Self-hosted AI for UK small businesses — without becoming a DevOps team

A few years ago, "self-hosted AI" meant a Linux server, a CUDA-capable GPU, a working knowledge of Docker, and a week of YAML files. That's no longer true. This post is the honest take on what self-hosted AI looks like for a UK SMB of 5-50 people in 2026.

What "self-hosted" actually buys you

Data residency. Your conversations live on a machine you control. No third party can see them.

Cost shape. One-off purchase plus optional support, not per-seat-per-month forever.

Resilience to vendor change. If OpenAI or Anthropic raises prices or changes terms, you keep your tool — you just point the configuration at a different provider.

Compliance story. Easier to answer "where does our client data go?" with "to a machine in our own office" than "to several US-based AI vendors".

What it doesn't buy you

No need for any AI provider. Frontier-quality models (Claude, GPT-4) still need an API call to their respective providers. Local-only is possible with Ollama, but the quality is a step below frontier.

No machine maintenance. The Windows box that runs the AI still needs power, updates, and occasional restarts. It's not a server — but it is a piece of always-on infrastructure.

No new-feature drip-feed. SaaS AI tools ship features weekly. Self-hosted gets updates when you update.

What setup actually looks like in 2026

Self-Install: download a Windows .EXE, double-click, answer 6-8 prompts (AI provider, Telegram bot token, each team member's Telegram ID). Setup takes ~20 minutes including downloading an AI model if you pick the local-Ollama path. No YAML, no Docker knowledge required (Docker Desktop installs itself if missing).

Managed Install: we run the install for you on a 2-hour remote session. Your IT only provides the API keys.

In both cases the result is the same: a Telegram bot the whole team can DM, with conversations stored on a Windows machine in your office.

Hardware requirements

Windows 10 or 11. 8 GB RAM. ~4 GB free disk space. Always-on (so the AI is reachable when the team needs it).

For local Ollama: add ~5 GB disk for a default model, more for larger models. A modern CPU is fine; a GPU is helpful but not required for the default 8B model.

No server-grade hardware needed. A 2-year-old office Windows desktop, repurposed, is plenty.

How AI In Your Pocket compares to rolling your own

You can self-host Open WebUI, LibreChat, or AnythingLLM for free. Many people do. If you have a sysadmin who enjoys this work, that's a perfectly reasonable choice.

AI In Your Pocket bundles the install, 21 UK SMB sector profiles, persistent per-user memory, multi-user Telegram allowlist, and ongoing support. The £1,997 buys you 20 minutes of setup vs an unknown number of evenings of YAML, plus someone to email when something breaks.

For the sector-specific view, see /for/[your-sector] from the homepage. For the setup checklist, see /setup-guide.