When a company cuts 1,300 jobs, it has to say something about why. Centrica’s chief executive Chris O’Shea chose an unusual answer: the customers asked for it. His exact words were that “AI isn’t driving these particular job reductions” and that the change was “mainly due to changing customer behaviour” — over 90 per cent of customers now use digital channels in the first instance, and calls have fallen 20 per cent. Writing in The Telegraph, Melanie McDonagh’s response was a single sceptical question: who are these people who prefer a chatbot? It is the right question, but the more useful one sits underneath it. The figures O’Shea cited are real. They just do not measure what the argument needs them to measure — and that substitution, usage data standing in for preference, is a pattern any leader signing off an automation business case should learn to recognise in their own numbers.
The measurement that isn’t there
Preference is a statement about what someone would choose if the options were equally available. Usage is a record of what they did given the options in front of them. These come apart whenever the options are not equally available — which, in customer service, is nearly always.
A customer who spends eleven minutes in a phone queue, is passed between two automated menus, and gives up to use web chat has not expressed a preference for web chat. They have expressed a preference for getting the answer. Route enough of those customers and the digital-first figure climbs on its own. Nothing about their view of chatbots changed.
Strategic Reality: Channel usage is downstream of channel design. A company that controls the queue length, the menu depth, and how prominently the phone number appears also controls the usage statistic it later cites as evidence of what customers want.
This is what makes the Centrica figures unusable as a justification. They are endogenous to the decision they are being used to defend. Set alongside the commercial position, the picture sharpens further.
| Figure | What it is offered as evidence of | What it actually records |
|---|---|---|
| Over 90% of customers use digital channels first | Preference for digital service | Which channel was easiest to reach at the point of need |
| 20% fall in call volume | Falling demand for human contact | Throughput on one channel, after that channel changed |
| 1,300 roles cut over two years | Capacity matched to demand | A committed cost reduction |
| 14% reduction across six UK sites | Proportionate adjustment | The scale of that commitment |
| Retail profit £346m, up £8m year on year | — | Margin improving |
| Domestic customers 7.45m, down from 7.5m | — | Volume falling while margin rises |
Profit rising as customer numbers fall is the signature of a business chasing margin, not one responding to a shift in what its customers want. O’Shea said as much himself in the same set of results: British Gas has focused on bigger margins from its fixed-price tariffs rather than “chase loss-making business”. That is a legitimate strategy, stated plainly. It is simply a different story from the one told about the call centres.
There is a second wrinkle. The company’s own framing of the 800 fresh cuts describes a “targeted deployment of AI tools” — so AI is present in the announcement even as the chief executive denies it is the driver. Both statements can be technically true at once. Together they show how much work the customer-behaviour explanation is being asked to do.
How the loop closes
The pattern has three steps, and each one is individually defensible.
First, the human channel is deprioritised — not abolished, just made slower and harder to find, usually for cost reasons that were never announced as strategy. Second, usage of that channel declines, which it must. Third, the decline is presented as revealed preference, justifying the removal of the capacity that was degraded in step one.
Nobody in this sequence has to lie. The queue times were a resourcing decision, the usage drop is genuine, and the inference from usage to preference is the sort of thing that passes without challenge in a results presentation. The circularity only becomes visible when you ask which step caused which — and by then the redundancy consultation has started.
What the story buys is not really cost. It is the reallocation of blame. A board that says “we are cutting 1,300 roles to reduce cost” owns the decision, and can be argued with. A board that says “our customers changed” has handed responsibility to an aggregate that cannot be cross-examined, does not give interviews, and has no representative in the room.
Critical Context: Note the inversion against the opposite move. Staffing group Adecco recently warned that some firms blame AI for layoffs actually driven by weak trading — AI as cover for commercial pressure. Centrica’s framing runs the other way, denying AI and crediting customers. Two opposite narratives, one shared function: move the decision away from management.
What the evidence actually says
If digital-first usage told us about preference, direct measurement of preference would agree with it. It does not.
Research by Pegasystems with YouGov, published in February 2026 and drawn from fieldwork with 4,748 adults in the UK and US between 4 and 13 November 2025, found that two-thirds of consumers prefer human-led support, 77 per cent say they always or often get better outcomes dealing only with a human, and 2 per cent — two — want to interact exclusively with generative AI chatbots. Nearly half, 46 per cent, said AI-powered interactions rarely or never produce a successful outcome. The figures are weighted evenly across the two countries rather than reported for the UK alone, so treat them as directional for a British supplier rather than exact.
Ask people what they use and you get the digital-first number. Ask them what they want and you get something close to its opposite. Both are true at once, and the gap between them is the artefact.
McDonagh gets at the same thing from lived experience rather than survey data. Her point is not that human service is uniformly good — she is scathing about the call that ends with a request to rate the experience, and about the encounter with BT that was pleasant and failed to achieve anything. Her point is that the good outcome, the one where a person finds a way around a problem that fits no category, is only available from a person. “We’ve beaten the Machine,” her housing-company contact said. That is a resolution no containment metric captures, because containment counts the call that never escalated, not the problem that got solved.
Reality Check: Preference is also not the operative test in a regulated market. Ofgem’s consumer standards require suppliers to be easy to contact by different methods, such as email and phone, at times that meet customer needs, and to publish their Citizens Advice star ratings. An access obligation is not discharged by a claim about what the average customer prefers.
