The boring reason most UK small businesses stall on AI

A panel goes up on LinkedIn this week arguing whether Claude has overtaken GPT for SMB workflows. Two posts down, someone announces prompt engineering is dead and agents are the future. Three posts further, a founder shares a demo of their voice agent. The bot cuts out at the eight-second mark because the venue wifi can't push the audio fast enough. Nobody in the comments mentions it.

That's the discourse. Here's the part nobody's saying out loud.

The boring reason most UK small businesses stall on AI isn't the model. It isn't the prompt. It isn't the workflow. It's the broadband line underneath all of it. We see this every day with our business customers. UK SME AI adoption is real, somewhere between 35 and 39% are now active users with another 24% planning to adopt this year [1]. The discourse is about which model to pick. The friction we watch from the ISP seat is whether the line can run any of them.

This piece does four things. It explains why the model wars are a distraction. It names the patterns we actually see across our SME customer base. It says, plainly, why we're the ones writing this and not somebody else. And it points at the boring fix.

The model wars are a distraction

Most of what gets written about AI for small business is written for people who already have the connection sorted. That's London tech press writing for London tech press. The cadence of the discourse tells you everything. Claude vs GPT, then GPT vs Gemini, then Copilot vs everyone, then agents vs prompts, then prompt engineering is dead, then prompt engineering is back. The wheel turns every six weeks.

Anthropic shipped Claude for Small Business on 13 May 2026 with 15 ready-to-run workflows across finance, ops, sales and customer service [2]. The big consumer model makers shipped their own SMB pushes around the same time. The tools are good enough. They've been good enough for a year. The differences between them, for a salon, an electrician, a corner cafe, are at the margins. You can run the workflow you actually need on any of the front-three options. You're going to be fine.

Meanwhile, a meaningful slice of that 35-39% adoption number stalls. Not because the model was wrong. Because the line behind it couldn't keep up. Voice agents stutter. Multi-megabyte uploads spin and die. The Claude session loses its place when the connection drops at 4:47 on a Tuesday because somebody started a Teams call. The owner doesn't blame the line. They blame the tool, then they blame themselves, then they quietly stop using it.

The model wars are a debate. The line under your desk is a decision. Let's be honest about who benefits from talking about the first thing and not the second.

What we actually see

We see this every day with our business customers. The shapes repeat. Four of them are doing most of the damage.

The first is voice agents stuttering on a copper line. The AI receptionist is brilliant on the demo. In production, on an FTTC (fibre to the cabinet) line with 15 Mbps upload and 30 to 40 ms latency, the turn-taking falls apart. The caller hears half a beat too much silence and starts talking over the bot. Two weeks in, the owner messages us to say AI receptionists aren't ready. The bot was fine. The pipe wasn't.

The second is document uploads stalling mid-prompt. A quote bundle with site photos, a batch of receipts for the VAT run. 30 megabytes going up on an FTTC line at 18 Mbps headline, real-world closer to 12. Claude shows a spinner. The owner closes the tab and goes back to typing the reply by hand. AI is "too slow". The model isn't slow. The model is sitting on a GPU somewhere waiting for the file.

The third is the office that kills its own connection. Three people on a Teams call, which Microsoft's own guidance puts at 4 Mbps upload per HD endpoint [3], one person running a Claude session with image inputs, the card terminal heartbeating in the background. On a single consumer-grade line that's saturated before lunch. The Claude session is the visible casualty because it's the slowest workflow to recover.

The fourth is the customer who messages us at 11pm to say our broadband's killing their AI tools. We run the line. Line's fine. The router is an eight-year-old consumer unit shoved on the back wall and the AI workstation is at the front of the shop on 2.4 GHz wifi through two breeze blocks. Sometimes the problem is the last 15 metres inside the building.

The first two are line problems. The third is a capacity problem. The fourth is a kit problem. All four get reported to us as AI problems. None of them are.

