AI agents for small business: what they do and what they cost

Every software product you already pay for now seems to have an "agent" in it. Your CRM has one. Your helpdesk has one. Your accounting package is threatening to get one.

Gartner estimates that only about 130 of the thousands of vendors selling agentic AI are the real thing. Plenty of the rest are rebadged chatbots and automation tools, a practice Gartner calls "agent washing". So before you buy AI agents for small business work, it's worth knowing what you're actually buying.

The short answer: an AI agent is software that takes a goal, decides the steps, uses your tools (email, calendar, CRM, spreadsheets) to carry them out, and checks its own progress. For most small firms, the cheapest start is an agent feature inside software you already pay for. A custom agent typically starts in the tens of thousands of dollars, and running it properly costs more than the AI model does.

Key points:

• Plenty of things sold as agents are chatbots or workflow automations with a new label.
• Published agency ranges for a custom single-workflow agent run from roughly $15,000 to $40,000 (RaftLabs), and those figures come from firms that sell the builds.
• Model costs are usually small. Supervision, fixes and integration upkeep are not.
• Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027 (Gartner).

What is an AI agent, and how is it different from a chatbot?

A chatbot answers. An agent acts.

Ask an agent for your next free slot and it opens your calendar, finds the slot, books it, emails the customer a confirmation and logs the booking in your CRM. It works out the steps itself, calls the tools it's allowed to use, and handles the case where the first slot has gone.

Workflow automation (Zapier, Make, Power Automate) sits in between. It follows the exact steps you drew. It's cheap, predictable and dull, which is a compliment. An agent earns its higher cost only when the work needs judgement: reading a messy email, deciding which of five things the customer actually wants, or choosing what to do when the data doesn't fit.

Gartner's own advice puts it neatly: use agents when decisions are needed, automation for routine workflows and assistants for simple retrieval. It also says plainly that many use cases sold as agentic don't need an agent at all.

What can AI agents for small business actually do?

The jobs that work well share a profile. They happen often, they follow a recognisable pattern, they involve reading unstructured text, and a mistake is cheap to catch. Here's where I'd look first.

• Inbound enquiries. Reading web forms and emails, answering the routine ones, qualifying the rest and booking calls.
• Phone answering. Taking calls out of hours, booking appointments and passing on messages. This is its own category now; see whether AI receptionists are worth it.
• Admin triage. Sorting the inbox, chasing missing paperwork, drafting replies for a human to send.
• Finance back office. Matching invoices to purchase orders, flagging anomalies and preparing (not approving) payment runs.
• Research and reporting. Pulling numbers from three systems into a weekly summary that someone used to spend Friday afternoon building.

Where they still struggle: negotiation, anything involving an upset customer, decisions with legal or financial consequences, and tasks where the "right" answer lives in somebody's head rather than in a system.

For context on where UK firms actually are, the Office for National Statistics found in June 2026 that 28% of businesses with 0 to 9 employees used at least one AI technology, against 49% of firms with 250 or more staff (ONS). The smallest adopters were also more likely to report using AI extensively. That is a self-reported depth-of-use measure, not evidence that every small adopter gets better results. More on getting started in AI for small business in the UK.

How much does an AI agent cost to build?

There are four examples below. The figures use different units and scopes, so they are not a like-for-like price ladder.

Two health warnings. Every custom range here comes from agencies that sell agent builds; I haven't found an independent survey of AI agent development cost. And these are US dollar figures, so UK and international prices vary with where the team sits.

What actually moves the number is integrations. Each system the agent has to read from or write to is its own piece of engineering, and older systems without decent APIs cost far more than modern ones. Compliance (health data, payments, regulated advice) adds another layer. For the wider budgeting picture, see how much AI implementation costs.

How much does an AI agent cost to run each month?

Less than you'd fear for the model, more than you'd expect for everything else.

Take a worked example. Say an enquiry agent handles 1,000 conversations a month, and each one uses about 20,000 tokens of input (instructions, context, tool results) and 2,000 tokens of output. Those volumes are my illustrative assumption, not a benchmark. On Anthropic's published API prices as of 10 October 2026, a mid-tier model (Claude Sonnet 5.5, $2 per million input tokens and $10 per million output) comes to about $60 a month. The smallest current model (Claude Haiku 5.5, $0.10 and $0.50) comes to about $3 (Anthropic pricing).

This calculation assumes the stated total input and output tokens per conversation at standard global rates. It excludes tool charges, hosting and extra model passes. Retries, repeated context and long prompts can change the bill; Haiku's quoted rate applies to prompts up to 100,000 tokens.

These assumed token volumes make inference cheap. Real workloads can be very different. The expensive part of an agent is the human who keeps it honest.

