UK Risks Falling Behind in the AI Race Without Faster Telecoms Upgrades, Warn Executives

On 29 August 2026 the networks director of the UK’s largest mobile operator said something that operators normally pay agencies to prevent them from saying. Speaking about the state of British mobile infrastructure, Andrea Donà of VodafoneThree put it plainly: “I am embarrassed today at the level of quality of our network[s]. The UK is worthy of what EE, O2 and VodafoneThree can offer, but we need reform.” That admission arrived alongside a warning with a specific timescale attached. The occasional network failure that British users already accept as normal — the dead signal in a packed stadium, the message that will not send at a mainline station, the video call that degrades the moment a train pulls into a busy platform — will, on his account, stop being occasional. Within the next two to three years, as artificial intelligence pushes a step change in the volume of data moving across mobile networks, that experience becomes the everyday one.

The warning does not stand on its own. It sits on top of a Which? analysis built on Opensignal measurement data which found that UK mobile coverage is worse than that of all 27 EU member states and every other G7 country. The headline placements are unflattering and specific: 57th globally for overall network performance, 70th for download speeds, and 55th on the composite reliability metric that captures whether calls connect, streams hold and games stay playable. Those are not rankings against emerging markets; they are rankings against the economies the UK competes with for investment, headquarters and data-intensive work. Today, 31 August 2026, two days after those comments were published, the question they raise is not really about consumer complaints. It is about whether a country that is betting heavily on artificial intelligence has built the connectivity that mass AI adoption will actually consume — and what a UK business should do about it while the answer is still being argued over in Whitehall.

What the operators actually said

Strip the story back to its load-bearing claims and there are four of them. The first is measurement: the Which? report, drawing on Opensignal’s crowdsourced network data, places the UK 57th in the world for network performance, 70th for download speeds and 55th on the reliability measure covering calls, streaming and gaming. The comparison set matters more than the absolute position. Being behind all 27 EU member states and every other G7 country removes the usual mitigations. This is not a story about geography, population density or an unusually rural landmass — the countries ahead of the UK include several with harder terrain and more dispersed populations. It is a story about how much capacity has been built and how quickly.

The second claim is the trajectory. Donà’s point is not that networks are failing now; it is that the demand curve is about to change shape. Today’s congestion is event-driven. It appears when tens of thousands of people converge on one cell — a stadium, a festival, a station concourse at 17:45 — and it disappears when they disperse. That pattern is tolerable because it is predictable and bounded. AI-driven traffic does not behave that way. Model inference at the edge, always-on assistants, real-time transcription and translation, agentic tools that poll and stream continuously, video understanding on device and in the cloud: these produce sustained, uplink-heavy, latency-sensitive load that does not go home at six o’clock. His forecast is that within two to three years the stadium experience becomes the ordinary experience, because the ordinary baseline will have risen to meet what today only happens in a crowd.

The third claim is about money and where it is already going. VodafoneThree has committed £11bn to upgrading masts to 5G standalone — the version of 5G that runs on a native 5G core rather than borrowing 4G infrastructure, and the version that actually delivers the low, consistent latency and network slicing that data-intensive applications need. Crucially, the operator is not asking to build more sites. It is reducing its nationwide mast count by roughly 30% while upgrading what remains. Around 96% of the upgrades it needs to carry out are on existing mast sites. This is the detail that reframes the whole debate: the capacity problem is not primarily a problem of new masts appearing in places people would rather they did not. It is a problem of existing masts being allowed to carry better equipment.

The fourth claim is the obstacle. Despite 96% of the work landing on sites that already exist, more than half of those upgrades still require planning permission — a process the industry describes as lengthy and bureaucratic, and one that applies per site, per authority, per application. Donà’s framing of the remedy is deliberately pitched at a Treasury audience: planning reform is “the lowest possible cost growth policy” because it requires no government spending at all. The capital is already committed and sitting on operator balance sheets. What is missing is permission to spend it quickly. He has urged the new prime minister to treat it as an early win. A second senior telecoms executive framed the stakes in different language: the intersection of AI and connectivity is currently a “nascent” issue, but it will become a “national competitiveness issue” — and the precedent they reached for was the UK’s slow full-fibre broadband rollout in previous years, a delay whose costs were paid quietly, by businesses, for a decade.

Why the timing is the whole story

The uncomfortable part of this warning is not the current ranking. Rankings move. It is the arithmetic of two curves crossing. On one side, a capacity programme with £11bn behind it, 96% of it on existing sites, more than half of it gated behind individual planning decisions that are measured in months rather than weeks and which any authority can extend. On the other, an adoption curve for AI tooling that is not waiting for anybody’s planning committee. Businesses are deploying assistants, copilots, transcription, summarisation and agentic automation now, on whatever connectivity they happen to have.

