AI Agents for Business
AI agents for business that do one job well, with a person approving what matters
An AI agent is software that works through a task in several steps and chooses what to do next. For a small business, the useful ones are narrow: they chase, draft, sort and report inside the tools you already use, and they stop for approval before anything that costs money or reaches a customer. This page explains what agents are, which jobs suit them and where the limits sit.
What is an AI agent?
An AI agent is software that works through a multi-step task, deciding its next step itself and using your tools, instead of answering one question.
Anthropic, the company behind the Claude models, draws the line clearly in its guide to building effective agents. Workflows are systems where the model and tools follow predefined code paths. Agents direct their own process and decide which tools to use. The same guide recommends starting with the simplest approach that works and adding agents only when simpler ones fall short, because agents trade speed and cost for better results on open-ended tasks.
In a small business that difference is practical. A reminder that goes out seven days after an invoice falls due is a fixed rule. Reading the customer's reply, working out whether it is a dispute, a promise to pay or a request for a copy, and choosing what to do next is agent work.
Most useful business agents mix the two: fixed steps where the rules are clear, a model where something has to be read, judged or written, and a person at the points where a mistake would be expensive.
What is the difference between an AI agent, a chatbot and an automation?
An automation follows fixed rules, a chatbot answers questions in conversation, and an agent carries out a task across several steps and tools.
| Comparison | Automation | AI chatbot | AI agent |
|---|---|---|---|
| What it does | Runs the same steps every time a trigger fires | Answers customer questions in a conversation on your website or a messaging channel | Works through a task, choosing the next step and using several tools |
| How it decides | Rules someone wrote down in advance | Your approved information, matched to what the customer asked | A goal, your rules and what it finds along the way |
| Small business example | A reminder email seven days after an invoice falls due | Answering questions about opening hours, delivery areas and returns | Reading replies to overdue invoices, sorting them and drafting the right response |
| Typical failure | Breaks quietly when a form or tool changes | Gives a confident answer that is out of date | Takes a wrong step and carries on from it |
| Where a person comes in | Changing the rules when the business changes | Any question it cannot answer from approved information | Approving anything that spends money, commits the business or reaches a customer |
Plenty of jobs sold as agents are really automations, and that is not a criticism: a fixed rule is cheaper to run and easier to check. If your problem is a process with clear rules, the AI automation agency page is the better starting point. If it is customers asking the same questions, see AI chatbot for business. An agent earns its place when the inputs vary too much for anyone to write every rule down.
Which AI agents suit a small business?
Agents for repeated jobs with messy inputs: following up leads, drafting quotes, chasing invoices, processing documents, reporting, scheduling, customer service and answering staff questions.
Lead follow-up
Reads a new web enquiry, checks it against the work you take on, replies within minutes with the right questions and logs it in your CRM.
A person approves: Replies outside your standard wording, and any lead it is unsure about.
Quoting
Pulls your price list, similar past jobs and the site notes together into a draft quote.
A person approves: Every quote, before it is sent.
Invoice chasing
Sends reminders on schedule, reads the replies and sorts them into disputes, promises to pay and requests for a copy.
A person approves: Disputes, payment plans and any reply that mentions a complaint.
Document processing
Reads the invoices, forms and certificates that arrive by email, pulls out the details and keys them into the right system.
A person approves: Documents it cannot read with confidence, and totals over a limit you set.
Weekly reporting
Collects the week's numbers from your accounts, job and sales systems and writes a short summary of what changed.
A person approves: The summary, before it is shared beyond the management team.
Scheduling
Matches jobs to people and calendars, proposes times and handles the back and forth with customers.
A person approves: Changes to confirmed jobs and anything that moves someone's working hours.
Customer service
Answers routine questions from approved information and looks up an order or booking status where it is connected.
A person approves: Refunds, complaints and anything it cannot answer from approved information.
Internal knowledge
Answers staff questions about how things are done, from your own procedures and documents.
A person approves: Changes to the documents it answers from.
Nobody needs all eight, and in an owner-managed business of 5 to 50 staff one well-chosen agent usually beats several half-watched ones. The best known single-job agent is the AI receptionist, which answers and qualifies phone calls; the jobs above are its desk-based equivalents.
What are some examples of AI agents for business?
Good AI agent examples are narrow: an agent that answers and qualifies enquiries, one that collects documents, one that drafts replies. Each has a person approving the risky step.
