Starting point
Where do I start with AI in my business?
The short answer
Start with one workflow your team already repeats every week, where the input is text and the output is a draft a person approves. Build that one thing properly in about six weeks rather than running five pilots, and you will know within the first week whether AI helps you or not.
Step 1: find the work that repeats
For a week, ask two or three people to note anything they do more than five times and would not miss. You are looking for reading, matching, summarising, drafting and chasing. Almost every SME finds the same shortlist: inbox triage, quote and invoice handling, answering the same customer questions, pulling numbers together for a report.
Rank the shortlist by hours per week times how similar each instance is. The top line is your first build. It is usually unglamorous, and that is the point.
Step 2: prove it in a week, on your own data
A prototype on your real documents settles arguments that a slide deck cannot. It shows you what the tool gets right, where it fails, and how long a human takes to check it. If the checking costs more than the doing, you have learned something cheaply and you stop.
Do not buy a platform at this stage. Week one is about whether the shape of the answer is right.
Step 3: put a person in the approval loop
The tool reads, matches and drafts. A person decides. Nothing sends, signs or spends on its own. That single rule removes most of the risk people worry about, and it is also what the EU AI Act expects of you as a deployer.
Log every decision the tool made and every override a human applied. That log is your evidence later, and your best source of improvements.
Step 4: train the people who will use it, while it is being built
Tools handed over cold get abandoned. The people who will use it daily should be testing it on real work in weeks three and four, so by handover it is already theirs. Article 4 of the EU AI Act also requires demonstrable AI literacy for staff using these systems, so the training earns its keep twice.
Step 5: only then decide what is next
One workflow live tells you more about your second build than any roadmap written beforehand. You will know your data is messier than you thought in one place and cleaner in another, and which team is keen. Sequence the rest from evidence.
Common questions
- Do I need a data strategy before I start with AI?
- No. You need one workflow and access to the documents or messages it already uses. A full data strategy is what a first build teaches you, not a prerequisite for it.
- Which model should I choose?
- It is almost never the deciding factor. The value sits in the workflow around the model: what it reads, what it may do alone, what a human signs off, and what gets logged. Models change every few months; the workflow design outlasts them.
- How do I know a workflow is a good first candidate?
- It happens many times a week, the steps are broadly the same each time, the inputs are already written down somewhere, and a person can check the output in under a minute.
- What should I not start with?
- Anything that sends, signs or spends without a human, anything touching special-category personal data, and anything where nobody internally can own the tool once it exists.
- How long until something is actually running?
- A working prototype on your own data inside a week, and one workflow live in your business in about six weeks. If the first week shows it will not pay back, stopping there is the right answer.
Next step
A 20-minute call is enough to tell you whether there is a workflow here worth building, and to say so plainly if there is not.