Who pays for the story
| Stakeholder | What the framing gives them | What it costs them |
|---|---|---|
| Board and executive | Converts a redundancy decision into a response to demand | Destroys the feedback signal that would reveal where automation fails |
| Frontline staff | Nothing | The cut is framed as their irrelevance rather than a cost choice, leaving nothing to contest |
| Customers | Nothing | Channel choice narrows while the narrative says it widened |
| Unions | A clear target: automation dressed as consumer choice | Pushes the argument onto contested statistics rather than the decision itself |
| Ofgem | Nothing | A preference claim offered where an access duty applies |
| AI vendors | A reference case | One built on containment rather than resolution, which weakens the next sale |
The board’s entry is the one that should worry commercial leaders most. A company that believes its customers wanted the chatbot has no reason to investigate the cases where the chatbot fails. Complaints become noise from a shrinking minority rather than evidence about the deployment. The story removes the organisation’s ability to learn from its own automation.
Testing your own channel data
Most organisations reading this are not Centrica, but most are looking at a channel-shift chart that points the same way. The question is whether yours is preference or artefact. Three tests separate them, and none requires new technology.
Measure stated preference directly. Ask customers what they would choose if both channels were equally quick. If the answer diverges sharply from your usage data, the usage data is describing your queue design, not their wishes.
Replace containment with resolution. Containment rate — the share of contacts handled without a human — rewards a system for refusing to escalate. First-contact resolution and repeat-contact rate within seven days measure whether the problem went away. A deployment that raises containment while raising repeat contacts is failing and reporting success.
Watch the exception cases. Aggregate satisfaction hides the pattern that matters: the queries that fit no category. Track escalations by reason, not just by volume, and note which categories fall to zero. A category that disappears has often been absorbed into customer abandonment rather than automated resolution.
Implementation Note: Run these before the business case, not after the cuts. Each of the three is cheap and answerable in a fortnight from data most contact centres already hold. Their value is entirely in the sequencing — after the headcount decision they become a defence exercise, and they will be conducted accordingly.
For organisations further along, a fourth test applies: hold the human channel at equal prominence for a fixed sample of customers for one quarter, and compare. If digital-first usage holds up under equal access, the preference is real and the business case is sound. If it collapses, the channel-shift chart was measuring the friction you built.
Four problems the story creates
The KPI inherits the error. Once automation is justified by preference, the metric that tracks it becomes containment, because containment is what preference implies. The organisation then optimises for calls that never reach a human, which is the same thing as optimising against escalation for the customers who most need it.
The claim becomes load-bearing in the wrong place. A preference statement made to the financial press has a way of resurfacing in a regulatory response, where it is not evidence of anything. Access obligations attach to what the supplier makes available, and a firm that has staked its position on customer wishes finds itself defending the wrong proposition.
Attrition removes the wrong people first. Centrica’s cuts run over two years, partly through unfilled resignations. Voluntary departure selects for the staff with the most options — typically the experienced ones who handle exceptions. The organisation loses precisely the capability that automation cannot replicate, and loses it invisibly, because headcount reduction looks the same on a chart regardless of who left.
Internal belief contaminates strategy. A claim repeated in enough decks stops being a claim. Within eighteen months “our customers prefer digital” is an assumption in the planning model rather than an assertion someone once made about a usage statistic, and nothing in the process is designed to re-test it. This is the most expensive of the four, and the hardest to reverse.
What to hold onto
The narrow lesson is about Centrica, and it is McDonagh’s: be sceptical of figures that so conveniently validate a decision already taken.
The broader one is about measurement discipline, and it applies whether or not you ever cut a job. Usage data describes the system you built. Preference data describes what people want. Automation business cases fail when the first is presented as the second, because everything downstream — the KPI, the escalation design, the investment case for the next phase — inherits an error that nobody can see any more.
Three things make the difference in practice:
- Separate the two measurements before the decision. Direct preference measurement is cheap, fast, and only useful before a commitment exists to defend.
- Pick metrics that can report failure. Containment cannot. Resolution and repeat contact can, which is exactly why they are less popular in board packs.
- Own the decision in the language you use. “We are reducing cost” is contestable and honest. “Our customers changed” is neither, and the organisation that says it loses the ability to find out whether it was true.
The UK debate about AI and employment is going to be dominated by this framing question rather than by the technology, because the technology is not what determines whose account of a job cut prevails. As we argued in our analysis of who actually loses when AI reaches knowledge work, compelling framing routinely points exposure analysis in the wrong direction. Centrica’s version is more consequential than most, because it comes with a number attached — and numbers are believed.
Sources: Melanie McDonagh, “Don’t fall for British Gas’s great AI ploy”, The Telegraph, 26 July 2026. Chris O’Shea’s quotes, the job-cut and site figures, and Centrica’s retail results are as reported by The Guardian, 23 July 2026, and covered in our news report, British Gas owner cuts 1,300 jobs, cites shift to chatbots. Consumer preference data from Pegasystems and YouGov, published 25 February 2026. Supplier obligations from Ofgem’s customer service standards.
Resultsense analyses UK AI developments for professionals and businesses. Read more at Insights or get in touch.