. Solution: Update documentation.

**Transfer due to low trust: ** the response exists but the agent has not reached the trust limit. Solution: Review calibration or rephrase content for clarity.

**Transfer by scope: ** the customer requested something that requires human action (example: cancellation, data change, case analysis). Solution: adjust flow to resolve the automatable part before transferring.

**Transfer requested by the customer: ** explicit request for human service. Solution: Ensure that the transfer happens quickly and with full context.

Build a weekly spreadsheet with these four categories. The distribution between them reveals where to invest effort. If 60% of the downloads are due to a content gap, prioritize updating the database. If 40% is due to low confidence, review calibration.

Review transshipment rules without locking up the customer

Transshipment rules define when the AI must pass on the conversation to a human. Adjusting these rules requires a balance between autonomy and security.

Set limits by type of demand. Frequently asked questions with a documented answer must have full autonomy. Requests involving judgment, exceptions, or critical decisions must be transferred immediately.

Set up attempts before transfer. In ambiguous questions, allow the AI to clarify the question before transferring. Example: customer asks “how much does it cost?”. The AI can answer “do you want to know the price of which plan?” instead of transferring direct.

Preserve context in the transfer. The attendant needs to see the full history: what the customer asked, what the AI answered, where the conversation stopped. This avoids repetition and reduces resolution time.

Test the flow before releasing. Simulate real conversations in an approval environment. Validate that the AI answers correctly in typical cases and transfers with context in complex cases.

Platforms that offer [chatbots with automation and configurable handoff] (URL0) allow you to adjust transshipment rules without depending on development. Set limits by demand category, test with controlled sample, and review weekly.

Update knowledge base after analysis

Content gaps appear in transferred conversations. Reviewing the weekly sample of transfers reveals questions that the AI should have answered but failed to answer.

Update the knowledge base by following this script:

Identify language variations. Customers ask the same question in different ways. Document synonyms, slang, and regional variations.

Rephrase ambiguous answers. If the AI transfers because the answer in the base is too generic or technical, rewrite clearly and objectively.

Cover complete use cases. Document not only the answer but the context in which it applies. Example: “Refund period is 5 to 7 business days after approval of the chargeback by the bank.”

Validate before publishing. Test the new version of the content with real customer questions. Confirm that the AI finds the answer and that it solves the question.

The update process must be continuous. Take weekly time to review transfers, identify gaps, and update content. Operations that do this in a structured way reduce transfers by 30 to 50% in the first three months.

For detailed guidance on structure and updating, see the guide on [knowledge base for AI] (URL1).

Test response before releasing to customers

Publishing new content without validation generates wrong responses in production. Test in a controlled environment before releasing.

Build test scenarios List real customer questions that the update should cover. Include language variations and different contexts.

Validate the generated response Confirm that the AI finds the correct content, that the answer is complete, and that the tone is appropriate.

Simulate long conversations. Test whether AI maintains context in conversations with multiple questions. Validate that she can resume the subject without losing her thread.

Review transfers after publication. Even after testing, follow real conversations for the first 48 hours. Identify false positives (AI answered but customer asked again) and adjust quickly.

Platforms that allow training AI agents with company documentation provide a testing environment before releasing for production. [BRANDAI can be trained with knowledge base] (URL2) and tested in restricted mode before activating for all customers.

Indicators to monitor deflection quality

Throughput alone doesn't reveal whether the AI is working well. Follow cross indicators:

**Transfer fee by type of demand: ** separate by category (product question, technical support, cancellation request). Compare month by month. Increase in a specific category indicates a content gap or inadequate calibration.

**Post-transfer satisfaction: ** measure CSAT of conversations that went through transfer. Low satisfaction indicates that the customer repeated information or that the AI did not provide useful context to the attendant.

**Resolution time after handoff: ** compare average resolution time in conversations that started directly with a human versus conversations that were transferred. If the time is longer in those transferred, the AI may be wasting attempts or providing incomplete context.

**Recontact for the same reason: ** a customer who reopens a conversation on the same topic within 24 hours indicates that the previous response did not resolve. Track recontact by category and adjust content.

**Volume of messages per conversation: ** transferred conversations that span many messages indicate a loss of context or lack of clarity in the passage to human.

Build a weekly dashboard with these five indicators. Their combination reveals where to adjust: content, calibration, transshipment rules, or transfer flow.

When to increase battery life and when to transfer faster

Reducing transfers doesn't mean eliminating them all. The purpose is to transfer only when necessary, with full context and at the right time.

**Increase autonomy when: ** the question is recurring, the answer is documented, there is no risk of critical error, and the customer shows no frustration. Configure the AI to respond directly.

**Transfer faster when: ** the customer explicitly asks for an attendant, the conversation involves judgment or exception, there is a risk of financial or reputational impact, or the customer shows irritation (repeated short messages, use of caps lock, negative emojis).

**Use a hybrid approach when: ** AI can solve part of the demand. Example: answer a question about delivery time and then transfer a request for an address change. This reduces transfer volume and delivers value to the customer.

Operations that balance autonomy with fast transfer maintain high deflection without affecting satisfaction. The criterion is: if AI can solve safely, it solves. If in doubt, transfer with context.

Frequently Asked Questions

**What is an acceptable transfer fee? ** There is no universal rate. It depends on the scope defined for the AI and the complexity of the demands. The most useful thing is to follow the evolution month by month: if transfers fall while satisfaction and resolution rise, the adjustment is on the right track.

**How do you know if the AI is downloading due to a configuration error? ** Review sample transfers marked as “low confidence” or “question not found.” If the answer was in the base but the AI didn't find it, there is a problem with calibration or content structure.

**Should I block human service requests to force deflection? ** Never. Blocking human access when the customer asks creates frustration and may violate service regulations. Set up triggers to transfer quickly when the customer requests.

**How long does it take to reduce transfers? ** It depends on the maturity of the knowledge base. Structured documentation operations reduce transfers by 30 to 50% within three months. Operations that start from scratch can take six months to achieve stable deflection.

**How to prevent the service team from losing space due to the deflection? ** Show that the role changed, not disappeared. Repetitive questions leave the human queue and the team begins to concentrate complex cases, exceptions, and clients at risk. Track average service time, satisfaction per agent, and volume of cases resolved in the first interaction.

**Can AI learn on its own with transfers? ** Not automatically. Transfer analysis requires human review to identify content gaps, adjust calibration, and validate that the new response is correct. Uncurated automation generates inaccurate answers.

References

  • [When the AI agent must transfer to a human] (URL3)
  • [Deflection in service: how to measure the real value of AI] (URL4)
  • [Chatbot with AI or chatbot with rules: which to choose for service?] (URL5)

Share

Open Instagram and paste the link in your post/story.