How do you implement AI in a small business?
Start small and specific. Don't begin with "we need an AI strategy". Begin with one job that is slow, repeated and annoying, and make that job better. Then do the next one.
AI suits small businesses well because the repeat work is easy to see, and a small team that stops doing it gets more done. You don't need a data scientist, perfect data or a big budget to begin. You need a clear problem and someone who cares about it.
Here are the five steps. Do them in order.
Step 1: How do you pick the right problem for AI?
Look for work with four traits:
- It happens often, weekly or daily.
- It follows a similar pattern each time.
- It eats real hours from people who could do better things.
- A wrong answer is fixable, because a person checks the result.
Ask your team one question: "What do you do every week that you wish you didn't?" Write down every answer. Then pick the one that scores best on the four traits. Don't pick the most exciting idea. Pick the most useful one.
Step 2: How do you map the workflow?
You can't improve what you can't describe. Take your chosen job and write it down on one page. Walk through it with the person who does it, not the person who manages them. Use this template:
- Trigger. What starts the job? An email, a form, a date?
- Steps. What happens, in order, with no skipping?
- People. Who touches it, and who approves it?
- Tools. Which systems, spreadsheets and inboxes are used?
- Output. What does "done" look like, and who receives it?
- Exceptions. When does the usual path break, and what happens then?
The exceptions matter most. They are where an AI build gets complicated, and where a person's judgement usually lives.
Step 3: How do you find and check the data?
Every AI job runs on information. Find where yours lives. Then answer four plain questions:
- Where is it? Which system, folder or inbox?
- Who owns it, and who can give access?
- Is it good enough? Open ten real examples and see if you could do the job from them.
- Is any of it sensitive, such as customer details or financials? That shapes where the AI can run.
Messy data is not a reason to stop. Scattered spreadsheets and scanned PDFs can be used. They just need work first, and it is far better to know that before you get a quote.
Step 4: Who should own an AI project?
One named person. Not a committee and not "the team". The owner doesn't write code. They decide what good looks like, answer questions quickly and approve the work.
Pick someone who knows the job and has the standing to say yes. In our projects, client time is about two hours a week each, mostly approvals. If nobody can spare that, the project is not ready yet.
Step 5: How do you set a success measure for AI?
Decide the test before you build, and write it down. A good measure is plain and uses your own numbers:
- Time taken by the job, before and after.
- How many go out without being fixed by hand.
- How fast a customer gets a reply.
- Whether the team would choose to keep using it.
Measure the current state this week, even roughly. Without a "before", nobody can say whether the "after" is better.
What does this look like in a real small business?
Imagine a small trades business where quotes go out days late. Walk it through the steps.
- Problem. Quotes are frequent, follow a pattern and take the owner's evenings.
- Workflow. A customer emails photos and a request. Someone reads it, checks past jobs, prices it, writes the quote and sends it.
- Data. Past quotes in a folder, a price list in a spreadsheet, customer emails in the inbox. Ten past quotes are enough to see if the pattern holds.
- Owner. The person who writes quotes now. They approve every draft.
- Measure. Time from enquiry to quote sent, and how many drafts go out without edits.
That is a clear brief. A partner can read it in five minutes and say what is possible.
How do you know if you are ready?
| Question | Ready | Not yet |
|---|---|---|
| Is the problem one job? | You can name it in a sentence | "We want AI everywhere" |
| Can you describe the steps? | One page, agreed by the person who does it | Each person does it differently |
| Do you know where the data is? | You can point to it | Nobody is sure who holds it |
| Is there an owner? | Named, with time to approve | "The team will sort it out" |
| Do you have a measure? | A before number, even a rough one | "We'll know it when we see it" |
Any "not yet" is a task for this week, not a reason to wait.
What should you bring to a first conversation with an AI partner?
Bring the one-page workflow, a few real examples of the input and the output, a list of the systems involved, the name of your owner and your success measure. A partner can give you a sensible plan from that, instead of a vague proposal. You'll find the same preparation useful for an AI discovery workshop.
What should you not do before you start?
- Don't buy tools first and look for problems after.
- Don't clean every dataset you own. Clean only what the chosen job needs.
- Don't start with a customer-facing job if a safer internal one exists.
- Don't skip the owner. Projects without one drift.
When you're ready, read how to choose an AI implementation partner. If your pilot stalls after the first demo, from AI pilot to production covers what comes next. For how we run the work, see how we work, or who we help.