The AI Efficiency Trap: Are You Automating Costs or Automating Away Customer Value?

We were recently discussing a website renovation project with a client when the subject of AI and automation came up.

Like many established businesses, they were considering where AI might fit into their customer journey, whether that meant customer service, website functionality, internal processes or some combination of the three.

The conversation reminded us of a presentation by the well known advertising executive and behavioural thinker Rory Sutherland.

One of Sutherland's recurring arguments is that businesses can become so focused on efficiency, automation and measurable cost savings that they accidentally remove things customers actually value.

It is an especially relevant warning now.

AI gives businesses extraordinary opportunities to reduce repetitive work, improve response times and make services more efficient. But there is an important question that should probably come before “What can we automate?”

What are our customers actually getting from this interaction, and could automation remove part of that value?

That question is at the heart of this article.

The pressure to introduce AI

Senior teams are being asked perfectly reasonable questions. Where can AI improve productivity? Which repetitive processes can be automated? Can customer service become faster? Can employees spend less time searching for information? Can a business provide useful support outside normal office hours?

The scale of adoption means these are no longer theoretical questions. Office for National Statistics research published in July 2026 found that self reported AI use among UK businesses with 10 or more employees had increased from around 12 per cent in late 2023 to around 35 per cent in 2026, with higher adoption among larger firms.

The 2026 UK Business Data Survey found a similar size effect. Among large businesses that handled digitised data, 82 per cent reported using AI for at least one purpose and 21 per cent reported using AI for customer service chatbots.

Many organisations are already experimenting. The more useful question is how they decide where AI belongs.

Cost reduction will inevitably form part of that discussion. It should. A tool that removes unnecessary administration, speeds up a routine process or helps an employee solve a problem faster can create genuine value.

The problem begins when cost reduction becomes the only measure of whether the project has succeeded.

Rory Sutherland and the Doorman Fallacy

Rory Sutherland, Vice Chairman of Ogilvy UK, used a simple hotel example at The Drum's Predictions 2026 event to explain what he calls the Doorman Fallacy.

Imagine a hotel looking at the person standing at its entrance. Management identifies the most obvious task: the doorman opens the door. An automatic door can perform that visible function more cheaply. The role disappears and the spreadsheet records a saving.

On paper, the logic looks impeccable.

But opening the door may have been only a small part of what the person was actually contributing. A good doorman might recognise returning guests, notice when someone needs assistance, provide reassurance late at night, help with luggage, give useful local advice, create a sense of security and make the hotel feel more personal from the moment somebody arrives.

Those benefits are harder to put into a cost saving model, but that does not make them worthless.

Sutherland's wider point is that businesses can become very good at measuring a visible function while overlooking the less visible value around it. In The Drum's account, he describes this as quantification bias: what is easy to measure attracts attention, while harder to measure value can be ignored.

A useful summary is simple:

Businesses often measure the visible function while overlooking the invisible value.

That is not an argument against automatic doors. It is an argument for understanding the whole job before deciding that the easiest part to describe is the whole source of value.

Why AI makes this mistake so tempting

AI makes the Doorman Fallacy especially tempting because many of its benefits are immediately measurable.

The cost of a software licence can be measured. So can the number of enquiries handled, minutes saved, cases processed and tasks completed without staff involvement. If a business reduces headcount, the saving appears quickly and clearly in the accounts.

Potential value loss is often slower and less tidy. It may appear as more repeat contacts, weaker conversion, more complaints, lower retention, higher escalation, abandoned purchases or employees spending time repairing failed automated interactions.

None of this means businesses should stop measuring efficiency. The lesson is the opposite. They should measure more of the outcome.

If an AI project saves time, ask what happens to resolution rates. If it reduces contact centre cost, ask what happens to repeat contacts. If it pushes customers towards self service, ask whether conversion, trust and retention change. If staff spend less time on routine work, ask whether that time is actually redirected towards more valuable work.

The right response to quantification bias is not to abandon numbers. It is to choose a broader set of numbers.

Customer service is where the problem becomes obvious

Customer service makes the distinction easy to see because not every request carries the same level of risk, emotion or judgement.

A customer asks:

“I need another copy of my invoice.”

That may be an excellent task for automation. If AI can authenticate the customer, retrieve the correct document and send it immediately, the customer gets a faster result and an employee avoids repetitive administration.

Now imagine a different message:

“We have been charged £4,800 incorrectly. I have contacted you three times already and if somebody does not resolve this today we are cancelling the contract.”

The second customer should not have to fight an automated assistant simply to find a person who can take responsibility.

This is where the distinction between replacement and augmentation matters. A sensible service design might allow AI to handle routine enquiries, identify intent, summarise the customer's history and present relevant information to an employee. A person then handles the judgement, exception, emotion and accountability.

That is not a failure to automate. It is automation being used in the part of the journey where it improves the outcome.

What current evidence tells us about digital customer service

Independent research gives businesses another reason to look beyond the cost of the channel itself.

In September 2026, Citizens Advice published Behind digital walls: how people are being locked out of essential services. The report is about digitalisation and access to essential services more broadly, not specifically about AI, so it should not be treated as an AI study.

Its findings are still highly relevant to automation design. Citizens Advice identified three recurring problems: erosion of phone and face to face support, people becoming locked into a particular channel, and poor digital design creating enough friction that users struggle or give up. It called for people to be able to reach a human when needed and to move more easily between digital and non digital support.

That is a useful design principle for AI customer service too. An automated channel should make straightforward interactions easier. It should not become a barrier that a customer must defeat before the business will listen to them.

The Competition and Markets Authority's 2026 research on agentic AI makes a related point from a different direction. It notes that current business deployments are still concentrated in relatively bounded areas such as customer operations, commerce workflows, IT and internal processes. In consumer facing settings, authority is often limited and escalation to humans remains common.

