The short answer
Today, AI agents can handle multi-step admin: sorting your inbox and drafting replies, chasing quotes and invoices, doing first-pass research, updating a CRM, booking and reshuffling meetings, and turning one piece of content into ten. They are good at repetitive, rules-based work with a human checking the important calls, not at unsupervised decisions.
Strip away the hype and an AI agent is just software that can carry out a multi-step task for you rather than only answering questions. For a small business, where everyone already does three jobs, that is genuinely useful. The trick is knowing which jobs to hand over and which to keep a hand on.
Ten real jobs an agent can do today
Where agents earn their keep in a small business
- Triage the inbox. Sort, prioritise and draft replies to routine email for you to approve.
- Chase invoices and quotes. Track what is outstanding and draft polite, firm follow-ups.
- First-pass research. Pull together a briefing on a prospect, a market or a competitor.
- Update the CRM. Log notes, contacts and next steps after meetings so nobody does data entry.
- Manage the diary. Book, reschedule and confirm meetings around your rules.
- Repurpose content. Turn one article into posts, an email and a summary for five channels.
- Prepare quotes and documents. Draft standard proposals and paperwork from your templates.
- Monitor and summarise. Watch a topic or a set of sources and send you a weekly digest.
- Handle first-line questions. Answer common customer queries and escalate the rest to a human.
- Reconcile the routine. Match and flag straightforward records, ready for you to sign off.
What are they not good at yet?
Judgement calls, anything high-stakes, and work where being confidently wrong causes real damage. An agent should not be making final decisions on money, people or legal matters on its own. It also needs clear instructions and clean access to your tools. Point it at a vague goal with the run of everything and it will disappoint you.
How should a small business start?
| Task | How an agent helps | Keep a human on |
|---|---|---|
| Invoice chasing | Drafts and tracks reminders | Sending to key clients |
| Inbox triage | Sorts and drafts replies | Anything sensitive or unusual |
| Research briefs | Gathers and summarises | Checking the facts before use |
| Content repurposing | Produces channel versions | Final tone and sign-off |
Pick one low-stakes task, give the agent narrow access to just what it needs, and keep yourself in the loop on anything that leaves the building. Prove it on that one job before you widen. Gartner expects over 40 percent of agentic projects to be scrapped by 2027, nearly always because someone started too big. Small and useful beats ambitious and abandoned.
Do agents replace staff?
For a small business, the honest answer is usually no. They remove the dull, repetitive edges of jobs so your people spend more time on the work only a person can do: the relationship, the judgement, the awkward phone call that saves the account. Think of an agent as the junior who does the admin, not the person who replaces the team.
How to scope a first agent so it cannot hurt you
The gap between a useful first agent and an expensive lesson is almost entirely about scope. The teams that get this right are boring about it on purpose, and they follow roughly the same progression.
Start read-only. For the first few weeks, let it look at things and tell you what it found. No sending, no updating, no spending. You learn how reliable it is on your actual data, at zero risk, and you will almost certainly discover it misunderstands something about your business that you would never have thought to specify.
Then let it draft, not send. Replies written and waiting in a folder. Invoices prepared but not issued. Reports built but unsigned. A person spends two minutes approving instead of twenty minutes writing, and nothing reaches a customer unseen.
Then allow reversible actions only. Scheduling a meeting that can be cancelled, adding a note that can be deleted, moving something between folders. Keep irreversible actions — payments, deletions, anything a customer sees immediately — behind human approval for as long as you sensibly can.
Give it the narrowest access that works. One mailbox rather than the whole tenancy. One folder rather than the whole drive. A test account first. If an agent only needs to read, it should not have permission to write, and if it only needs this year's records it should not be able to see the last ten.
The jobs that look perfect and are not
Some tasks have exactly the shape people associate with automation and turn out to be poor first candidates. It is worth knowing which, because failing on one of these early tends to sour a whole team on the idea.
