Have we got AI completely the wrong way around?
I have a question for every BPO leader, outsourcing client and AI technology provider. What problem does your AI actually fix? Not what it can do. Not how many conversations it can handle, how many languages it speaks or how quickly it can produce a response. What problem does it fix, why does that problem matter, and what evidence do you have that fixing it has improved the business?
I ask because I am increasingly concerned that we have become so enthusiastic about implementing AI that we have forgotten some of the fundamental principles of running a successful customer operation. I use AI every day. I believe its potential is extraordinary. But I don't believe we are getting the best from it. I suspect too many implementations are delivering less value than they could because we are approaching adoption from the wrong direction.
Want the rest of it?
Tell me who you are and you can read the whole thing, then download it to keep. You only need to do this once.
We identify activities that can be automated rather than establish which business problems need solving. We measure the performance of technology rather than the performance of the operation. And we celebrate improvements to processes that perhaps should have been redesigned or eliminated before we considered automating them.
The debate seems to have become a choice between two extremes. AI is either the answer to everything or the death knell of the contact centre and BPO industry. I think both positions miss the point. AI is a remarkably powerful technology. Its value depends on what we do with it. And that is where I believe the industry needs to ask more uncomfortable questions.
We are implementing AI because we can, rather than because we should
I have witnessed situations where the desire to implement AI was the starting point, with the business justification following afterwards. One involved a small BPO developing a chatbot to provide quotations for a specialist UK insurance company. Quotations were selected primarily because they appeared relatively straightforward to automate, with the prospect of reducing costs. The solution was piloted with real customers and prospects. Insufficient discovery meant the chatbot could not accommodate some additional requirements. Customers found themselves in repetitive loops, some resorted to telephoning the business, and the project was abandoned. The vendor lost credibility; the potential value of lost opportunities remained unknown.
The issue wasn't simply that a pilot failed. Pilots sometimes fail, and that is part of learning. The issue was that the organisation had selected something to automate before establishing why it needed automating.
On another occasion, I encountered a client whose manager had set an annual objective to implement AI in three business processes. Which processes? That hadn't been decided. What problems needed solving? That hadn't been established. What benefits were expected? That hadn't been discussed. The objective was simply to implement AI three times. I find that extraordinary. Imagine instructing an operational leader to change three processes without specifying why they needed changing. We would question the logic immediately. Attach AI to the instruction and suddenly it becomes an innovation objective.
These are individual experiences, not proof that every organisation behaves this way. But they raise a serious question about how projects are commissioned. In June 2025 Gartner warned that many agentic AI projects were driven by hype, unclear business value and inadequate consideration of implementation costs and risks. Its prediction that more than 40% would be cancelled by the end of 2027 is a forecast, not an observed failure rate, and is not specific to BPO. But the reasons it identified deserve attention.
I suspect part of the problem is that AI has become something executives feel they must be seen to embrace. The instruction comes from above. The organisation needs an AI strategy. The BPO wants to demonstrate innovation. The vendor has a solution to sell. Somewhere in that process, we start looking for activities that might accommodate the technology. Surely the starting point should be a business problem worth solving, not a technology looking for somewhere to live.
Fix the process before you automate it
This is the argument I feel most strongly about. If a customer contacts a business because a delivery has failed, a payment has been processed incorrectly or a product doesn't work as promised, what have we achieved by automating the conversation? We may have reduced the cost of handling the enquiry. We may have provided a faster response. We may even have improved the experience of contacting us. All worthwhile possibilities. But the delivery still failed. The payment was still wrong. The product still didn't work.
Take a retailer receiving thousands of enquiries about delivery dates. An AI agent could provide tracking information, answer questions and reduce the number of contacts handled by employees. That might be a perfectly sensible investment. But what if the real problem is unreliable delivery information, inconsistent promises at checkout or a failure to communicate proactively when something goes wrong? Wouldn't the more valuable exercise be to examine the entire fulfilment and communication process?
Perhaps AI could identify orders likely to be delayed, initiate corrective action or alert customers before they need to ask. Perhaps better system integration or clearer delivery promises would remove the demand. Perhaps the best intervention would not require AI at all. This is an illustration, not a documented deployment, but it shows the difference between automating the contact and improving the operation that creates it.
The same principle applies to billing disputes, repeat technical faults, complaints and collections. How much of the work entering our contact centres is created by failures elsewhere in the business? How much is generated by confusing policies, disconnected systems, poor communication or processes designed around departments rather than customers? And how much could be prevented?
We have spent years improving the efficiency of handling contacts. AI gives us the opportunity to do that on an entirely different scale. But I wonder whether we are concentrating on the wrong end of the problem. Instead of asking how cheaply we can handle a contact, perhaps we should first ask why the contact was necessary.
This doesn't mean AI must wait until every process is perfect. AI may be one of the most useful tools available for analysing contact reasons, identifying patterns and helping us design better operations. But diagnosis should drive the technology decision. The customer journey should determine where AI belongs. AI should not determine how the customer journey must work. If the journey is broken, automating it may simply allow us to deliver the same poor experience more efficiently.
