For a small business, artificial intelligence can look like something for large corporations with big budgets and dedicated departments. In reality many tools now cost about as much as an ordinary software subscription and run in the browser. The risk, if anything, is the opposite: trying everything, finishing nothing and ending up with company data scattered across different services. What you need is an orderly path, small but concrete.
Start from problems, not tools
The most common mistake is to ask "which AI should I use?". The right question is: which activities waste the most time every week and consist mainly of text, documents or repetitive data? Those are where generative AI and automation pay off most.
A useful exercise: for one week, everyone notes the repetitive tasks they perform and how long they take. At the end you will have a concrete list to reason about instead of impressions.
Use cases that often work
Here are some applications that, in small organisations, deliver results quickly.
- Drafting communications. Replies to frequent customer requests, polite payment reminders, notices of changed opening hours. A person reviews and sends.
- Summarising documents. Minutes, specifications, supply terms: a first summary helps you see where to focus your reading.
- Product descriptions and data sheets, starting from data already held in your management software.
- Working translations of emails and documents for foreign customers or suppliers, with review for important texts.
- Spreadsheet analysis. Asking it to identify customers who have not ordered for months or products with shrinking margins.
- Transcribing meetings and calls with a list of action items, provided all participants have been informed.
- Help writing internal procedures, for example how to open a case file or handle a return.
Example: an installation company
A five-person company receives quote requests by email every day, often incomplete. With an AI assistant you can prepare, in a few seconds, a reply that thanks the customer, recaps what they wrote and lists the missing information (measurements, photos, address). The technician checks it and sends it. The gain on a single email is not spectacular, but over a hundred requests a month it becomes real time.
What it really costs
| Item | What to consider |
|---|---|
| Subscriptions | Per-user plans; business versions often offer stronger data protections |
| Learning time | A few hours per person in the first weeks |
| Integrations | Connecting AI to management software or email may require a consultant |
| Review | The time needed to check results must not be forgotten |
An honest calculation compares the time saved with the overall cost, including checking. If a task takes longer to verify the AI's work than to do by hand, it is not a good candidate.
Internal rules before you start
Even in a three-person business it is worth writing down a few clear rules:
- Which tools are authorised. Avoid everyone using a different service with their own personal account.
- Which data is never entered. For example health data, identity documents, credentials and confidential customer information, unless the tools come with adequate contractual guarantees.
- Who is responsible for the result. Text produced with AI is the responsibility of whoever sends it.
- When to disclose the use of AI, for example if a customer is interacting with an automated assistant rather than a person.
The GDPR still applies in full: if you enter personal data into an external service, that service effectively becomes a supplier processing data on your behalf and must be chosen with care.
A four-step plan
- Pick a single process from those identified, frequent and low-risk.
- Trial it for a month with one or two people, measuring time before and after.
- Collect successful examples: prompts that work, reply templates, mistakes to avoid. They become your internal "manual".
- Expand gradually to other processes and colleagues, only if the first experiment produced measurable results.
Training: the deciding factor
The same tool, in different hands, produces very different results. A few hours of hands-on training are enough to teach people to write clear requests, recognise typical errors and understand when AI is not the right fit. It is also the best way to reduce fears: those who understand how the tool works see it as a help rather than a threat.
When it is not worth it
AI is not the right answer for rare processes, for decisions that require full professional responsibility with no way of checking, or when the real problem is organisational. If information is scattered across ten different places, no assistant will put it in order by itself: the process has to be fixed first.
In short
- Start from the repetitive tasks that waste time, not from choosing a tool.
- Drafts, summaries, translations and spreadsheet analysis are quick-win use cases.
- Set written rules on tools, permitted data and responsibility.
- Experiment on a single process, measure the results and only then expand.
- Training your people matters as much as the tool you choose.