Anything where the rules live in someone's head. Quoting is the classic example. It looks mechanical — dimensions in, price out — and then you discover the price depends on who the customer is, how busy the workshop is and whether they were difficult last time. Until that is written down, an agent cannot do it, and the act of writing it down is often more valuable than the automation.
Anything with a hard compliance boundary. Regulated advice, safety-critical instructions, medical or legal guidance. Not because the technology cannot produce plausible text, but because plausible is precisely the danger and the liability sits with you regardless.
Anything where the exceptions are the job. If nine in ten cases are routine and the tenth is where all the value and risk sits, automating the nine can mean the tenth is handled by someone who has lost the habit of noticing. Keep a human across the whole flow, or design the handover deliberately.
Anything customer-facing and unsupervised, early on. Your most fragile asset is the relationship. Earn confidence internally before you point an agent at the people who pay you.
Who owns it, and what it costs to keep running
An agent is not a purchase, it is a small piece of infrastructure, and infrastructure needs an owner. Before you start, name the person responsible for checking it still works, and give them actual time for it rather than adding it to a full week.
The running costs come in three parts. Usage is the obvious one, and it is higher than people expect because an agent makes many model calls per completed task rather than one. Maintenance is the one that bites: your systems change, an integration breaks, a password expires, and the agent fails quietly. Supervision is the third, heaviest at the start and reducing as you tighten the scope.
None of that is a reason to avoid agents. It is a reason to start with one job that clearly wastes hours every week, prove it properly, and expand from a position of knowing what it actually takes. A single well-scoped agent that reliably saves a morning a week is worth more than four ambitious ones nobody trusts.
What AI Agents Can Do – frequently asked questions
What can AI agents do for a small business?
Handle repetitive, multi-step admin: inbox triage, invoice chasing, first-pass research, CRM updates, diary management, content repurposing, drafting standard documents and answering first-line queries, with a human approving anything that matters.
Are AI agents worth it for a small business?
They can be, if you pick a genuinely repetitive task and supervise it. The value is the hours returned on dull work. Start with one job and prove it before you spend more.
What tasks should I not give an AI agent?
Final decisions on money, people or legal matters, anything high-stakes, and work where being confidently wrong causes real harm. Keep a human firmly on those.
How do I start with AI agents?
Choose one low-stakes, repetitive task, give the agent narrow access to only what it needs, keep yourself in the loop on anything that leaves the building, and widen only once it has earned trust.
Do AI agents replace staff in a small business?
Usually not. They remove the repetitive edges of jobs so your people focus on judgement, relationships and the work only a person can do. Think capable junior, not replacement.
Can an AI agent connect to my email and accounting software?
Often yes, either through built-in integrations or an automation platform sitting in between. The important question is not whether it can connect but how much access it gets: one mailbox rather than the whole system, read-only before read-write, and a test account before live data. Anything touching customer or financial records also needs a look at your UK GDPR position and your supplier's terms.
Who is responsible if an AI agent makes a mistake?
You are. If your business acts on an agent's output, or an agent acts on your behalf, the responsibility to the customer sits with you — vendor terms almost universally disclaim liability for output. That is why scope, approval steps and logging matter: they are not bureaucracy, they are how you keep a manageable problem from becoming a serious one.
How long does it take to set up a useful AI agent?
For a well-chosen job using tools you already have, a first read-only version can often be working within days. Getting to something you trust unsupervised takes longer — usually weeks of running it alongside a person and tightening the instructions as you find the gaps. Projects that aim for full autonomy on day one are the ones that tend to stall.
Which Oxford AI School training fits this need?
Pick the right first task and learn to run it well. Practical sessions for small businesses, in person or live online.
Want to put an agent to work, safely?
We help small businesses pick the right first task, set it up sensibly and build the skills to run it, without handing over the keys to everything.
Training for your business Are AI agents safe?
Harry Lang runs The Oxford AI School, teaching teams and businesses the basics of AI in a fun, easy and practical way. A marketing leader turned AI trainer, he can be reached at Learn@TheOxfordAISchool.com.