If the capacity curve is slower than the demand curve, the shortfall does not distribute itself evenly or fairly. It concentrates in exactly the places where economic activity is densest — city centres, business districts, transport hubs, industrial parks — because that is where the cells are already busiest. The organisations that feel it first will be the ones whose work is mobile, distributed and time-sensitive: field service, construction, logistics, healthcare in the community, events, retail with mobile point-of-sale. These are not fringe use cases. They are a substantial slice of the UK economy, and almost none of them have a fixed line to fall back on at the point of work.

There is a second-order effect that gets less attention. When mobile capacity is tight, the applications that suffer first are the ones with the least tolerance for jitter and packet loss — real-time voice, video, and any interactive AI feature with a streaming response. Those also happen to be the applications a business is most likely to have adopted in the last eighteen months and least likely to have instrumented. When a hosted phone call breaks up on a 5G tether, almost nobody logs a network fault. They blame the phone system. When an AI assistant takes eleven seconds to start answering instead of one, nobody opens a ticket about uplink contention. They stop using the assistant. The cost is real and it is invisible in every dashboard a typical UK business currently looks at.

How the UK arrived at 57th — and what happens next

The chronology below is the shape of the problem rather than a list of announcements. It matters because every stage in it was individually defensible, and because the final three entries are the ones a business can still plan around.

Which parts of a UK business actually depend on mobile capacity

Most organisations underestimate their mobile dependency because it is spread across functions and nobody owns the total. The assessment below is Cloudswitched’s own indicative view of how exposed each common business function is to sustained mobile congestion — not a measurement from the Which? or Opensignal data, but a working ranking we use when we scope connectivity resilience for clients. Read it as a prompt for where to look first, not as a statistic.

Two entries in that list deserve unpacking. Broadband failover sits near the top because it is the single most common piece of connectivity design a UK business has in place, and the one least likely to have been tested under realistic conditions. A 4G or 5G backup circuit is bought as insurance, sits idle for months, and is expected to carry a full site’s traffic on the day the primary line fails. If the local cell is already contended at peak, the failover does not fail cleanly — it comes up, passes a ping test, and then delivers a fraction of the throughput the site needs. Staff conclude that the internet is “still down”. Technically it is up. Practically it is useless. That distinction is exactly what sustained congestion turns from an edge case into a pattern.

Real-time AI assistants sit high for a subtler reason. Their traffic profile is unusual: small requests, large streaming responses, frequent round trips, and increasingly a meaningful uplink component when audio, images or screen context are being sent. Latency and jitter matter more than raw bandwidth. A connection that comfortably streams video can still deliver a poor experience for an interactive assistant, because video buffers and conversation does not. This is the specific mismatch behind the two-to-three-year warning: the workloads arriving next are not simply bigger than the ones the network was sized for, they are shaped differently.

The number that reframes the planning argument

If there is one figure in this story worth carrying into a conversation with a local authority, an MP or a landlord, it is this one. The overwhelming majority of the upgrade work required to modernise the UK’s largest mobile network does not involve a new mast anywhere.

Public debate about mobile infrastructure is overwhelmingly a debate about new masts: where they go, what they look like, who objects. That debate is legitimate and it is also, on these numbers, largely beside the point. Ninety-six per cent of the work is swapping and augmenting equipment on structures that already exist, on land already leased, in locations already accepted. The operator is not asking to expand its footprint — it is shrinking that footprint by around 30% while making the remaining sites substantially more capable. And yet more than half of those same-site upgrades still enter a permitting process designed principally for new development.

That is the gap Donà is pointing at when he calls planning reform the lowest-cost growth policy available. There is no subsidy to argue about, no capital allocation to find, no new spending line to defend at a fiscal event. The £11bn already exists. The constraint is administrative throughput: how many same-site upgrades can be consented, by how many separate authorities, in how many months. Reform here is unusually cheap in the literal sense — it changes the speed at which private capital can be deployed without changing what the public purse pays. Whether that argument lands is a political question. Whether the delay has a cost is not.

Where UK businesses are most exposed right now

National capacity is not something an individual organisation controls. What it does control is how much of its own operation is sitting on an assumption about mobile performance that nobody has tested. The grid below reflects what we typically find when we audit connectivity for a UK SME — the exposures are ranked by how often they turn out to be present and unmanaged, not by how technically serious they sound in isolation.

