Most of the teams we work with are curious about AI and cautious about it in equal measure. They've seen what the assistants can do, and they've also heard about made-up answers, leaked data and bills that grow without warning. Both reactions are reasonable.
The good news is that "adding AI" isn't one big decision. There are several quite different ways to use it, each with its own costs and risks, and you can start with the smallest. Here are the four we think about, from the most contained to the most ambitious, and then the question that matters as much as any of them: where the AI actually runs.
1. Search that understands the question
Traditional search matches words. Ask "which jobs ran over budget last spring?" and it looks for those words, finds nothing useful, and gives up. AI-assisted search works out what you mean, turns it into a proper query against your data, and shows you the records it found.
It's a good first step because it only reads. Nothing is changed, the answers are records you can click into rather than paragraphs to trust, and if it misunderstands you, you can see that it has.
2. AI inside the workflow
A surprising amount of office work is reading one document and typing what it says into another: a delivery note, a referral letter, an invoice, an emailed order. AI is very good at that reading. It can pull the details out of a document, check them against your data, and present a completed record for someone to confirm.
The important word is confirm. The person stays in charge, the system records what the AI suggested and what was accepted, and the checks your system already applies to anything typed in still apply. Done this way, AI takes the drudgery out of a process without taking the judgement out of it.
Let the AI do the reading and the typing. Keep the deciding with people.
3. Connecting your system to the assistant you already use
Instead of building AI into your application, you can do the reverse: let the AI assistant your team already uses, such as Claude or Copilot, work with your application. A standard called the Model Context Protocol (MCP) lets a system describe what it can do, such as "find a customer" or "draft an order", so that an assistant can use those abilities on your behalf, with the same permissions as the person asking.
We use this ourselves. Our own content platform has an MCP connection, and much of this blog was put together through it: an assistant drafting pages and making changes, every one of them logged, reversible and limited to the sites it's allowed to touch. It's the most powerful of the four options, and it's also the one that most needs careful design: what the assistant may read, what it may change, and how every action is recorded.
4. Where the AI runs
Whichever of these you choose, there's a second decision that matters just as much.
- A cloud AI service. The most capable models are run by providers such as Anthropic, OpenAI and Microsoft. You pay per use, quality is excellent and there's nothing to install. Your data is sent to the provider to be processed, so the terms matter: business agreements commonly rule out training on your data, and some providers can keep processing within the UK or EU.
- A model on your own machines. Smaller "open" models can run on your own servers, or even a well-equipped PC. Your data never leaves the building, and the cost is the hardware rather than a bill per request. The trade-off is capability: they're very good at narrow, well-defined jobs like reading a standard form, and less good at open-ended reasoning.
For sensitive information, such as patient or research data, the second option, or a cloud service with the right agreements in place, is often the deciding factor. For everyday business documents, a cloud service is usually the simplest and best choice.
Keeping costs and risks in proportion
- Start where mistakes are cheap. Search, then suggestions a person confirms, before anything that acts on its own.
- Use it where it earns its keep. A request to a cloud model costs a fraction of a penny, but thousands a day add up. Design so the AI runs when it's useful, not on every page view.
- Keep a record. Know what the AI suggested, who accepted it, and when.
- Make it optional. The system should work perfectly well without it, so that a change of provider, price or policy is an inconvenience rather than a crisis.
AI won't replace a well-designed system, but it can make a good one noticeably easier to use. If you're wondering where it might help your team, or whether it should at all, we're happy to talk it through, with no obligation to add any.