Quick answer: Start with one job you already do, not with a tool. Pick something that comes round every week and where you can tell a good result from a bad one at a glance, then do it yourself with an AI assistant, on your own material, until you trust it. Write down what worked, check every answer before it leaves your desk, and keep personal and confidential data out. Only once the job works reliably by hand is it worth automating, and if it never does, that is a useful answer too.
Where should a business start with AI?
With a job, not a product. Choosing software before you know what it is for is like buying a van before you know what you are delivering.
If you are not sure where to begin, you are in good company. When the Office for National Statistics looked at what stops or delays UK businesses adopting AI, "difficulty identifying business use cases" was among the most commonly reported reasons, alongside cost and a lack of expertise (ONS, Artificial intelligence in UK businesses: 2023 to 2026, released 20 July 2026).
The same release found 28% of businesses with 0 to 9 employees reporting use of at least one AI technology, against 49% of those with 250 or more. Plenty of small firms have not started either, and there is no deadline. The real question is whether there is a job in your week that AI would genuinely help with.
How do you choose your first AI task?
Choose a job that comes round often, that you can check quickly yourself, that needs no personal data, and where a mistake is caught before it reaches a customer. Four questions. A task that fails any one of them can wait.
How some common candidates fare, judged under SBW Labs' method (judgements, not measured results):
| Candidate job | Good first task? | Why |
|---|---|---|
| Drafting replies to the questions customers ask most | Yes | It comes round weekly, you know a good reply on sight, and you read it before it is sent |
| Rewriting a long service description into a shorter one | Yes | The material is already public and easy to judge |
| Turning your own meeting notes into a list of actions | Yes, with care | Easy to check against your memory, once names and anything sensitive are taken out |
| A chatbot answering customers on your website | Not first | Wrong answers reach customers with nobody reading them first |
| Anything built on customer, patient or staff records | Not yet | Needs a decision about tools, terms and data protection before any experiment |
Notice the pattern. The good first tasks are drafting and summarising, on material you can see and check. That is deliberate: confidence comes from checking, and you can only check what you can read.
What are the steps from a first task to automation?
SBW Labs works in a fixed order: choose one useful task, check it suits AI, practise it by hand, write a working guide, and only then automate it, if you want to. It is our own method, the same five steps behind our practical AI training, and the order matters more than any tool.
- Choose one useful task. Something you or your team already do, that takes time, and where you know a good result when you see one.
- Check it suits AI. Whether AI helps with it at all, what data and tools it would need, and what could go wrong.
- Practise it by hand. Do the task with an AI assistant, on your own material, until good and bad answers are easy to tell apart. Before each run, jot down two or three lines on what a good result must contain, so a fluent answer cannot talk you round.
- Write the working guide. The prompts, checks and steps that worked, on one page someone else in the business could follow.
- Automate it, if you want to. A workflow in your own tools, with checks on what it produces. That is a separate piece of work, and plenty of tasks never need it.
Why not jump straight to step five? Because you cannot check an automation you have never done by hand. The practice teaches you what a wrong answer looks like, and the guide is what any automation is built from. Automate first and you have a machine producing work nobody knows how to judge.
What should you never put into an AI tool?
Keep out anything you would not be comfortable seeing in public, and keep out personal data until you have checked the terms of the exact tool and plan you use.
The National Cyber Security Centre recommends "not to include sensitive information in queries to public LLMs" and "not to submit queries to public LLMs that would lead to issues were they made public", and says the terms of use and privacy policy "need to be thoroughly understood before asking sensitive questions" (NCSC, ChatGPT and large language models: what's the risk?, 14 March 2023).
The Information Commissioner's Office is just as plain: "There is no 'AI exemption' to data protection law", and "if an organisation is processing personal data, then data protection law will be applicable" (ICO, Tackling misconceptions). Pasting a customer's details into a chatbot is processing their data, with all the usual duties attached.
For a first task, that boils down to this:
| Keep it out | Usually fine |
|---|---|
| Names, contact details and records of customers, patients or staff | Text you have already published, such as your website or a brochure |
| Health information and other sensitive personal data | A template or example with the details stripped out |
| Passwords, access keys and bank details | Notes on how a job is done, with no names in them |
| Contracts, unannounced prices and anything under a confidentiality agreement | General questions about how to do something |
If the task you most want to try fails the left-hand column, that is your cue to decide on tools and terms first, not to experiment.
How do you know the AI is actually helping?
Compare each answer with the few lines you wrote down before you ran it, not with how impressive it sounds. AI can be fluent and wrong in the same sentence, so check names, dates, prices and claims against something you trust.
A rule of thumb worth keeping: if checking the answer takes longer than doing the job yourself, the task is not suited to AI, at least not yet. Stop, pick another, or use something simpler: a template, a checklist or a person. We hold our own work to the same rule: if a task does not suit AI, we say so and recommend the simpler approach. Using AI is not the goal; getting the job done well is.
What else do people ask about starting with AI?
Do you need to buy an AI tool first?
No. Start with a general-purpose AI assistant you already have access to, check its data terms against the table above, and pay for something more only once a task has shown it is worth it.
Do you need an AI strategy before you start?
Not for the first task. A strategy written before anyone has used AI on real work is a guess. Let it grow from what one real job teaches you.
Should you train the whole team at once?
Usually not at first. Get one task working, with its guide, then teach it to the people who do that job. There is more on that in how to learn AI for your business.
Is it too late to start?
No. Among businesses with 10 or more employees, the ONS describes AI adoption to date as "relatively shallow", with the average number of AI technologies per adopting business rising only from around 1.4 to 1.6 since late 2023 (ONS). Starting small, with one job, is a perfectly ordinary place to begin.
Should your first AI project be an automation?
Rarely. Automation is step five for a reason: once a job has been done by hand, written up and checked, an automation is worth scoping.
Want a hand with the first one? Practical AI training is built around one real job from your own business, and you finish with a guide you keep. Rather compare the ways of working first? How we work sets them side by side.