An agent is helping a customer whose replacement device has failed. The knowledge tool returns a warranty article. The customer has already received one replacement, and the standard instructions do not explain whether another is allowed.

The article is relevant. The next step is still unclear.

Useful agent assistance needs to close that gap. It should help the employee understand which guidance applies, what remains uncertain and what action is available within their authority.

Put the recommendation in the customer's context

The agent needs the product, purchase date and previous replacement history. A recommendation built without those facts can sound helpful while directing the employee down the wrong path.

Show the policy source and its effective date. Distinguish information retrieved from the account from conclusions the tool has drawn. If an exception requires approval, make that boundary visible before the agent promises a replacement.

Presentation matters as much as retrieval. An agent handling a frustrated customer cannot carefully read several competing summaries at once. Put the immediate decision first, with supporting information available when needed.

Then connect that decision to the work: prepare the request, carry forward verified details and show what must be checked before submission.

Expect different agents to benefit differently

A study published in 2025 in the Quarterly Journal of Economics examined AI assistance for 5,172 customer-support agents at one software company during a rollout mainly in 2020 and 2021. It found a 15% average increase in issues resolved per hour, with the largest benefits among less-experienced and lower-skilled workers. Experienced, highly skilled workers saw much smaller benefits, with small declines in conversation quality among the most skilled. Brynjolfsson, Li and Raymond

A single firm's result is a reason to test the opportunity in the operation where the tool will be used. A new starter may need help recognizing the exception. An experienced colleague may already know the answer and find an intrusive suggestion slows the conversation.

Test with both groups. Give agents an easy way to reject incorrect guidance and explain why. An override can be useful evidence about an outdated policy, missing account detail or poorly timed suggestion.

Remove work without removing the check

For the failed replacement, assistance could assemble the relevant history and prepare an escalation. The employee still checks whether the request matches the facts and explains the next step to the customer.

That final review takes time. It belongs in the workflow, especially when a wrong action would be difficult to reverse.

Assess whether the tool reduces searching, repeated entry and avoidable holds while protecting the quality of the answer. Review completed requests as well as how often agents use the feature. Frequent use can reflect dependence or mandatory process, not usefulness.

The best evidence comes from the conversation itself: the agent understands the situation sooner and can move the customer's request forward with confidence.