The real running costs are these:

• Someone reviewing what the agent did, especially in the first months.
• Fixing it when a connected system changes its API, a form gets a new field or a price list moves.
• Updating its instructions and knowledge as your business changes.
• Hosting, logging and monitoring, if it's custom.

Ask any supplier for a 12-month total cost, with a named person responsible for maintenance. If they can only quote the build, they're quoting half a project.

Why do so many AI agent projects get cancelled?

Gartner's prediction is that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. Its analyst described most current projects as early experiments "mostly driven by hype and are often misapplied".

All three causes show up in small businesses, just at smaller scale.

Escalating costs usually means the integration work was underestimated, or nobody budgeted for the person who has to supervise the thing. Unclear value means nobody measured the "before" state, so nobody can prove the "after". Inadequate risk controls means the agent was given permission to send, pay or delete before anyone had watched it make decisions for a month.

That last one is the scary one. An agent with write access to your email can embarrass you at machine speed. The fix is boring: start read-only, add draft-only, and give it the power to act on its own only for low-risk steps you've seen it get right many times. The pattern behind most failures is covered in why AI projects fail.

Do you need an agent, or just automation?

Run your candidate task through three questions.

1. Do the steps change from case to case? If they're always the same, a workflow automation is cheaper and more reliable.
2. Does the input arrive as messy human language? Emails, call transcripts and scanned documents are where AI earns its place. Clean form data rarely needs it.
3. Is a wrong answer cheap to catch? If an error costs you a customer or a regulator's letter, keep a human approving every action.

If the first two answers are yes and the risks can be contained, an agent may be worth trialling. An unsafe failure mode is a stop condition, whatever the other answers say. Fewer than that and you probably want automation with a small AI step in the middle, such as summarising or classifying. That hybrid is the unglamorous workhorse for most small firms. Whether to buy it or build it is a separate decision, covered in build vs buy AI.

What I've seen running my own agents

I run my own laboratory of AI agents, Ascendancy IQ, doing real work under human direction. It's an evolving experiment rather than something I sell, and it has taught me more about agents than any vendor demo.

Memory is the first thing that breaks. Out of the box, every session starts cold. The agent forgets yesterday's decision and cheerfully contradicts it. I ended up building a shared memory layer just to stop repeating myself, which I wrote up in I fixed the dumbest thing about AI agents.

Approval gates are the feature. My agents can draft, research and prepare almost anything. The actions that matter still wait for a human yes. That's slower than full autonomy and far cheaper than cleaning up after it.

Agents need a manager. The common mistake I see is treating an agent like software you install and forget. It behaves more like a capable junior hire with no common sense: brilliant at volume, needs clear instructions, and needs its work checked. If nobody in the business owns the agent, the agent owns you.

How to introduce your first AI agent

1. Pick one task that happens at least daily, involves reading messy text and has a cheap failure mode.
2. Measure it now: volume, time per item, error rate, response time. This is your baseline.
3. Check whether a tool you already pay for has an agent feature that covers it. Try that before building anything.
4. Run it read-only or draft-only for two to four weeks. A human approves every action and logs every mistake.
5. Write down the permissions it will get, the steps it may take alone and the ones that always need a human.
6. Name an owner who reviews its work weekly and fixes it when something changes.
7. Compare against the baseline after a month. Expand, fix or switch it off. All three are good outcomes if you decide on evidence.

If the trial needs to convince a sceptical co-founder or board, structure it as a proper AI proof of concept with success criteria agreed up front.

Frequently asked questions

What is the best AI agent for a small business? The best first agent is usually one built into software you already use, because the integration work is done for you. Start from the task you want done, then find the tool, and trial it on one workflow before paying for more seats.

How much does an AI agent cost per month? It depends on the route. One agency puts DIY and low-code agents at roughly $50 to $2,000 a month in platform and usage fees. For a custom agent, the model itself can cost a few dollars a month, while supervision, maintenance and platform fees make up most of the bill.

Are AI agents safe to use with customer data? They can be, if you control what they can see and do. Use business plans that don't train on your data, give agents the minimum access they need, keep logs and tell customers when they're dealing with AI. The AI governance framework guide covers the basics.

Will an AI agent replace my staff? It may change tasks, staffing needs or both. Headcount statistics cannot tell you how your team will use the capacity. Agree the intended outcome with staff, then measure service quality and workload as well as cost.

Sources

• Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, Gartner press release, 25 June 2025
• Artificial intelligence in UK businesses: 2023 to 2026, Office for National Statistics, 20 July 2026
• Claude API pricing, Anthropic, checked 10 October 2026
• AI Agent Development Cost in 2026, RaftLabs (agency estimate)
• AI Agents for Small Business: A Practical Guide, LOW/CODE Agency (agency estimate)

If you're weighing up your first AI agent and want a straight answer on whether it's worth building, let's talk.