Successful automation is not necessarily business transformation
There is genuine evidence that AI improves operational performance. Research by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published in The Quarterly Journal of Economics in 2025, examined generative AI assistance across 5,172 customer-support agents. It found an average 15% improvement in issues resolved per hour. Benefits varied substantially: less experienced employees generally gained more than highly experienced employees. The study also found improvements in some aspects of customer and employee experience.
That is a substantial result and deserves recognition. But it demonstrates that AI assistance can improve the productivity of people handling enquiries. It does not establish that the underlying reasons for those enquiries have been addressed or that the entire operation has been transformed. We should not pretend that it does.
Consider a business that reduces average handling time by 15%. What happened to the released capacity? Did expenditure fall, did the business avoid recruitment, did availability improve, or was the benefit absorbed elsewhere? Did resolution improve? Did customers need to contact the business again? Did the improvement affect retention or profitability? There may be valuable answers, but the handling-time figure alone does not provide them.
The same applies to automation rates. If an AI agent handles 70% of enquiries, that tells us something about utilisation. It does not automatically establish whether customers got what they needed, whether they subsequently required assistance or whether the overall operation became less expensive.
There are substantial company-reported benefits worth examining. Klarna's 2025 annual reporting states that its AI assistant handled 80% of customer-service chats, with reported satisfaction comparable to human-assisted service, and estimates approximately $59 million in customer-service cost savings for the year. These are impressive disclosures. They are also company estimates, rather than an independent causal assessment of every implementation cost and customer consequence. We should neither dismiss them nor stretch them beyond what they establish.
I want to know whether AI is cheaper than the human alternative when we compare the complete cost of delivering the same outcome. Not just the cost of a bot conversation against the cost of an employee conversation. I mean implementation, integration, licensing, infrastructure, ongoing usage, maintenance, governance, oversight, exceptions, failed interactions and additional work created elsewhere. Then compare that with the fully loaded cost of an equivalent human operation and the quality of the result.
I am not suggesting AI cannot be cheaper. In many applications it may be considerably cheaper. I am suggesting that the answer should be demonstrated rather than assumed. A lower cost per interaction is not necessarily a better business result if conversion falls, repeat contacts rise or customers have to work harder. Equally, a change in customer satisfaction is not automatically commercially material. We should investigate its scale and consequences rather than make assumptions in either direction.
I want to know whether the operation is better, not simply whether the technology is performing well.
BPOs, clients and vendors are all part of this
I don't believe responsibility rests with one part of the industry. BPOs want to demonstrate innovation and remain competitive. Clients want efficiency, lower costs and improved experiences. Vendors want to develop and sell increasingly capable products. All understandable ambitions. But their interaction can encourage questionable decisions. The client announces that AI must feature in its strategy. The BPO proposes a use case. The vendor demonstrates a platform. Everyone is enthusiastic. Who has established whether this addresses the most important problem in the operation? And who is prepared to say the answer might not involve AI?
For BPOs there is a commercial contradiction we cannot ignore. A substantial part of the traditional industry earns revenue from supplying people, hours and capacity. Now we promote technology whose attraction is often that it reduces the hours required. It's rather like turkeys voting for Christmas. If AI removes substantial work, a provider paid for that work may be financially worse off.
I am not suggesting that BPOs deliberately create unnecessary demand or that every contract rewards inefficiency. But the incentive problem is real enough to discuss. We cannot promote eliminating unnecessary work while pretending that the commercial arrangements rewarding its delivery are irrelevant. This is not a paper about pricing models. It is about AI adoption. The two cannot be kept mutually exclusive.
BPOs have expertise that could be more valuable than simply supplying people to handle demand. They understand operational delivery, performance, customer interactions and the consequences of poor processes. Why shouldn't they sell the expertise to redesign an operation, challenge the client's policies and prevent avoidable work? That requires courage, because it may reduce the volume of activity they traditionally sell. Some clients will embrace the transition. Others may take work in-house or move to another provider. Neither choice automatically fixes the underlying process.
Clients have responsibilities too. They cannot demand frictionless customer journeys while refusing to change the policies, products and systems generating customer frustration. They cannot expect a BPO to eliminate failure demand if they only allow it to handle the resulting contacts. And they cannot insist that a familiar hourly rate is the only form of commercial certainty. A well-designed managed service can provide transparent cost for an agreed scope, with sensible provisions for material changes. It is not the answer to everything, but it deserves consideration when the goal is to remove unnecessary activity.
Vendors also need to raise their standard of proof. An impressive demonstration is not a business case. A system that can retrieve information, conduct a conversation or complete a transaction has demonstrated a capability. It has not necessarily demonstrated that the capability addresses a problem worth solving. Tell me, in two sentences, which operational problem your product fixes and what evidence shows the value of fixing it. Then tell me what happens when the customer takes an unexpected path, an exception arises or human intervention is required.
A vendor should be prepared to challenge an unsuitable use case. A BPO should be prepared to challenge a client's demand to automate something that doesn't need automating. A client should be prepared to hear that its biggest problem may be a policy, process or product failure. That is what partnership should mean.