The eight jobs above describe what agents do. The table below shows what the same idea looks like in five kinds of small business, with the step a person keeps. These are the patterns we scope most often, not case studies, and the sector pages go into each one in more detail.
| Business | Example AI agent | A person approves |
|---|---|---|
| Trades and construction | Reads each web or phone enquiry, checks it against the jobs you take on, asks for photos and the postcode, and books a site visit slot | Unusual jobs and any price mentioned to the customer |
| Accountants | Requests year-end documents, reads what comes back, ticks off the checklist and chases only what is still missing | Anything that goes into a client file or a filing |
| Law firms | Triages new enquiries by matter type, gathers the details a conflict check needs and drafts the first reply | Every reply before it is sent, and all conflict decisions |
| Real estate and estate agents | Answers portal enquiries within minutes, qualifies the buyer or tenant and offers viewing slots from the agent's calendar | Offers, anything about price or legal terms, and listing copy |
| Solar installers | Collects the documents a grant or connection application needs and tells the customer exactly what is still outstanding | The submitted application |
A first-hand example from my own business, Shamrock Electrical in Dublin: the first two workflows I built were invoice chasing and the weekly management report. The reminders themselves run on fixed rules, which is automation rather than an agent, and that was the right place to start. Reading and sorting the replies is the agent step, and it is only worth adding once the volume of replies justifies it. That is the honest pattern for most small businesses: automate the rule, then add an agent where the inputs vary.
What should an AI agent not be allowed to do on its own?
Spend money, send anything binding, change records that matter, make decisions about people, or delete data. Those steps wait for a person's approval.
The OWASP GenAI Security Project lists excessive agency among the main risks of systems built on large language models, and traces it to three causes: too much functionality, too many permissions and too much autonomy. Its recommended controls are the ones we build in: give an agent only the tools and access the job needs, and use a human in the loop to approve high-impact actions before they happen.
In practice, an agent that chases invoices can read the ledger but cannot issue a credit note. One that drafts quotes prepares the quote, and a person presses send. One that sorts the inbox can label and draft replies but cannot delete anything.
Decisions about people deserve a separate mention. Under Article 22 of the GDPR, people have the right not to be subject to a decision based solely on automated processing that has legal or similarly significant effects on them. Screening job applicants, refusing credit terms or closing a customer's account stay with a person, whatever the agent recommends.
Agents also make ordinary mistakes: they misread a scanned document, or take a wrong step and carry on from it. Approval points go where a mistake would cost the most, and a log records what the agent did so a person can check it.
How do we build and maintain AI agents inside your existing tools?
We map the job, give the agent narrow access to the tools it needs, test it on real work beside a person, then maintain it.
- Map the job. Every step a person takes today, the tools they touch, the decisions they make and what a good result looks like.
- Set the boundaries. What the agent may read, what it may change and where it must stop for approval, written down before anything is built.
- Build inside your tools. The agent works in your inbox, CRM, accounts package or job system through the connections those tools already offer, and puts its drafts and questions where your team already looks.
- Run it beside a person. On live work, with someone checking every output until you know how often it gets things wrong.
- Maintain it. Tools update, prices change and suppliers change their paperwork. We read the logs, fix what drifts and adjust the rules as the business changes.
No new software to learn and no developer on your staff. The full stage by stage process, with timings and what you provide, is on AI implementation services. A first workflow is usually live within two to four weeks after scoping.
How is customer data protected when an AI agent uses it?
By giving the agent the least access the job needs, using business-tier AI services under a data processing contract, and logging what it touches.
An agent that reads your inbox or accounts is processing personal data, so the usual rules apply. Article 28 of the GDPR says a business may only use processors that give sufficient guarantees of appropriate technical and organisational measures, and that the processing must be governed by a contract setting out what data is processed, why, for how long and for whom. The AI service behind an agent is one of those processors, which is why we set agents up on business accounts with proper terms rather than personal or free consumer accounts.
The practical questions are settled at scoping: which systems the agent can see, whether it needs customer names at all, where its logs are kept and for how long, who can read them, and whether your privacy notice needs updating.
The rules differ between the countries you might trade in, and we are not lawyers. Confirm the position with your own adviser before an agent handles customer or staff data; we will build to what they tell you.
First-hand example
Our own AI Readiness Review is rules plus a person, on purpose
The free AI Readiness Review on this site is an automation, not an agent, and it shows where the line sits. You answer six questions. Fixed rules check team size, revenue and country, and the instant report is assembled from workflows written in advance for each industry. Nothing in that path needs judgement, so nothing in it is left to a model.
When you submit, a copy of the report is emailed to you and every submission is sent to Ger, with strong fits flagged. The follow-up within 24 hours is written by Ger personally, because reading your answers and deciding what to say is the part where judgement matters. It is the same split we use when building agents: rules where rules work, a model only where something has to be read or written, and a person wherever a mistake would cost a customer.
