Eighty-four per cent of businesses in Asia Pacific believe their AI agents recognise returning customers and their history. Seven in ten consumers in the same region say they routinely have to start a new AI conversation from scratch.
Both numbers come from the same survey, published by Twilio in September 2026. One side is reporting what its systems tell it. The other side is reporting what happened to them.
That gap is the most useful thing a manager can take from the research, and it has very little to do with artificial intelligence.
The failure is quiet, which is the problem
Software failures are usually loud. The page does not load, the invoice does not send, the phone rings.
This one is silent on the inside. The system logs a completed conversation. The dashboard counts a resolved query. Nothing anywhere records that the customer had to explain who they were for the third time, gave up, and went to a competitor.
Twilio’s figures put some shape on the cost. Seventy per cent of consumers in the region abandoned an AI service interaction because it lacked basic information about them or their previous contact. Five per cent ended their relationship with the brand entirely after an exchange like that.
Five per cent is a small number attached to a large one. Whatever your customer base is worth, one in twenty of the people who hit that wall does not come back.
The counterweight belongs here too, because the picture is not all bad. Sixty-eight per cent of the same consumers say AI bots improved over the past year. The tools are getting better. The measurement is not.
A second gap of exactly the same shape
The survey found one more pair of numbers worth keeping.
Seventy per cent of consumers say AI agents should identify themselves as AI. Twenty-two per cent of businesses say theirs do.
Again: what the customer wants, and what the business believes it is providing, measured separately and pointing in different directions. And again, nothing inside the company would ever surface it. No dashboard has a field for “the customer was not sure who they were talking to”.
Why the two sides disagree
It is tempting to read the 84% as businesses kidding themselves. That is not quite what is happening, and the real explanation is more useful.
The system genuinely does recognise the returning customer, in the sense that the record exists somewhere. What it often cannot do is carry that record into the conversation the customer is having right now, on the channel they chose, after the last exchange ended. The capability is present in the product and absent at the point of contact.
So the business answers honestly. It bought the thing that remembers customers, the supplier demonstrated it remembering customers, and nothing since has told anyone otherwise. Meanwhile the customer meets a system that opens every conversation as though it has never heard of them.
Both answers are true. They are answers to different questions, and only one of the two questions is being asked inside the company.
What a manager can do about it, before buying anything
The practical question is not which tool to choose. It is how you will find out what the customer experienced, rather than what the system reported.
That is an old management instrument pointed at a new place. Contact centres have always known they cannot read everything: traditional quality assurance reviews something like two to five per cent of interactions, which is a sample, not a view.
Michael Cowell of the outsourcer TDCX offers a sharper test for whether AI has actually been built into a process or bolted onto one. His question: “if you switched the AI off tomorrow, what would break?” In an operation that has simply added tools, things get slower. In one designed around the AI, the process stops.
Ask it about your own customer journey. If switching the AI off would change nothing the customer notices, you have not automated the work, you have added a layer. If it would stop the process, you should know exactly who is watching that process and what tells them it is going wrong.
Three checks that cost nothing
Be the customer on Monday morning. Go through your own AI-touched path as a returning customer would. Not a test account with perfect records. A real one, with a history and a half-finished order. Most teams have never done this, and it takes twenty minutes.
Read the handovers, not the resolutions. Sixty-five per cent of consumers in the survey had to repeat themselves after being passed from an AI agent to a person. The handover is where the joins show, and it is almost never where the reporting looks. Pull ten transcripts from the point of transfer and read what the person had to ask again.
Decide who owns the outside view. Name a person whose job is knowing what customers experienced this month, with time set aside for it. If nobody owns it, the dashboard wins by default, because it is the only account of events anyone is reading.
Being straight about the source
Twilio sells customer-communication and contact-centre software, including the products that give AI agents exactly the customer memory this research says they lack. It has a commercial interest in the finding. Name it when you use the numbers, and treat its recommendation, which is to unify your customer data on a platform like Twilio’s, as an interested party’s advice rather than a neutral conclusion.
The sample is stated, which is why the research is worth citing at all: 7,652 consumers and 660 business leaders across 18 countries. All the figures quoted here are the Asia Pacific responses, not global ones. The fieldwork agency and dates are not published, so the figures deserve a “reportedly” rather than a “research shows”.
None of that changes the shape of the finding. Two groups were asked about the same thing and gave different answers, and only one of those groups has the experience first-hand.
The belief inside the business is not evidence. It is a hypothesis that nobody has tested, and the people who could settle it are the ones quietly leaving.