Is your business broadband ready for AI? A 7-point checklist

It's Wednesday morning. The AI receptionist on the front desk has talked over a caller for the third time this week. The file you tried to upload to Claude is still spinning after forty seconds. The Teams call with the client in Glasgow has dropped audio twice. The owner asks the obvious question. Is it the tools, or is it the line?

For UK small businesses that adopted AI in 2024 and 2025, it's almost always the line. Standard fibre to the cabinet (FTTC) was specced for browsing and email, not for AI workloads that hammer the upload channel and demand consistent jitter. The friction you're feeling isn't a personal IT failure, it's structural. Around 560,000 UK businesses, roughly 370,000 of them SMEs, are running on connections that can't cope with modern workloads [1].

This is the companion piece to 5 ways UK small businesses are using Claude right now. That post was the workflows. This one is the connection check that decides whether those workflows feel instant or feel broken.

Seven checks. Each one takes under five minutes. By the end you'll know whether your line is ready, whether you need an upgrade conversation this quarter, or whether you're already losing AI productivity every day.

1. Upload speed: the metric your broadband ad never shows you

A plumbing firm in Croydon uploads a 35 MB PDF survey to Claude for a fixed-price quote. On their 18 Mbps FTTC upload, the file takes 22 seconds before the AI even starts reading. Twelve quotes a week, that's a lost evening. A marketing agency batch-uploading client brand assets to an AI image tool feels the same problem at scale.

Every AI tool runs on someone else's GPU. The prompt and the attached file go up before anything comes back. AI workloads are upload-heavy in a way browsing and email never were, and the headline download speed on the broadband ad doesn't tell you what's about to happen.

What good looks like. Microsoft's bandwidth guidance puts Teams at 2.5 Mbps upload per HD user, with 4 Mbps for best performance, and Copilot stacks on top [2]. A team of ten on a Monday all-hands with Copilot in the meeting, cloud file sync running, and one person uploading a contract can easily need 40 to 60 Mbps of upload combined. Standard FTTC tops out around 20 Mbps up. Fibre to the premises (FTTP) typically delivers 100 Mbps symmetric or more, which is the realistic SME target.

How to test it (2 minutes). Plug a laptop into the router with an ethernet cable, because Wi-Fi adds noise. Go to speed.cloudflare.com. Note the upload number. Repeat at 9am, midday, and 3pm. If your upload sits below 30 Mbps during peak, your line is already throttling your AI workflows.

The failure mode is quiet. Files stall mid-upload, AI prompts time out, Teams calls go choppy the moment someone shares their screen. The line isn't down, it just isn't keeping up.

2. Latency: the millisecond your customer hears as awkward

A dental surgery in Bristol has an AI receptionist taking out-of-hours bookings. The receptionist works, but callers keep saying "sorry, you go" because the AI keeps talking over them. The dental nurse who set it up is convinced the AI model is broken. It isn't. It's the line.

Every conversational AI runs a pipeline: audio goes up, speech-to-text converts it, the language model thinks, text-to-speech generates the reply, audio comes back. Each hop adds milliseconds. The network is the only piece of that pipeline you control, and if you get it wrong the AI sounds like it's on a satellite call.

What good looks like. Voice AI is perceived as instant when end-to-end latency stays under 300ms, and the conversation breaks down above 1,500ms [3]. Your network has to fit inside the budget the AI pipeline leaves it, which means under 50ms on a wired connection, jitter under 15ms, and zero packet loss. FTTP typically delivers sub-20ms idle latency. Decent FTTC sits at 20 to 40ms. Anything riding ADSL or a busy 4G hotspot is already in the danger zone.

How to test it (3 minutes). speed.cloudflare.com gives you both idle latency and loaded latency, which is the latency you feel when the line is busy. Loaded latency above 100ms on a wired connection is an ISP problem, not yours. For voice agents specifically, run packetlosstest.com too. Any packet loss above zero, any jitter above 20ms, and your voice agent is going to sound robotic.

The failure mode is the awkward beat. Callers talk over the AI, streaming Claude responses arrive in chunks instead of flowing, Teams meetings get the frozen-for-two-seconds, then catch-up feeling.

3. Uptime: the SLA that pays you back when the line drops

A small accountancy practice in Sheffield is filing a client VAT return on the 31st. The line drops. Stays down for six hours. Ofcom's Automatic Compensation Scheme covers residential customers; the practice gets nothing. The client gets a late submission. The accountant absorbs the penalty.

AI workloads are constantly chatty. Cloud sync runs in the background, the AI receptionist polls for calls, CRM automations fire on a schedule. A six-hour outage on a residential line is annoying. The same outage on a business running AI customer service is six hours of missed conversations, dropped leads, and broken automations you'll spend half the next day reconciling.