Do you need AI agents, or would simple automation do?
Start with simple automation when the rules can be written down. Use an agent only where inputs vary and a person currently reads and decides.
Agents cost more to run and to check than fixed rules, because each step involves the model and each output needs watching until it has earned trust. Much of the marketing around AI agents skips that part.
For an owner-managed firm the honest order is usually: switch on the AI features already inside the software you pay for, automate the parts with clear rules, then add an agent for the step that still needs someone to read and decide. The free review names the two or three workflows worth doing first for a business like yours, before anyone decides whether they need an agent.
This page covers that last step. Choosing which processes to automate at all, and getting any workflow live, are covered on the automation agency and implementation pages linked above.
How much does an AI agent cost per month?
Monthly running cost is usage or subscription fees for the AI service plus maintenance, and both depend on how much work the agent handles.
An AI agent has three kinds of cost. Two of them recur every month, and one is paid once.
- Model or platform usage and subscriptions (monthly). An agent built on an AI provider's API pays for the text it reads and writes, measured in tokens. On 14 September 2026 Anthropic's pricing page listed its Claude Sonnet 5 model at $2 per million input tokens and $10 per million output tokens, before tax. That bill rises and falls with how much the agent reads and writes, so an agent handling a few enquiries a day and one processing documents all day cost very different amounts. Other agents run inside a tool paid for per user instead: the same Anthropic page listed a Claude Team standard seat at $25 per seat per month billed monthly, or $20 billed annually, for teams of 2 to 150, and OpenAI's ChatGPT pricing page says its paid plans are priced per user per month. If the agent also runs on an automation platform or connector service, that plan is another line on the bill.
- The build (once). Mapping the job, setting the boundaries, connecting your tools and running the agent beside a person on real work, as described above.
- Maintenance (monthly). Reading the logs, fixing what drifts when a connected tool updates, and adjusting the rules as prices, suppliers and the business change.
Provider prices are shown in US dollars as listed and change often, so check the provider's own page before you budget. For our own work on the build and maintenance, pricing follows the three points below.
01
Why there is no single figure
An agent that drafts replies to web enquiries in one inbox is a different job from one that reads supplier invoices, matches them to orders and keys them into your accounts. Price follows the tools it connects to, the approval steps, how much work it handles and the running costs of the AI service behind it.
02
What you get free first
The review, the instant report naming your highest-impact workflows with estimated hours saved per week, and a follow-up from Ger within 24 hours. You find out whether an agent is the right tool, or whether a simpler automation would do, before any money changes hands.
03
Who it is for
Owner-managed businesses with roughly 5 to 50 staff where someone spends hours every week reading, sorting, keying and chasing: trades, professional services, ecommerce, distribution and agencies.
Where we work
Clever Merchants is based in Dublin and builds AI agents remotely, inside your own tools, for businesses in Ireland, the UK, Australia, New Zealand, Canada and the United States.
Questions people ask about AI agents
Are AI agents safe to use in a small business?
They are safe when they have narrow access and a person approves anything that spends money, commits the business or reaches a customer. Most problems come from agents given more permissions than the job needs.
What is the best AI agent for a small business?
The best one does a specific, repeated job inside the tools you already use, such as chasing invoices or following up leads. Choose it the way you would hire for a role: name the job first, then check it connects to your systems.
Can an AI agent work with the software we already use?
Usually, yes, through the connections your email, CRM, accounts and job software already offer. Where a tool has no practical way in, you hear about it during scoping, before anything is built.
How much does an AI agent cost per month?
There is no single monthly figure, because running costs depend on how much work the agent handles and which tools it connects to. The AI Readiness Review is free, and the build and running costs are scoped after it.
What are some examples of AI agents for business?
Common examples are agents that follow up new leads, draft quotes, chase invoices and sort the replies, process emailed documents, write the weekly report, schedule jobs and answer routine customer questions, each with a person approving the steps that matter.
Do you offer AI agent consulting?
Yes. AI agent consulting here means deciding whether a job needs an agent at all, designing its access and approval points, then building and maintaining it. It starts with the free AI Readiness Review, and plenty of jobs turn out to need a simpler automation.
Will AI agents replace my staff?
In an owner-managed business they usually take over the repetitive part of a role, such as chasing, keying in data and first drafts. People keep the judgement calls, the approvals and the customer relationships.
Find out whether an agent should come first
Six questions, about three minutes, an instant report naming the workflows with the highest impact in your industry, and a personal follow-up from Ger within 24 hours. Free, no card, no obligation.
Take the free AI Readiness Review