AI is changing the work behind the headset. Whether it produces better service or simply a cheaper one will depend on what employers do with the time it saves.
A man paid off the debt on his late wife’s account. Seven months later, a debt purchaser began chasing him for money he no longer owed. The original creditor eventually apologised and sent £30 in compensation.
The apology and cheque were addressed to his wife.
The Financial Conduct Authority recorded this case in its 2025 review of support for customers in vulnerable circumstances. It did not identify AI as the cause. The company had managed to turn a settled debt into another problem for a bereaved customer. FCA review.
This is the service AI is being asked to improve. There is plenty of room to do so. Nobody should have to spend a lunch break chasing a refund. On the other side of the conversation, an adviser may have spent all afternoon explaining the same broken process. A tool that finds the right information and gets both people on with their day has an obvious purpose.
The difficulty begins when answering a message is treated as the whole job. Customer service is also where a company discovers what its promises cost. A parcel has gone missing. An account has been frozen. Someone has died, and the computer is still sending bills. At that point, the person making contact needs the organisation to do something.
AI will change who answers, how much they know and how quickly they can work. It will also expose a choice that predates the technology: how much help is a business prepared to give?
A better first day
For someone new to the job, a customer’s question can set off a small, private scramble: find the policy, check the account, work out what happened, compose a reply. All while the customer waits. Being pleasant is only one part of it. Knowing where to look can take months.
One of the strongest studies of workplace AI offers some encouragement here. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the introduction of an assistant among 5,172 customer-support agents, mostly based in the Philippines and serving a US software company’s customers. Issues resolved per hour rose by about 15% on average, with larger benefits for less experienced and lower-performing workers. The rollout mainly took place in 2020–21; the peer-reviewed paper appeared in 2025. This was assistance for people, who could reject suggestions, rather than a test of replacing them. Generative AI at Work.
There is something hopeful in that result. Experience need not remain locked in the head of the colleague everyone asks for help. If more of it is available when a customer needs it, a new recruit has a chance to feel useful sooner.
For a trainee, the tool is most useful alongside a colleague who can explain why an answer fits. They need to learn what an exception looks like and when to stop following the usual steps. Otherwise, they may become fluent in answers they do not understand.
For experienced advisers, the question is different. They may already know the answer. An extra suggestion can become another interruption.
A 2026 working paper on Alibaba’s after-sales service makes that concern concrete. In a four-week experiment in 2024 involving 5,940 agents with less than a year’s tenure, AI assistance improved speed and customer ratings overall, without a significant improvement in repeat-contact rates. Lower-performing agents gained most; top performers experienced worse quality outcomes. It is a working paper about one platform and a particular tool, rather than a verdict on every assistant. It does argue for testing separately with different groups of staff. Ni and colleagues.
When another apology won’t do
AI can help an adviser sound more considerate. That possibility deserves a fair hearing. People do not always find the right words under pressure, especially when they are tired, working in a second language or dealing with the fifth angry customer in succession.
Shunyuan Zhang and Das Narayandas tested AI-generated suggestions at a meal-delivery company. Their randomised field experiment found faster responses and improved customer sentiment, particularly with less experienced agents. But the benefits varied. Subscription-cancellation conversations improved; repeat complaints were harder to help because the underlying problems lay beyond what the suggested replies could fix. The study was published online in Management Science in 2025. Zhang and Narayandas.
Imagine a customer who cannot find a cancellation option: a clear explanation may settle the matter. Now imagine someone whose delivery has failed for the third time. A beautifully phrased apology leaves dinner just as absent.
For complaints handlers, retention teams and technical specialists, this is where the work becomes more demanding. The answer may require a refund, an investigation, a replacement or a change elsewhere in the business. Someone has to decide what is fair and then make it happen.
That requires authority. A human adviser who can only repeat the same refusal offers limited relief after a bot has failed. A retention team should be able to distinguish a problem worth solving from a customer’s settled wish to leave. A technical specialist needs access to the people who can fix a fault, as well as language that explains it.
There is also a cost to waiting too long to bring someone in. A separate 2026 Alibaba working paper examined an autonomous service system supervised by human workers. Chats became shorter, but ratings fell for the issues eligible for automation. In its analysis of handovers, service quality held up better for technical difficulties than for situations already marked by frustration; earlier intervention was associated with better outcomes. The experiment ran for 17 days in August 2024, and only a small share of chats qualified for the AI system. It is an early, bounded result. Wang and colleagues.
