AI customer service with source attribution

AI customer service becomes verifiable only when every proposed answer can be traced to approved knowledge or current operational data. Source attribution does not automatically prevent every incorrect answer, but it does show what the system relies on, how current the information is and where an employee should be cautious.

This matters because customer enquiries often combine two kinds of knowledge. A customer may want to know the returns policy and whether their particular return has already been received. The first answer comes from managed policy information; the second from a current order system. An AI model must not confuse those sources or fill in missing facts itself.

Source attribution starts before the answer

Placing a source beneath generated text is the final step, not the foundation. First, the organisation needs to decide which information the system may use at all. Distinguish three layers:

1. Approved customer knowledge: product information, terms, procedures and explanations that may be shared with customers.
2. Current operational facts: order status, payment, stock, planning or file status from the authoritative system.
3. Internal decision information: exception rules, permissions and commercial considerations an employee may need but which must not appear in the customer answer.

This separation prevents internal instructions from accidentally being paraphrased as customer-facing text. Yindle Desk can place a proposed answer beside the relevant service context so an employee assesses the source and outcome together.

Use different sources for different questions

Not every fact belongs in a knowledge base. Prices, availability and order status change too quickly and should be retrieved from the current source wherever possible. Explanations of a process or product characteristic may be better held in managed content.

Type of question	Preferred source	Main check
How do returns work?	Approved policy information	Version and applicable market
What does this product cost now?	Online store or product source	Variant and freshness
Has my order been dispatched?	Order and logistics context	Identity, status source and timestamp
Is an exception permitted?	Internal procedure plus human judgement	Authority and customer impact
What was agreed previously?	Authorised customer timeline	Correct customer and relevant context

This model prevents two familiar errors: manually copying changing facts into documents and giving sensitive internal information to the answer generator.

How to produce a verifiable answer

1. Identify the question. Determine the subject, language and whether current customer or order data is needed.
2. Select suitable knowledge. Search only approved content for the correct market and category.
3. Retrieve live facts selectively. Request only the minimum operational data relevant to this customer and question.
4. Create a draft. Have the model preserve the distinction between fact, explanation and proposed next step.
5. Show the source. Give the employee the source, title, version or system status alongside the relevant passage.
6. Check for contradictions. Block or flag an answer when sources contradict each other or essential information is missing.
7. Record the review. Retain which sources were used and which corrections the employee made.

This approach aligns with the principle of the NIST AI Risk Management Framework: AI risks are not checked off once, but continually governed, measured and managed throughout the lifecycle.

What should a source reference show?

A bare filename is rarely enough. The employee needs to assess quickly whether the source is usable. Where possible, show:

• the title or a recognisable description of the source;
• the relevant passage, rather than the full document;
• the applicable market, product group or customer category;
• the version or date when the information was updated;
• for live data: the source system and retrieval time;
• a warning when several sources do not match.

Customers do not always need to see this technical detail. A service employee does need it to check the answer responsibly. For direct self-service, a simple reference such as ‘according to the current order status’ or a link to the applicable terms may be appropriate.

A citation does not prove that a conclusion is correct

A model can misinterpret a reliable source. It may combine two valid passages that concern different situations or apply a general policy to an exception. Checking therefore needs to go beyond ‘a source is included’.

Test at least these failure patterns:

• the source mentions the subject but does not answer the question;
• the passage applies to another market or product variant;
• the operational status is older than a recent customer message;
• two sources give different instructions;
• the answer adds a timeframe, price or commitment that appears nowhere in the sources;
• an internal procedure is presented as external policy.

Use test questions with incomplete, ambiguous and contradictory information. A good system does not then confidently state what is probably true; it identifies what is missing or asks an employee to review it.

Design a clear route for uncertainty

When no suitable source is found, several safe outcomes are possible:

1. ask the customer a focused follow-up question;
2. create a draft stating which check is still needed;
3. assign the case to a subject specialist;
4. block the answer when the risk of an incorrect commitment is high;
5. record the missing subject as a knowledge gap.

The choice depends on the consequences. A missing detail in a general product question calls for a different response from uncertainty about cancellation, payment or warranty.

Manage the sources as a product

An AI knowledge base is not an archive that is finished after the initial import. Assign a content owner for each subject, record which source takes precedence and remove duplicate wording. Take changing facts out of documents when a current system can provide them more reliably.

Use employees’ corrections as a signal. If drafts in one category are frequently changed, the cause may be missing knowledge, poor discoverability, outdated content or an unclear procedure. Also read how to automate customer service in a controlled way and retain human judgement at the right moments.

Checklist before launching an AI answer feature

• Has every source suitable for customers been approved for its content?
• Are internal and customer-facing knowledge technically separated?
• Do prices, stock and statuses come from current systems?
• Can the employee see the passage used and how current it is?
• Are contradictory or missing sources clearly flagged?
• Is it recorded which answers always require human review?
• Are corrections fed back into the knowledge and testing cycle?

Make reliability visible

Source attribution turns AI from an opaque text generator into an assistant that is easier to check. Real quality comes from the interaction of approved knowledge, current process data, clear routes for uncertainty and an employee who can intervene in high-risk questions.

Want to investigate which sources are suitable for AI-assisted service? Discuss a scoped Desk application with Yindle and start with one question category and its actual information sources.