The opportunity is bigger than automating conversations
I believe we are underestimating what AI could enable customer operations to become. Chatbots, voice agents, agent assistance, quality assurance, summarisation and intelligent routing all have legitimate uses. But the larger opportunity may be in understanding and preventing the demand that creates those interactions.
Imagine using contact intelligence to identify recurring product faults before they become widespread, recognising patterns that show a policy is confusing customers, or finding where information repeatedly disappears between departments. Imagine using AI to help diagnose the cause and support a redesigned journey, rather than simply automating the next conversation about the same failure.
Some of these capabilities are already emerging; others require better data, integration, governance and operational change. But the direction matters. Instead of asking which contacts can be automated, ask which can be prevented. Instead of asking how quickly a process can be completed, ask whether it needs to exist in its current form. Instead of asking how many people can be replaced, ask what those people could achieve if unnecessary work were removed.
Start by understanding the complete customer journey, including what happens before and after contact with the BPO. Examine the reasons for contact, where customers struggle, what creates repeat demand and which client policies or delivery practices contribute. Challenge the design. Then decide where AI, other technology, human expertise and simpler processes belong. Sometimes the right answer will be an autonomous agent. Sometimes it will be an employee with better information. Sometimes it will be a changed policy or a repaired system.
I am working with developing BPO businesses where I am encouraging a more cohesive approach to service design, technology and commercial arrangements, and I am speaking with prospective investors considering different approaches to entering the industry. Those conversations are first-hand observations, not proof of an industry-wide shift. But they remind me that new entrants need not inherit every assumption of the traditional model. Established providers have experience and relationships that matter. The question is whether they will use those advantages differently. Wishing the tide away will not stop it.
So is AI delivering the value promised?
My answer is that it is delivering real value in some applications, but I remain unconvinced that the industry as a whole is approaching adoption in a way that will realise its full potential. That is my opinion, informed by operational experience, published research and the questions that remain insufficiently answered. I am not claiming every implementation is misguided, every BPO protects the status quo or every vendor oversells its product.
I believe we have allowed enthusiasm for AI to get ahead of the operational thinking required to use it effectively. Businesses may invest in technology aimed at the wrong problem. Customers may encounter more complicated journeys. BPOs may undermine their traditional revenue without establishing a more valuable role. Vendors may lose credibility when demonstrations fail to become meaningful results. Most importantly, we may miss the opportunity to prevent customer problems rather than merely handle their consequences.
Fix the process before you automate it. Not because automation is bad, but because understanding what needs fixing is the only sensible place to start.
I want you to challenge me. But bring evidence.
I have deliberately taken a position. I believe too much AI adoption is driven by enthusiasm for the technology rather than a clear understanding of the business problem. I believe we are too quick to celebrate automation and productivity measures without establishing their contribution to the wider operation. And I believe we are missing opportunities to prevent customer problems rather than simply handle them.
You may disagree. Good. If you believe the industry is already approaching AI strategically, show me the evidence. If you believe it is delivering the transformational financial benefits being promised, explain how those benefits have been measured. If you believe automating an existing process without first challenging its design is the right approach, tell me why. And if you have examples of AI helping to redesign customer journeys, eliminate unnecessary demand and improve business performance, I genuinely want to hear about them.
I am not asking anyone to prove that my preferred approach has already worked in every situation. I am advancing an argument about how we should approach customer operations and technology. Challenge the reasoning. Challenge the evidence. Offer a better alternative. But let's stop treating opinions, forecasts and marketing claims as established facts. That applies to my arguments as much as anyone else's.
AI is not a panacea. Nor do I believe it is necessarily the death knell of the BPO industry. It is an extraordinary technology that offers us the opportunity to reconsider how customer operations should be designed and delivered. My concern is that we are so busy celebrating what it can automate that we are overlooking what it could help us change.
Before you tell me what your AI can do, tell me what problem it fixes. And before you tell me how much you've saved, show me how you've made the business better.
Let's continue the conversation
Through Laughing Leadership, I work with BPOs, outsourcing clients, investors and technology providers to challenge established thinking and identify opportunities to improve customer operations. That includes reviewing customer journeys, operational performance, processes, policies, technology adoption and the commercial arrangements that influence service delivery.
I am also available to contribute to board discussions, leadership meetings, conferences and industry events where a challenging, practical conversation about AI, BPO and the future of customer operations would be valuable. If you disagree with my argument, I would welcome the discussion. If you recognise some of these challenges in your own business, perhaps we should examine them together.
Dave D’Arcy | Founder, Laughing Leadership | dave@davedarcy.co.uk | www.davedarcy.co.uk
Research and source notes
Brynjolfsson, E., Li, D. and Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. https://doi.org/10.1093/qje/qjae044
Klarna Group plc (2025). Annual Report 2025. Company disclosures concerning customer-service AI, reported automation, satisfaction and estimated savings. https://www.sec.gov/Archives/edgar/data/2003292/000200329226000007/klar-20251231.htm
Gartner (25 June 2025). Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
The insurance chatbot pilot and executive AI implementation objective are anonymised first-hand examples, not representative statistical evidence. This is an opinion paper: the author's judgements are distinct from published research findings and company-reported results.