What good looks like. Ofcom's Automatic Compensation Scheme pays £9.33 per day for total loss of service on residential broadband; it explicitly excludes business customers [4]. Businesses need a written service level agreement (SLA) in the contract. Business-grade FTTP typically commits to 99.9% uptime, which is under 8.8 hours of downtime per year. A leased line commits to 99.99%, which is under 53 minutes per year. A standard residential line carries no SLA at all.

How to test it (5 minutes). Find your current broadband contract and search the PDF for "SLA" or "service level". If nothing comes up, you don't have one. Find the target fix time and the compensation-per-day clause while you're there. If neither exists, there's no protection coming when the line goes down.

The failure mode is the four-day fix window. Line drops, ISP gives you "by end of Friday", you eat the lost trading days with no payout to soften it.

4. Capacity headroom: enough lane for everyone at once

A ten-person creative agency in Manchester runs an all-hands on Teams every Monday at 10. Two of the team are also dialled in on AI voice agents handling client intake. By 10:05 the agents are talking over callers and the Teams call is buffering. The IT-savvy partner blames Microsoft. It isn't Microsoft.

AI tools don't queue politely. A voice agent needs constant low-jitter bandwidth or it starts dropping syllables. A Teams call with Copilot active uploads at 2.5 Mbps per user. A document upload to Claude can spike to 50 Mbps for ten seconds. Stack them and a 40 Mbps shared line is already in the red.

What good looks like. Apply Microsoft's per-user numbers to the headcount you actually have, then add 30% headroom for peak bursts. A ten-person office running Teams plus Copilot plus cloud sync needs 40 to 60 Mbps of upload as a baseline, with room for the days when someone's also pushing a 200 MB asset to the cloud. Voice agents need jitter under 15ms continuously, not just on average.

How to test it (5 minutes). Run speed.cloudflare.com at 9am on a Monday with the whole team logged in. If your upload drops more than 30% from your idle test, you're bandwidth-constrained at peak. Count concurrent users x 2.5 Mbps for Teams, add 5 Mbps per Copilot user, add 10 to 20 Mbps for cloud sync. Compare against the peak figure. If you're inside 80% of your line, you've no headroom for growth.

The failure mode is the 10am dip. Voice agents sound robotic at 10am only, calls drop when someone shares a screen, cloud backups crawl to a stop and fall over again next Monday.

5. Static IP: the address your security tools trust

A logistics SME in Leeds whitelists their office IP in their CRM and their payment gateway for security. Tuesday morning, nobody can log in. The router pulled a new IP overnight. The whitelist now points to a stranger's home broadband in Newcastle. Customer service is down until the IT contact gets the new IP added to two SaaS dashboards.

AI SaaS tools, payment gateways, API consoles, and remote access tools increasingly use IP allowlisting as a security control [5]. If your office IP rotates every few weeks, which is what a dynamic IP does, every allowlist breaks silently. The bigger your AI stack, the more silent failure modes you've stacked up.

What good looks like. A single, static IPv4 address assigned to your line that doesn't change. Most business broadband providers either include this in the package or charge a small monthly fee. If you're hosting any customer-facing AI service, taking card payments via an API, or running IP-based VPN access for remote staff, a static IP is the baseline rather than the upgrade.

How to test it (2 minutes). Visit whatismyip.com on a Monday morning. Make a note. Check again on the following Monday. If the IP has changed, you're on a dynamic IP. Then list every SaaS tool your team uses and check its security settings. Any "IP allowlist", "firewall rules", or "trusted IPs" field is a silent failure point waiting for the next rotation.

The failure mode is the unexplained Monday outage. Logins fail, APIs return 403s, the payment gateway blocks access mid-transaction. Each one looks unrelated. The actual root cause is that the IP changed at 3am and nobody told the security tools.

6. Failover: a backup line for the day the cable gets cut

A serviced-office company in Birmingham loses its primary line at 11am on a Thursday. The AI receptionist that takes meeting room bookings goes silent. Forty-three calls land in the void over the next four hours. The fix arrives Friday lunchtime. Three of those leads are gone for good.

When the AI is doing the answering, an outage isn't downtime for staff (who can pick up a phone). It's downtime for customers, who get nothing and don't usually call back. A backup line on a different physical path, whether 4G, 5G, or a separate fibre route, keeps the AI talking while the primary is being fixed.

What good looks like. A failover-capable router with a 4G or 5G SIM as a secondary WAN, configured to cut over automatically when the primary drops. On business-grade kit, failover should happen in well under a minute, often inside a few seconds. Customers don't see the failure, and you get to deal with the fault on a calmer timeline.

How to test it (5 minutes). Phone your IT support or your ISP and ask one question: "If our broadband fails at 10am tomorrow, how long are we offline?" If the answer is "until an engineer arrives" (one to four days), you've no failover. While you're there, ask whether they can supply a 4G backup. Most business ISPs can, very few include it by default.

The failure mode is the silent void. Customer-facing AI tools go quiet for hours. Card payments fail in-store. Cloud CRM access dies until the engineer arrives. You don't get to retry the missed leads, because the customer already went elsewhere.