The pattern in that grid is worth naming. Five of the eight exposures are not equipment problems — they are evidence problems. The business owns the right hardware and pays for the right services, but nobody can currently answer a simple question with a document: what happens to this site if the fixed line drops at eleven o’clock on a Tuesday, and how much of what we do depends on a mobile connection whose performance we have never measured? That question is answerable in an afternoon per site. It is almost never answered before the day it matters.

The third row deserves particular attention because it is the one most often introduced by accident. A business buys a leased line or FTTP circuit and adds a 4G or 5G backup, believing it now has diverse paths. In a meaningful number of cases, the mobile backup terminates on a mast whose backhaul runs through the same exchange, along the same duct, as the primary. Physical diversity was purchased. Logical diversity was not delivered. Nobody discovers this until a single street works incident takes out both, and the post-incident review finds that the resilience design was correct on paper and wrong in the ground.

What connectivity resilience actually costs a UK business

The figures below are indicative monthly ranges we use for scoping conversations with UK organisations, not quotations, and they exclude one-off installation and hardware. They are included because the most common reason a business does nothing about connectivity resilience is an assumption that doing something is expensive. For most bands, the recurring cost of a properly designed and tested arrangement is smaller than the cost of a single day of degraded operation.

Two observations about that table. First, the largest single improvement in most bands is not a bigger circuit — it is a tested arrangement rather than an assumed one, which costs almost nothing beyond the discipline of scheduling it. Second, the line item that changes most as an organisation grows is not bandwidth but visibility. A 200-person business with field teams is not primarily buying more capacity; it is buying the ability to know, per site and per application, whether the connection is meeting what the work requires. That is precisely the capability that becomes decisive if national mobile capacity tightens over the next two to three years, because it is the difference between diagnosing a problem and arguing about it.

Assuming the network versus engineering around it

Nothing in the right-hand column requires waiting for planning reform, a general election cycle or a new mast. It is entirely composed of work an organisation can commission now, at a cost that sits in the ranges above, and its value increases precisely in the scenario the operators are warning about. That asymmetry is the practical argument for doing it. If national capacity improves quickly, a business with engineered connectivity has lost very little. If it does not, that business is one of the few that can tell the difference between a network problem and a supplier problem on the day it matters.

That score is our own judgement rather than a published figure, and it is deliberately not catastrophic. Most UK businesses have bought adequate connectivity; what they have not done is establish what it delivers. The points are lost almost entirely on the evidence side of the ledger: no tested failover, no application dependency map, no performance baseline, no verified path diversity. Those four items are cheap to fix and they account for the majority of the gap. The organisations that score well are rarely the ones spending the most on bandwidth. They are the ones where somebody once sat down with a list of sites and applications, measured what each actually needed, and then put a date in the calendar to check it again.

At a glance

The pattern this fits into

Read alongside what we have covered over the past few months, this is a familiar shape: infrastructure that everybody depends on, nobody owns end to end, and few organisations can produce evidence about. Our reporting on the NCSC advisory on internet-exposed edge devices made the point in a security register — the recurring failure is a device at the boundary of the organisation, trusted more than its management warrants, absent from any register. Connectivity is the same category of asset viewed from the performance side rather than the risk side. The router, the backup SIM and the failover policy are all things a business owns, depends on daily, and has rarely tested.

The same evidence gap runs through our analysis of the Manchester Airports Group breach affecting 8.7 million people, where the exposure sat in a layer nobody had classified as critical, and through the follow-up on what that incident says about penetration testing scope — organisations examine what is in the scope document, and the things that hurt them are reliably the things that were not. A mobile failover circuit is almost never in anybody’s scope document. It is bought as insurance and then treated as a settled matter.

Two further threads connect directly. Our coverage of the ransomware high-water mark recorded in July 2026 is a reminder of why connectivity resilience is not only a productivity question: an organisation running on a degraded backup link during an incident has fewer options for isolation, recovery and remote support at exactly the moment it needs more. And the Claude Code prompt-injection research published in August is a useful counterpoint to the optimism around AI rollouts — the tooling arriving in UK businesses this year brings new dependencies as well as new capability, and the network underneath it is one of them. Finally, for anyone whose sites are already being visited by engineers, the PSTN switch-off deadline facing UK businesses is quietly pushing new connectivity hardware into thousands of premises — lift lines, alarm circuits and door entry systems moving onto data and mobile connections, each one a fresh dependency that ought to arrive on somebody’s register rather than merely in somebody’s comms cupboard.

Frequently asked questions

CloudSwitched

London-based managed IT services provider offering support, cloud solutions and cybersecurity for SMEs.