AI can do more. Your business should still start small.

The news, briefly

DeepSeek has released DeepInfer, open-source software for running its models on specialist hardware. Meanwhile, a DeepMind paper published on 15 October—not a new release today—describes four AI agents completing Minecraft over more than 27 hours. One development opens up infrastructure; the other demonstrates sustained, multi-step work in a game.

Our take: capability is not the same as business value

Our interpretation is that the conversation should move from what AI can demonstrate to what a business can use reliably. Open software gives technical teams more options, and long-running agents suggest possibilities beyond a single question and answer. Neither development tells a shop owner what to buy tomorrow.

For an SME, the meaningful unit is a completed, checked job: a usable supplier enquiry, an accurate booking draft, or a useful stock summary. A system that appears impressive but needs extensive corrections may not improve that job. These announcements are reasons to explore carefully, not evidence that every business should hand over whole processes. We would judge progress by the quality, time and cost of the final result, including the human checking.

What this could mean for a smaller business

Consider an illustrative small hotel, not a reported client result. Staff receive guest enquiries containing dates, room preferences and special requests. AI could turn those messages into a structured draft for a receptionist to check against the original. The possible benefit is less repetitive copying; the important limitation is that dates, prices and availability still need verification.

That is a much smaller commitment than letting an agent confirm rooms, change prices or take payment. Keep those decisions with staff during a trial. The Minecraft result does not predict how often a booking assistant will make mistakes, and it cannot establish savings for this hotel. DeepInfer is also infrastructure software, not a ready-made hotel application. Most owners can evaluate an existing tool without buying specialist servers or building their own hosting system.

A useful two-week test

Choose one internal drafting task and keep the usual process available as a comparison. For the hotel example, use anonymised past enquiries rather than live guest details. Have a staff member check every AI draft before it goes anywhere. Do not connect the experiment to payments or automatic booking confirmation.

Over two weeks, record minutes per completed task, corrections per draft, and tool cost per approved result. Include review time, not just generation time. Compare those figures with the current manual process on similar enquiries. Stop if checking and rewriting takes longer than doing the job normally, or if you cannot prevent unchecked drafts from reaching customers. If there is no clear improvement, waiting is a reasonable outcome—not a failed digital transformation.

What could change our view

The research demonstrates one game run, not dependable performance in every business setting. DeepSeek's documentation describes specialist hardware and additional components needed for a complete service; open code alone does not remove operating or maintenance costs. We should not assume cheaper software means a cheaper finished workflow.

We would become more confident as repeatable tests show useful results on ordinary tasks, with the checking effort and full costs included. Until then, protect customer information, keep human approval where mistakes matter, and ask vendors to demonstrate the workflow you actually need. Bigger capabilities are interesting. A small improvement you can verify is more useful.