7. Security posture: the router the AI sits behind

A four-person legal firm in Reading has a consumer-grade router from a residential provider sitting on the front desk. Behind it: a Claude integration handling client intake, a cloud CRM with sensitive case notes, and an AI receptionist taking calls. The router hasn't had a firmware update since it came out of the box in 2022, and every device on the network shares whatever DNS the router happens to pick up.

AI workflows touch sensitive data: client records, payment details, customer conversations, internal case notes. The router is the front door, and one in two UK small businesses suffers a cyber incident every year [6]. Consumer routers lack QoS for prioritising real-time AI traffic, lack DNS filtering, lack VLAN separation to keep customer-facing AI away from finance laptops, and lack an active firmware update policy.

What good looks like. A business-grade router with quality of service (QoS) to prioritise voice and AI streaming. Protective DNS (PDNS) at the resolver to block malware, command-and-control, and phishing domains before they reach the browser; the NCSC explicitly recommends PDNS for the private sector [7]. VLAN separation between staff devices, customer-facing AI, and guest Wi-Fi. An active firmware update policy from whoever supplied the kit.

How to test it (5 minutes). Three questions to your IT supplier (or to yourself, if you're the IT supplier). One: what model of router is on the front desk and when was it last patched? Two: what DNS resolver is the network using and is it filtering? Three: is the customer-facing AI on the same network as the finance laptop? If any answer is "I don't know", you have a security posture problem before you even start on AI.

The failure mode is the breach you find out about from somebody else. A phishing domain gets through the unfiltered DNS, the AI receptionist's client data is exposed, and the breach notification clock starts running.

Score yourself: where you are in 30 seconds

Count the checks you'd pass today. Each item is one point.

- 6 to 7 out of 7. Your line is ready. Your AI workflows will scale as you add tools and headcount, and the only thing holding you back is your prompt library, not your broadband. - 4 to 5 out of 7. Upgrade conversation this quarter. The line works today but you're paying a productivity tax, and as AI usage grows you'll feel it on the worst days first. - 2 to 3 out of 7. You're losing AI productivity every day. Calls drop, files stall, voice agents stutter. You're blaming the tools when you should be looking at the line. - 0 to 1 out of 7. It's the line, not the tool. Whatever AI you switch to next, on this connection, will feel the same as the last one.

If you scored anywhere from 0 to 5, the next step is a proper business-grade line. Inspire's business broadband is the challenger product in this market, built around fixed-price protection, a written SLA, and real UK support.

Frequently asked questions

How fast does my broadband actually need to be for AI tools?

The honest answer is in symmetric upload and low latency, not a single download number. For a single user on text-based AI, 30 Mbps symmetric upload and sub-20ms idle latency is the practical floor. For voice agents, jitter under 15ms is the bar that actually matters. For a team of five or more sharing the line with Teams and Copilot, 100 Mbps symmetric is the realistic minimum.

I have 100 Mbps download. Why are my AI tools still slow?

Download is the wrong metric for AI. The prompt and the file go up before the model can think, so upload speed, loaded latency, and jitter are what decide whether the tool feels instant or feels broken. Run speed.cloudflare.com (covered in check 2) to see your upload and loaded latency. If loaded latency jumps above 100ms when the line is busy, the bottleneck is the line.

Do I really need a static IP if I'm just using ChatGPT and Copilot?

For basic AI use, no. The moment you start using IP allowlisting in any SaaS tool, whether that's your CRM, your payment gateway, or your remote access tool, the answer flips to yes. Check the security settings in each tool you rely on. If any of them has a trusted-IP field, a dynamic IP will break that quietly on a Monday morning.

What's the simplest way to make my broadband AI-ready without ripping it out?

Wire your AI workstation in via ethernet, add a 4G failover SIM, switch on Protective DNS at the router, and ask your ISP for a static IP. Those four moves together close most of the everyday gaps without changing your contract. If your current line still fails the scoring rubric, the next step is a proper FTTP business line. Inspire's business broadband is the challenger product here, and Inspire is the UK's #1 internet provider on Trustpilot from 600+ reviews.

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

Sources

1. IT Pro — Over half a million UK businesses struggle with insufficient bandwidth, February 2025. 2. Microsoft Learn — Prepare your organization's network for Microsoft Teams, bandwidth requirements per user. 3. Hamming AI — Voice AI latency: what's fast, what's slow, how to fix it, January 2026. 4. Ofcom — Automatic Compensation Scheme: residential broadband and landline customers, £9.33 per day for total loss of service, residential only. 5. Microsoft Learn — Microsoft 365 network connectivity principles, guidance on IP-based access controls. 6. NCSC — Small Organisations Guide to Cyber Security, UK National Cyber Security Centre. 7. NCSC — Protective DNS for the private sector, recommended resolver-level filtering.

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.