The employee is still the unit of capacity
In my previous article I argued that decisions about AI in customer service are usually made from inside the organisation, and that the customer’s own view rarely comes up. In this one I want to look at the approach most large service organisations are choosing, because I think it shows that problem most clearly.
The approach is usually called AI-assisted customer service, and it is often sold as a copilot. An employee still handles every request, and AI works beside them. It finds the right guide in the knowledge base, pulls up the customer’s account, suggests a reply, writes the summary afterwards and files the notes. The employee spends less energy on the systems and more on the customer. It is a sensible approach. It is easy to introduce and asks nothing difficult of the organisation, and I understand why so many organisations choose it.
The approach has been measured
It is also one of the few ideas in our field that has been studied properly. A study of 5,172 customer support employees, published in the Quarterly Journal of Economics in 2025, followed what happened when they were given an AI assistant of exactly this kind. They resolved about 15 per cent more issues per hour, which is roughly a seventh more. The newest and least experienced employees gained the most, and I think that is a genuinely good result. Customers wrote to them in a more positive tone, and fewer asked to speak to a manager.
What did not measurably move was the score customers gave. Customer satisfaction, and the share of problems that were solved, were statistically the same as before.
+15%
What the employee gained
15 per cent more requests resolved per hour, and a 9 per cent shorter handling time.
Unchanged
What the customer noticed
No measurable change in customer satisfaction or in the share of problems solved.
Always open
What changes with oHallo
The requests oHallo resolves are taken up as they arrive, under the same procedures every time.
Brynjolfsson, Li and Raymond, Generative AI at Work, Quarterly Journal of Economics, May 2025. 5,172 customer support employees.
I think that is an honest description of what the approach does. The employee became faster and the exchange became friendlier, and the customer still had to wait for an employee to be free.
The answer still waits for an employee to be free
The reason, in my view, is structural. In this approach a request can only be answered when an employee is available to answer it, and almost everything the customer notices about the service follows from that. The opening hours are the hours employees work. The wait on the Monday morning after the bills have gone out is set by how many employees are on shift. The email that arrives on Friday afternoon is answered on Monday, or on Tuesday.
A seventh more capacity changes the size of that constraint a little. At a busy hour it can shorten the queue noticeably, and I do not want to understate that. It is still a queue, though. It still closes at four, and its length still depends on how many employees the organisation decided to have.
AI-assisted customer service, an example
AI-resolved with oHallo
The same requests, each taken up by oHallo as it arrives, always open
There is also the question of where the gain goes. A productivity gain can be spent in two ways. The organisation can serve the same customers with fewer hours of work, or it can serve them sooner. Only the second reaches the customer, and either way the cost of customer service still grows with every employee the organisation adds.
I would change the unit of capacity
When AI resolves the request itself, in the organisation’s own systems, the unit of capacity is no longer an employee. The fiftieth caller at nine in the morning is picked up at the same moment as the first. The request that arrives at eleven at night is handled at eleven at night. The customer who wants a meter reading registered or a payment date moved does not wait for anyone, and the employees are no longer what that customer is waiting for.
This is the difference I care about. AI-assisted customer service makes the organisation’s employees more productive, and that is worth having. AI-resolved customer service removes the wait for a free employee from every request it can resolve, and that is what the customer actually notices.
AI-assisted customer service
Human in the loop
Every request passes through an employee.
AI-resolved customer service
Human on the loop
Only the requests that need a person reach an employee.
The team still needs good tools
None of this makes tools for the team pointless. The employees who take over a request need the whole history in front of them, and they need it at once. At oHallo, when a request is passed to an employee, they see everything said so far, the customer’s account and everything oHallo has already checked, so the customer does not have to start again. That is assistance too. It is aimed at the requests that need a person, which is where I think assistance does the most good.
One question for each approach
If you are choosing between the two, I would put one question to each of them. What happens when a customer gets in touch at nine in the evening? In AI-assisted customer service, the honest answer is that they wait until the morning. With oHallo, the requests that can be resolved are done that evening, and the ones that need a person have a follow-up booked for the morning. I do not think the first answer is good enough any more, and I doubt that your customers do either.
Related
Nobody asked the customer
Many organisations keep an employee on every customer request and give AI the work around it, because customer service is their last personal contact with customers. I think that view underestimates the customer, and I think the customer would say so if anyone asked.
27 Sept 2026Customers want to know a person is within reach
Many customers say they would choose a person even if the outcome and the wait were the same. I take that seriously, and I think it tells us what an organisation has to promise.