What AI agents can (and can't yet) do for a small business
An honest look at where AI agents help a café, a practice or an agency today, where they still fall short, and why a person should approve every step that leaves the company.
- Agents are strongest at analysing, planning and drafting on your own data.
- Anything that leaves your company, like an e-mail to a customer, should be approved by a person.
- Start with one recurring task, check the results, then expand step by step.
What people mean by an AI agent
An AI agent is a language model that does more than answer a single question. It works through a task in several steps: it reads the data you give it, picks out the relevant parts, compares, calculates and writes a result. For a small business, that might mean looking at last quarter's sales and costs and drafting a short analysis, or reading the conditions of a funding programme and listing which documents are still missing.
The important word is task. An agent is useful when the job is clearly described and the data it needs is available. It is far less useful when the goal is vague or the information only exists in someone's head.
Where agents already help
Analysis is the most reliable area. A café owner can ask for a comparison of weekday and weekend sales, a hotel can get a SWOT analysis based on its own figures, an agency can have its client list grouped by revenue and last contact. The agent does not get tired of reading tables, and it explains its reasoning in plain language.
Planning and drafting come second. A go-to-market outline, a first version of a financial plan or business plan, a list of the week's open deadlines, a polite reminder to a customer whose invoice is overdue: these are documents that would otherwise cost you an evening. An agent produces a solid first draft quickly, and you spend your time correcting and deciding instead of starting from a blank page.
Research is the third area. Searching funding programmes, checking basic eligibility criteria against your company profile and building a document checklist is tedious work that agents handle well, as long as you verify the result against the official programme page.
What they can't do (yet)
Agents do not know what they have not been told. If your stock list is outdated, the reorder suggestion will be wrong. If your business plan lacks your prices, a well-built agent asks for them; a poor setup simply guesses. Missing information is the most common reason for weak results.
Agents also make mistakes that sound convincing. A number can be misread, a rule can be summarised too broadly, a funding condition can have changed since the source was written. Treat results as drafts from a capable assistant, not as final answers.
And they are no substitute for judgement or professional advice. An agent can prepare the questions for your tax adviser or lawyer, but it should not replace them. It produces text documents, not polished slide decks or finished designs; the presentation is still your job.
Why human approval matters
There is a big difference between an agent that drafts an e-mail and one that sends it. A wrong draft costs nothing. A sent message with the wrong amount, the wrong tone or the wrong recipient costs trust. That is why every step that leaves your company should pass through a person: e-mails, offers, applications, anything that carries your name.
Approval does not have to be slow. Reading a prepared draft and clicking approve takes a fraction of the time it takes to write it. What matters is that someone who knows the customer and the context has looked at it. Approval also leaves a record of who decided what, which helps when questions come up later.
How to start small
Pick one recurring task that annoys you and has clear inputs. Good candidates are a weekly overview of open invoices, a monthly sales summary or a check of upcoming deadlines. Give the agent the relevant data, look closely at the first results and correct what is off.
Once that task works reliably, add the next one. Keep a short list of what the agent may do on its own (read, analyse, draft) and what always needs your approval (anything sent, signed or submitted). That list is worth more than any general AI strategy.
This is also how we built NEXUM: agents work on your own company data, ask follow-up questions when something is missing, and prepare documents and drafts. A person approves every outbound step, and no e-mail is sent without that approval.