Why nobody else is saying this

Every AI piece you've read about small business has been written by someone whose business is selling AI, integrating AI, or training people on AI. None of them have a commercial reason to point at the broadband line. Their gate is the model, the prompt, the workshop, the consultancy retainer. The conversation they want is the conversation that ends with a quote attached.

We're an ISP. We sell the pipe AI runs on. So we have a commercial reason to point this out, and we should say so, because pretending we don't have an interest would be the dishonest move. The reader can weigh that for themselves.

Here's why we're still the right people to say it. We're the only category in this conversation whose interest aligns with yours on this specific question. The vendor wants you on the model. The consultant wants you in the workshop. We want you on a line that can actually run the model. If the line's fine for what you're doing, we'll tell you. If it isn't, we'll tell you. Every other AI thinkpiece is written by people selling AI. We sell the pipe. That difference is the whole article.

The fix is boring

The fix is boring, which is why it works.

The exciting questions in AI right now are which agent framework, which model, which prompt pattern, which integration. The boring question is whether your line can run the answer. The boring question is the one the businesses winning at AI are quietly answering first. They aren't posting about it on LinkedIn, because it doesn't make for a sharable graphic. They've just got 30 Mbps symmetric upload, a sub-20 ms latency line, a router that isn't a decade old, and a backup connection for the workflows that matter.

The shape of the boring question is upload speed, latency, packet loss, and what happens at 3pm on a Tuesday when half the office is on a call. The concrete tick-list version is the next guide in this cluster. This piece is the argument for why you'd bother doing it. The businesses that win AI in the next two years won't be the ones who picked the cleverest model. They'll be the ones who paid attention to the foundation first. If you want to talk to us about whether your foundation's there, our business broadband is the page to read next.

Frequently asked questions

Isn't this just a broadband sales pitch dressed up as an opinion piece?

Fair challenge. We're an ISP, so yes, we have a commercial interest in this argument being true. The alignment cuts both ways. If your line is fine for what you're running, we'll say so. Most of the connectivity advice in this piece would apply if you bought your broadband from anyone. The argument we won't make is that model choice is the problem, when the data we see from the ISP seat says it usually isn't.

So what broadband do I actually need to run AI tools properly?

The honest short answer is full fibre (FTTP) on a business-grade product with a service level agreement (SLA) on uptime, 30 Mbps symmetric upload as the floor, sub-20 ms latency, and a backup connection if AI sits on the critical path. The longer version with self-checks you can run on your own line is in the next guide in this cluster.

What if my postcode can't get full fibre yet?

Most can. UK full fibre coverage crossed 78% of premises in 2025 [4]. If yours isn't there yet, a 4G or 5G alternative for business is worth costing up, especially in the gaps where the rollout hasn't reached. We're happy to talk you through it without trying to sell you something that doesn't work at your address.

Is the model I pick really irrelevant?

No. Pick the tool that does the workflow you actually need. The point of this piece isn't that model choice doesn't matter. It's that model choice isn't the gate. The gate is whether the line under your desk can run the tool reliably enough that you keep using it past the first week.

Sources

1. Mole Valley Chamber of Commerce, UK SME AI Adoption Report 2026. 2. Anthropic, Introducing Claude for Small Business, 13 May 2026. 3. Microsoft, Prepare your organization's network for Microsoft Teams. 4. Ofcom, Connected Nations 2025.

A note from our founder

"We use AI to enable and empower our team. That's a deliberate choice, because Inspire is a human-first company and AI doesn't change that. We also know the businesses ignoring AI will get left behind, so we use it to make our people faster. The aim is to do both well."

Ray Bennett, Founder of Inspire Telecom

A note on what this is and isn't. This guide is general information for UK small business owners exploring AI tools. It isn't legal, tax, or financial advice. The tools named here are third-party products with their own terms, pricing, and limits, and we don't control their outputs. For any decision with real money, contracts, or HMRC attached, run it past your accountant, solicitor, or the relevant qualified professional before you act on what an AI has told you. Your name's on the document, not ours and not the AI's.