which tasks can a small business realistically automate with AI?
Seven tasks a small or mid-sized business can automate with AI: what each one saves, where a human must review, which data needs care, and three criteria for choosing the first.
Talk of AI automation tends to conjure up large projects. In small and mid-sized businesses the gains usually sit in small tasks repeated every day: the same answer typed to the same question, rows copied between spreadsheets, proposals written from scratch each time. Below are seven tasks that can realistically be automated today, where each needs human review, and which data calls for care.
First, know where AI gets things wrong
Language models produce text by predicting the next likely word. Asked something they do not know, they may give a plausible-sounding answer without admitting it. The errors you will meet:
- They invent information they do not have: prices, stock levels, delivery times, legal clauses.
- They get numbers wrong and can mix up totals, dates and units.
- They do not know your exceptions and cannot apply a rule nobody wrote down.
- They state a wrong answer as confidently as a right one, so errors do not stand out.
The rule is simple: AI drafts, classifies and summarises. A person approves anything binding, such as prices, commitments, legal wording and the final text a customer receives.
Tasks to automate in customer communication
Answering frequently asked questions (website and WhatsApp assistant)
What it saves: questions about opening hours, delivery, returns and appointments are answered at once, out of hours too, and the team stops retyping the same reply.
Human review: the assistant should answer only from content you supply (FAQs, delivery and returns policy, service list) and hand over to staff when it cannot find the answer. Complaints, refund requests and price negotiation go straight to a person. Read the conversation logs in the first weeks; wrong answers show up there.
Data: tell users clearly that they are talking to an AI assistant. It should never ask for ID numbers, health or card details.
Form and email replies
What it saves: incoming messages sorted by subject (quote request, support, complaint), routed to the right person, with a draft reply waiting.
Human review: the draft should not send itself. Staff read, correct and send it. Only a short acknowledgement of receipt goes out automatically.
Data: names, phone numbers and email addresses from a form are personal data; cover this processing in your privacy notice.
Tasks to automate in content and documents
Proposal and report drafts
What it saves: a first proposal draft built from meeting notes and your template, and monthly report data turned into prose.
Human review: price, scope, delivery dates and payment terms are not left to the model; they should come fixed from your price list. Check every figure in a report against its source.
Data: proposals contain your client's commercial information. Check any confidentiality agreement before sending it to a third-party service.
Product descriptions
What it saves: consistent, readable descriptions for hundreds of products from a specification table, plus other language versions.
Human review: the model may add a feature it was never given, such as fabric composition, dimensions, warranty period or a waterproofing claim. Technical details should come from the data, with the model only writing the sentence. Check every field that makes a claim before publishing; in content production with AI this step is what decides quality.
Data: usually no personal data, so this is among the lowest-risk tasks.
Summarising reviews
What it saves: hundreds of customer reviews grouped into themes: the most praised feature, the most frequent complaint, a recurring delivery or sizing problem.
Human review: the summary is an input to a decision and cannot be the decision. A model can let a rare but critical complaint, a safety issue for example, dissolve into a generally positive picture. Read the reviews behind the summary before you act.
Data: strip out names, phone numbers and order numbers before processing.
Tasks to automate in data and internal operations
Moving data between spreadsheets and the CRM
What it saves: information from forms, emails or PDFs reaches the spreadsheet or CRM without manual copying. Most of this is conventional automation; AI is needed only to pull fields such as company name, request and date out of free text.
Human review: a record with an empty required field or the wrong format goes to an approval queue. Do not hand deletion or bulk updates to automation.
Data: customer records are personal data. Give the automation access only to the fields it needs.
Meeting notes
What it saves: a summary of the recording, the decisions taken and who does what by when.
Human review: figures can be misheard and remarks misattributed. Do not circulate the summary until an attendee has approved it.
Data: a voice recording is personal data; tell all participants beforehand and obtain their consent.
Personal data and KVKK: what to watch
KVKK is Turkey's personal data protection law; GDPR sets similar duties in the EU and UK. This section is not legal advice, so confirm your position with a data protection adviser. In practice:
- Minimise data. Do not send the model personal data the task does not need; names, phone numbers and ID numbers can be masked for most jobs.
- Keep special category data such as health, biometrics, religion and trade union membership out of general-purpose AI tools.
- Find out where your provider processes data, how long it keeps it and whether it uses it to train models. Servers abroad mean a cross-border data transfer with obligations of its own.
- Update your privacy notice, and explicit consent processes where required, to reflect the new processing.
- Set a written rule that stops staff pasting customer data into tools on their personal accounts.
Which task should you start with?
Judge a first candidate against three criteria:
| Criterion | Question to ask | A good candidate answers |
|---|---|---|
| Repeats often | How many times a week is this done? | Every day, in a similar form |
| Has clear rules | Could we explain it to a new starter on one page? | Yes, with few exceptions |
| Low cost of error | What happens if the output is wrong? | Easy to correct, touches neither money nor the customer |
A task that meets all three is your first candidate. In most businesses that is internal work someone already looks over, such as sorting incoming messages or meeting notes. A customer-facing assistant is more visible, but its cost of error is higher; it suits a second step.
A workable order:
- For one week, write down the team's repetitive tasks and roughly how long each takes.
- Score the list against the three criteria and pick one task.
- Define what a correct output looks like using real examples.
- At first, have a person approve every output and log the errors.
- If errors fall to an acceptable level, widen the scope. If not, keep that task out of automation.
If you cannot explain a task in writing to a new member of staff, you cannot explain it to an AI either. Make the process itself clear before automating it.
Connecting the chosen task to your existing systems (website, email, spreadsheet, CRM) is the job of AI integration and automation. For the team to use these tools safely, a short training session often comes before any software.
To assess together which of your tasks suits automation, write to us through the contact page with a list of your repetitive work.