That is a much more realistic picture of sensible adoption than the idea that every human interaction should disappear.

Where businesses should use AI

There are many places where AI can create obvious value without sacrificing the customer experience.

Inside a business, AI can summarise meetings, search internal knowledge, analyse documents, classify enquiries, identify patterns in feedback, support research and handle routine administration. Customer facing uses can also be valuable, from simple FAQs and appointment information to order status and initial enquiry triage.

The most interesting opportunity may be employee augmentation.

Instead of asking “Can AI replace this employee?”, ask “Can AI make this employee significantly better at helping the customer?”

Imagine an adviser receiving a difficult enquiry. Before they respond, AI can assemble the relevant customer history, previous conversations, policies, product information and likely options. The employee spends less time hunting through systems and more time understanding the problem and making a good decision.

That can produce what businesses wanted from automation in the first place: faster service with less wasted effort, while using efficiency to support judgement rather than replace it blindly.

The questions businesses should ask before automating

Before removing a human step from a customer journey, it is worth slowing the decision down long enough to ask what that step is actually doing.

This framework does not make automation slower for the sake of it. It reduces the risk of optimising a task while damaging the wider journey.

Measure value, not just savings

Consider a fictional board presentation announcing:

“Our AI customer service project saved £600,000.”

That may be excellent news. It may also be only half the analysis.

What happened to first contact resolution? Did repeat contact increase? Were more complaints escalated? Did conversion change? What happened to retention, abandonment, refunds, satisfaction, resolution time and employee workload?

The £600,000 figure in this example is deliberately fictional. The point is not that a saving hides a loss. The point is that a saving alone cannot tell you whether the system improved the business.

A project that reduces operating costs while causing a larger loss in sales, trust, retention or productivity elsewhere is not necessarily an efficiency improvement. It may simply have moved cost into a place the original business case did not measure.

Do not use AI to paper over a poor website

This matters particularly when a business is considering adding AI to its website.

Suppose customers keep asking where to find a particular service. The immediate response may be to add a chatbot that can answer the question.

But the underlying problem might be confusing navigation, badly organised services, weak information architecture, poor search, unclear calls to action, missing answers, difficult forms, weak mobile usability or content that has not been maintained.

Adding AI can help in some of those situations. It can also create another interface through which customers have to navigate before reaching information that should have been clear in the first place.

Automating a confusing process can simply create a faster confusing process.

That is Toolbelt Websites' interpretation, not a Rory Sutherland quotation.

Before introducing another layer of technology, established businesses may benefit from first reviewing the structure, content and conversion journey of the website itself. In some cases, renovating an existing website and fixing the underlying customer journey will solve problems that another piece of software would otherwise be asked to hide.

Do not accidentally automate your competitive advantage

There is another strategic risk in making automation the default answer. If competing organisations use similar models for support replies, sales emails, onboarding, marketing content and customer communications, competence may become easier to reproduce. Businesses can become faster while also becoming harder to distinguish.

Sutherland makes a related argument about differentiation. In a world full of automation, doing something differently can become more valuable precisely because fewer competitors are willing to do it. As he put it at the 2026 event, “Different is better than better.”

That does not mean every customer needs a telephone call or dedicated account manager. It means businesses should identify the moments where human involvement creates competitive value: taking ownership of an unusual problem, remembering important context, giving thoughtful expert advice or recognising when reassurance matters more than another automated follow up.

Efficiency may become abundant. Difference may become scarce.

A simple framework for sensible AI adoption

A useful way to summarise the whole argument is with three principles.

1. Automate the repetitive

Use machines for predictable, repeatable work where automation genuinely improves speed, consistency or convenience.

2. Augment the valuable

Use AI to give good employees better information, better tools and more time to solve the problems that deserve attention.

3. Protect the human moments

Deliberately preserve human involvement where judgement, reassurance, accountability, emotion or relationships materially affect the outcome.

How Toolbelt Websites thinks about AI

At Toolbelt Websites, we see AI as infrastructure rather than a substitute for judgement. We use modern technology where it can make research, production and repetitive work faster or more consistent, while retaining human involvement where understanding the business, its customers and its objectives affects the quality of the finished result.

That same principle applies to a professionally managed website service. Technology should make the service more effective. It should not become an excuse to stop thinking about the person using the website.

We are not an enterprise AI consultancy. Our relevance is narrower: websites, digital customer journeys, marketing, conversion and how technology changes the experience customers have with a business.

Conclusion: do not mistake efficiency for progress

Rory Sutherland's Doorman Fallacy is useful because it forces a deceptively simple question: are we measuring the whole source of value, or only the part that happens to be easy to count?

A company can become faster, cheaper, more automated and more technologically sophisticated while simultaneously becoming worse for customers to deal with.

The question therefore should not simply be:

“Can we automate this?”

It should be:

“What value is being created here, and how can AI improve it without accidentally removing the part that matters?”

If you are reviewing how AI, automation or a website redesign fits into your wider customer journey, talk to Toolbelt Websites about your existing setup. We can help you look at the website you already have and identify what is worth keeping, what needs improving and where technology can genuinely make the experience better.

Sources and further reading

• Rory Sutherland, 2026 Predictions, The Drum Labs, YouTube.
• The Drum, Rory Sutherland: Why marketing's biggest risk in 2026 is mistaking efficiency for progress, 29 January 2026.
• Office for National Statistics, Artificial intelligence in UK businesses: 2023 to 2026, 20 July 2026.
• Department for Science, Innovation and Technology, UK Business Data Survey 2026, 18 June 2026.
• Competition and Markets Authority, Agentic AI and consumers, 9 March 2026.
• Citizens Advice, Behind digital walls: how people are being locked out of essential services, 22 September 2026.