AI
AI agents, explained: what they can actually do for a small business today
Strip away the demos and the funding froth, and AI agents genuinely do about six things well for a small business as of early 2026. Here they are — along with the failure modes nobody puts in the sales deck.
I spend my working week installing automation in small and mid-sized firms, which gives me a useful immunity to demo videos. So let's do this properly: what is an AI agent, what will one actually do for a business of five to fifty people, and where does it fall over?
First, the definition — because it matters
A chatbot answers you. An agent acts for you. Technically, an agent is a language model wired up to tools — email, a browser, your accounts package, a calendar — and allowed to take a sequence of steps towards a goal without asking permission at every turn. "Find the three unpaid invoices over thirty days, draft polite chasers in our tone, and queue them for my approval" is agent work. The approval step at the end is not decoration; it is the load-bearing wall.
The six jobs agents do well right now
1. Research and triage. Reading fifty supplier websites, comparing prices, summarising what changed in a regulation, sifting an inbox into act-now, delegate and ignore. Agents are tireless and fast, and mistakes here are cheap because you review the summary, not the raw material.
2. Document processing. Pulling data out of PDFs, invoices, delivery notes and contracts into a spreadsheet or your accounting system. This was the dullest job in your office; it's now the most automatable. Accuracy is high on clean documents and degrades on photographed crumpled receipts, exactly as you'd expect.
3. Customer-service drafting. An agent that reads the incoming query, checks your FAQ and order system, and drafts a reply for a human to approve will absorb a startling share of routine tickets. Fully autonomous replies are still a mistake for most small firms — one confidently wrong answer to a customer costs more than a year of saved minutes.
4. Bookkeeping preparation. Categorising transactions, matching receipts to card statements, flagging anomalies for your accountant. The agent does the sorting; the human does the judgement. Your accountant will not miss the shoebox.
5. Meeting follow-ups. Transcribe, summarise, extract actions, draft the follow-up email, create the tasks. This one is close to a solved problem, and much of it can now run on-device — see our piece on small local models for why that matters for client confidentiality.
6. Repetitive web tasks. Form-filling, listing updates, price checks across marketplaces. Browser-driving agents remain the flakiest of the six — websites change under them — but for tolerant, checkable tasks they already pay their way.
The pattern across all six: the agent does the volume, a human does the judgement.
Where it goes wrong
Three failure modes recur. First, compounding errors: an agent that is 95% reliable per step is only about 77% reliable across five steps, which is why long unsupervised chains disappoint. Second, confident nonsense: agents inherit their models' habit of asserting things that are not so, and an agent with tools can act on its hallucination. Third, brittle integrations: the agent is only as good as its access, and half of small-business "agent projects" die in the plumbing between the model and a ten-year-old CRM.
None of these is fatal. All of them argue for the same design: short chains, narrow scopes, human sign-off on anything that touches money, customers or the law.
What it costs, and how to start
As of early 2026, useful agent capability arrives three ways: built into software you already pay for (check what your accounts and helpdesk vendors have shipped this year before buying anything new), via off-the-shelf agent products typically priced per task or per usage rather than per seat — a shift we've analysed separately — or custom-built, which rarely makes sense below a few hundred hours of saved labour a year.
Start with one process that is frequent, boring, checkable and low-stakes. Run the agent alongside the human for a month. Measure. Only then expand. The firms getting real value are the ones treating this as workflow redesign rather than software shopping — a theme that runs through what "AI-first" companies actually change.
The honest bottom line
Agents in early 2026 are roughly where good spreadsheet macros were in the 1990s: transformative for the tedious middle of office work, dangerous when treated as infallible, and destined to become invisible plumbing within a decade. Ignore the people telling you they'll run your company. Ignore equally the people telling you it's all hype. Six real jobs, done with supervision, is not hype. It's just work — done by something that never gets bored.