Every organisation is being told it needs an AI strategy. Most do not need a strategy; they need two or three uses of AI that save real time this year, built safely. Here is where we see AI working today, and where it is still the wrong tool.
Where AI works today
Reading documents
Invoices, receipts, delivery notes, application forms and identity documents arrive in every shape. AI models can now read them and turn them into structured data: supplier, date, items, totals, names and numbers. A person reviews what the model is unsure about instead of typing everything. For finance, procurement and onboarding teams this is often the fastest win.
Sorting what comes in
Emails, enquiries, complaints and support requests can be read as they arrive, sorted by type and urgency, and sent to the right team. Spam and sales pitches are separated from genuine requests. Nothing is decided automatically; the queue simply arrives organised.
Answering from your own records
An assistant connected to your policies, product catalogue, procedures and records can answer staff and customer questions in seconds, and show where the answer came from. The key word is connected: the assistant answers from your information, not from whatever it learned on the internet. When it does not know, it says so and hands over to a person.
Summarising and drafting
Long case files, meeting notes, reports and contracts can be summarised in a paragraph, and first drafts of letters and replies written in your style. People stay responsible for what is sent; they simply start from a draft instead of a blank page.
Seeing patterns in your data
Models can flag unusual transactions, predict which loans or invoices are likely to be late, and forecast demand from your sales history. These work best as early warnings that a person reviews, not as automatic decisions.
Where AI is the wrong tool
- Where a clear rule exists. If the rule is “orders above a set value need a manager’s approval”, write the rule. Rules are predictable, cheap and auditable.
- Where a mistake is serious and cannot be checked. Medical, legal and credit decisions need a qualified person to decide, with AI as an assistant at most.
- Where the data is not there. AI cannot answer from records that live on paper or in someone’s memory. Sometimes the first AI project is getting the data into a system.
Five rules for using AI safely
- Ground it in your data. Connect the model to your own documents and records, and make it show its sources.
- Keep a person in the loop wherever a decision affects money, health, safety or someone’s rights.
- Send only what is needed. Give a model the minimum information required for the task, and keep sensitive records inside your systems.
- Choose providers carefully. Use models whose providers commit, in writing, not to train on your data. Know where the processing happens.
- Measure accuracy before you rely on it. Test with a few hundred real examples, in the languages you use, and keep measuring after launch.
How to start
Pick one task that is high in volume and low in risk, such as reading supplier invoices or sorting incoming enquiries. Measure how it is done today, build the AI version with a person checking its work, and compare after a month. If it works, widen it; if it does not, you have lost weeks, not years.
We build AI into the systems we deliver, always with these rules. If you have a task in mind, tell us about it and we will tell you honestly whether AI is the right tool for it.