An AI agent can answer a customer call without making your service better. The test is whether it resolves the request, recognizes when it cannot, and hands the rest to a person who knows what has already happened. That is the business decision behind OpenAI’s September 23 account of Ringg, a company building voice and chat agents for customer operations. Ringg says its agents resolve up to 65% of routine inquiries without human involvement. That is a useful result. It is not a reason to design the remaining work as an afterthought.
If an agent takes the easy calls and leaves the hard ones to people, the human job changes. It becomes less about repeating answers and more about repairing exceptions, making judgments, and keeping a promise the automated conversation could not finish.
TL;DR
A call handled without a person is not the same thing as a customer problem solved. Measure completed requests, repeat contacts, and the quality of handoffs. Give agents a narrow job they can finish reliably; give people the authority and context to finish the rest. The handoff is part of the service, not evidence that the AI failed.
The percentage has a denominator
Ringg reports that, across its customers, agents resolve up to 65% of routine inquiries without a person. OpenAI’s account also says that Policybazaar, an Indian insurance platform using Ringg, connected more than 57,000 customer requests and handled 67% of calls without human intervention. Those are two different descriptions of two different populations. Neither figure tells us that 65% or 67% of every company’s customer issues can safely be automated.
“Up to” is doing work in the first claim. So is “routine.” A status request, a scheduling change, and a dispute about coverage are not interchangeable calls. The first may have an answer in a record. The last may require judgment, an explanation of terms, or a person with authority to make an exception.
A leader looking at this news should ask for the boundaries behind the number. Which calls counted as routine? What counted as handled: an answered call, a completed action, or a customer who did not need to call again? How were transfers counted? OpenAI’s story describes the outcomes, but does not establish those definitions for your business. The right lesson is to set them before buying a resolution claim.
Design the transfer before the greeting
Imagine a customer calling about a payment that appears twice. An agent can verify the account, locate the transaction, and explain what the system shows. If the account data is inconsistent, it should not improvise a refund policy or trap the caller in another explanation. It should send the case to a person with the account identifier, the disputed transaction, what was checked, and the question still open.
That example is a proposed workflow, not a reported Ringg deployment. It illustrates a common design choice: where the agent stops. If the person who receives the call has to ask the customer to repeat everything, the organization has moved work rather than removed it.
Ringg’s platform, according to OpenAI’s case study, can divide work among agents for qualification, support, verification, scheduling, and escalation while keeping a conversation consistent across voice, chat, WhatsApp, and the web. The interesting part for a manager is not the number of agents. It is whether the customer can cross those boundaries without losing the thread.
That takes an agreement between the service team and whoever owns the system. What information should follow a transfer? When does a person take over automatically? Can the person correct an agent’s account of what happened? Is there a clear owner when the agent cannot complete an action but marks the case closed?
Agents become useful collaborators when they leave their human colleagues with a better starting point, not a mystery to clean up.
Give the human team a different scorecard
A vendor can show a high percentage of calls handled without people while your staff still spends its day on repeat contacts and confused escalations. Do not let a single automation figure become the service scorecard.
I would track completed customer requests separately from calls ended without a transfer. I would also look at how often customers contact the team again about the same issue and whether a person receiving a transfer can pick up without asking for the story again. Keep a sample of transferred conversations for review. The numbers tell you where to look; the conversations tell you what the numbers missed.
Then ask the staff what changed. Did the agent remove repetitive work, or did it route a steady stream of cases that now require more time to untangle? Are people empowered to fix the issue, or merely to apologize for a system they cannot override? Those answers determine whether you have improved the operation or just moved the queue.
None of this requires a grand redesign on day one. Pick one category of request with a clear completion condition. Write down the conditions that force a handoff. Give the receiving person the conversation history and authority to act. Review a small batch of finished and transferred cases together each week. Only then widen the job.
The promise is a completed request
Ringg’s result matters because it points to real work getting done in customer channels, not merely a bot greeting a caller. But the part worth copying is not a headline percentage. It is the discipline of designing the whole journey, including the point where the agent needs a colleague.
If your customer has to start over with a person, the AI did not really hand anything off. It just answered first. The practical standard is simpler: the customer reaches an answer, and the team knows who owns the next move.
Research and structure: Mai. Direction and voice: John Lipe.