A practical map of the tasks AI can take over in sales, quoting, support, basic bookkeeping and HR, and the ones that stay with people.
When people say “AI in a small business”, many owners picture either a robot replacing staff or a marketing toy. The everyday reality is duller and more useful: language models are good at tasks with a lot of text, loose rules and repetitive volume. They read an email, work out what is being asked, search documents, fill in a form and write a first draft. They are not good at taking responsibility for decisions that carry risk.
Below is a map by department: what can be automated, what stays with a person, and how to pick a first step. We do not promise exact savings, because they depend on how each company works; an article on estimating the payoff comes later in this series.
The basic rule: what is worth automating
A task is a good candidate if it happens often, arrives in a predictable format, produces a result that is easy to check, and a mistake is cheap to fix. If it does not tick at least three of those, start with something else.
Conversely, a rare, high-stakes task with no clear criteria (negotiating an important contract, say) stays with people. AI can prepare the paperwork, but it does not decide.
Sales: from message to CRM
Enquiries arrive through forms, email, phone and messaging apps. An assistant can read the message, pull out the essentials (who, which product, what deadline), enter them in the CRM and suggest a first reply that a person approves. Lead scoring against criteria you set and follow-up reminders belong here too.
What stays with a person: the real conversation with the customer, negotiation, the relationship. A badly scored lead is no tragedy if a person can correct it easily; an email sent automatically, unchecked, in the company’s name can be.
Quotes: a first draft, not the final document
A small firm’s quote usually looks like the previous ones: same structure, different quantities and a different customer. An assistant can assemble a quote from existing templates, the price list and the customer’s requirements, then flag what it could not find. A person checks prices, terms and deadlines before it goes out.
The common mistake is letting the AI “invent” a price or a commercial condition. Prices must come from a controlled source (a spreadsheet, an ERP), not from the model’s memory.
Customer support: answers to repetitive questions
If most messages are the same three questions, that says something about your website rather than your customers, but until you fix it an assistant can answer from your own documents (opening hours, returns, delivery, warranty) and hand the conversation to a person when it does not know or the customer is unhappy. We cover how to do this without damaging your reputation in a separate support article.
Basic bookkeeping: documents, not tax decisions
Here AI helps with the work that happens before the accountant: reading incoming invoices (PDFs, photos), extracting supplier, amounts and dates, matching them to orders and suggesting an expense category. The output is a list ready for review, not a ledger entry made on its own.
Tax treatment, VAT decisions and filings stay with the accountant. Some tasks, such as sending invoices to the tax authority, are better solved with classic integrations than with AI; we cover that in a separate e-invoicing article.
HR: organisation, with care around selection
Answering staff questions about internal rules, drafting job adverts, scheduling interviews and summarising a CV in a uniform format are all suitable tasks. Be careful with selection, though. The EU Artificial Intelligence Act (AI Act) treats systems used for recruitment and candidate selection as high-risk systems with specific obligations. If you plan to let a system rank or reject candidates, talk to a lawyer first.
What not to automate
- Decisions with a significant legal or financial effect on a person.
- Sensitive communications (dismissals, serious complaints, disputes).
- Any task where you cannot check the result.
- Processes you have never described: if the person doing it cannot say exactly what they do, no tool will guess.
How to choose a first step
- For one week, note the repetitive tasks your people do and roughly how long they take.
- Pick a single one, with clear input and an easily checked result.
- Run the automation alongside the person for a few weeks and compare results.
- Only then let it work on its own, with a person spot-checking and a simple way to switch it off.
- Expand gradually, one task at a time.
A hypothetical example: a distribution company receives orders by email in different formats. The first automation reads the email, fills in an order in the system and sends it to a person to confirm. It does not change the process; it only removes the retyping.
One more practical point: every automation needs an owner in the company, someone who watches it, corrects it and knows how to stop it. Without one, even a good automation degrades slowly without anyone noticing.
In short: AI does not replace good processes and does not fix chaotic ones. It works best where the work is already described and a person remains accountable.
If you want to see what would suit your business, ME Digital provides AI implementation for small and medium-sized companies, from identifying tasks to putting them into service; you can reach us through the contact page.