For a service team, the practical implication is to watch the handover closely. Has the next adviser received the history? Can they act? Is someone available before the customer has exhausted their patience? Those questions deserve as much attention as the bot’s first reply.
The harder queue
If automation takes more straightforward enquiries, the people who remain could spend more of their day dealing with exceptions, distress and disputes. That is a plausible consequence of changing the mix of work, rather than a finding that every contact centre has already reached this point.
It could make the job more satisfying. Less copying and pasting; more opportunities to solve a problem properly. It could also make a shift punishing if the simpler conversations disappear while targets stay the same.
Imagine two teams handling the same number of contacts. One deals mostly with delivery updates. The other deals with disputed payments, repeated faults and customers in difficult circumstances. Counting contacts alone would tell a manager very little about the effort involved. The same problem arises when yesterday’s target is applied to tomorrow’s more complicated queue.
Workforce planners and analysts will need to look beyond total volume and average handling time. Which cases return? How much work follows the conversation? Where does a failed automated exchange add to the next adviser’s workload? A lower cost per chat can coexist with a higher cost of putting the whole matter right.
Quality teams face a similar choice. They can use automation to search for missing steps, flag possible errors and find conversations worth reviewing. They should also check whether a correct-sounding answer led to the promised action. A refund mentioned in a transcript and a refund received by a customer are different events.
For team leaders, coaching may shift towards judgement: why the policy failed this person, why the suggested answer was incomplete, why an apparently long conversation was time well spent. Staff need room to raise those concerns without every departure from the script counting against them.
What becomes of the career?
It would be comforting to promise that better tools simply mean better jobs. The evidence does not justify that promise.
The US Bureau of Labor Statistics projects customer-service representative employment to fall by 5% between 2025 and 2035 as self-service and other technology take on more tasks. It also projects about 289,500 openings a year, all from replacing people who leave the occupation or the workforce. These are US projections, not observed AI job losses or a forecast for Britain. They show how a shrinking occupation can still be hiring. BLS occupational outlook.
For an individual worker, the route forward deserves more substance than advice to “learn AI”. A useful ambition is to become better at understanding a problem from beginning to end: the customer’s account, the relevant policy, the system that carries out the action and the reason things sometimes fail.
New advisers need practice checking suggested answers against the source, with feedback from someone who knows the work. Experienced staff should have routes into specialist casework, coaching, quality, knowledge management or service improvement where those roles exist. None offers guaranteed protection from job cuts. Each builds on knowledge that is useful beyond producing a reply.
Knowledge managers, in particular, have a consequential job. They need to establish which policy is current, who owns it and what happens when two sources disagree. An assistant makes unreliable information easier to distribute as readily as reliable information. Keeping the source material sound deserves time and recognition.
There is something for workers to bring to their next conversation with a manager, too. Choose a few cases in which the suggested answer failed. Explain what you noticed, what you changed and what happened afterwards. That is concrete evidence of expertise. It also gives the organisation a chance to improve the service rather than quietly relying on staff to repair it.
Who gets the saved time?
At the top of the department, the choice becomes financial. A head of customer service or chief customer officer can use faster routine work to reduce waiting, train staff, investigate recurring faults or lower staffing costs. In practice, they may pursue several of these at once. The allocation should be explicit.
Before removing capacity, leaders should follow a sample of problems all the way through. Include customers who gave up, came back or needed somebody to reverse an earlier answer. Check whether an apparent saving has become unpaid effort for the customer or extra work for another team.
The FCA’s review also recorded customers having to repeat difficult personal circumstances or submit documents more than once. Its findings concern financial services, but offer any service designer a useful test: can a person get help without repeatedly proving how difficult their life has become? FCA review.
AI could make a worthwhile difference here. A checked summary could spare someone another retelling. Reliable information could save an adviser a long search. A completed routine request could leave both people with a little more time. Each deserves to be judged by whether it actually helps.
The man in the FCA’s case needed the demands to stop. A new layer of software would have earned its place by helping somebody make that happen.
That is the promise worth holding AI to. When someone asks for help, the answer should leave them with less to carry.
Research note: Desk-researched analysis, checked 1 November 2026. The opening is a documented FCA case, not an interview conducted by this publication. Two Alibaba studies are working papers; publication dates and experiment dates are distinguished above. The strongest experimental evidence discussed concerns digital chat, so its results should not be assumed to apply unchanged to telephone or face-to-face service. Career and management implications are the publication’s analysis. Research and drafting were assisted by AI.
Continue the series: HR’s new power struggle.