Insights
What is an AI agent, and can a small business use one?
- AI
- Automation
- Small business
The short answer
An AI agent is a language model given a goal, a set of tools and permission to keep working in a loop until the goal is met. For small businesses it suits short, repetitive, checkable tasks of three to five steps, and it is a poor fit for anything irreversible, customer facing, or long.
An AI agent is software that gives a language model a goal and a set of tools, then lets it work through several steps on its own until the goal is met or it gives up. A chat assistant answers one question and waits; an agent keeps going, chooses what to do next, and acts. For most small businesses the useful answer is narrow: agents suit short, repetitive, checkable tasks of three to five steps, and they are a poor fit for anything irreversible or long.
What is an AI agent, in plain terms?
Two terms are worth fixing first. A language model is a program that predicts text one piece at a time, based on everything it has been shown, and the same technology sits behind everyday chat assistants. A tool, in this context, is a small piece of ordinary software the model is allowed to call, such as checking the stock level for a product or creating a calendar entry.
Without tools, an agent is a chatbot with a longer prompt. Tools are what let it change something rather than describe it, and they are also what make the risk real. A model that can write to your order system can write the wrong thing to your order system.
How does an AI agent actually work?
The loop is easy to describe. The software sends the model the goal and the history so far. The model replies with either an action or a final answer. The software carries the action out, appends the result, and asks again. That repeats until the model says it has finished or a step limit stops it.
A useful picture is a contractor handed a job each morning with no memory of yesterday, but the picture matters less than the mechanism: the model holds no state between turns. Every turn, the entire running transcript of the task is pasted back into the prompt from the beginning.
That single fact explains most agent behaviour. A longer task means a longer transcript, which means more input on every turn, higher cost, slower replies, and a growing chance the model repeats work or forgets a constraint it was given earlier.
Why do AI agents fail at long tasks?
Errors compound. If each step succeeds 95% of the time, a five step task finishes unaided around 77% of the time, ten steps around 60%, and twenty steps around 36%. That is multiplication rather than opinion, and it is why a three step demo looks far better than a fifteen step system in daily use.
Failure is not only about stopping. A wrong value produced at step two gets handed to step three, which acts on it confidently. Models rarely announce that they are working from bad input.
There is also a hard ceiling. A model can only consider the text inside its context window, the fixed amount of material it can hold at once, measured in tokens, where a token is roughly three quarters of a word. Once a task transcript runs past that limit, older details are summarised or dropped.
The practical rule: if a process needs more than about five steps, split it into separate jobs with a check between them rather than asking one agent to work straight through.
What are AI agents genuinely good at in a small business?
Short, repetitive, checkable work where a mistake is cheap to spot is the sweet spot.
- Sorting inbound enquiries into categories and drafting a reply for someone to approve.
- Moving data between two systems that have no integration, such as copying order details from an email into a spreadsheet row.
- Gathering information from a fixed list of sources, for example pulling the key figures out of ten supplier quotes into one comparison.
- Checking a set of conditions once a day and raising a flag, such as a competitor price dropping below a threshold.
- Reformatting, translating or summarising routine content so it matches a house style.
The pattern is consistent: few steps, a clear finish line, and an output a person can verify in seconds.
What is the difference between an AI agent and workflow automation?
Workflow automation follows rules a person wrote in advance. When an invoice email arrives, save the attachment and add a row to a spreadsheet. That approach is cheap, predictable and fast, and for a fixed process it beats an agent almost every time.
An agent decides the steps itself, which helps when the input varies. Reading a messy supplier email that could be a quote, a delivery note or a complaint is a job for a model. Pulling a known field out of a known format is not.
The test is whether you can describe the rule in advance. If you can, automate it. If you cannot, and the task is short, an agent is worth trialling.
Where should a small business not use an AI agent?
Do not give an agent the ability to move money, delete records, publish to your live website, or message customers without review. A failure rate near 5% per step is manageable when a person checks the output, and unacceptable when the output is already live or already paid for.
Personal data needs care as well. Every agent turn sends the contents of the task to a third party model provider, so if that content includes customer details you are exporting them, and you need a lawful basis and a contract that covers it.
Skip agents entirely where a wrong answer is worse than no answer. Pricing promises, legal and medical statements, and anything a customer will read as advice should stay with a person.
What does an AI agent cost to run?
Two costs matter, and only one of them tends to appear in a demo.
Assistant products with agent features typically sit at £15 to £40 per user per month, which covers light use. Pay-as-you-go model access is charged per token, and rates vary by tier, but competent text models sit broadly in the region of £1 to £15 per million input tokens, with output costing several times more.
The multiplier that surprises people is the transcript. A single chat question might use around 1,000 tokens. A ten step agent task can use 20,000 to 200,000, because the whole history is resent on every turn.
A task might therefore cost a fraction of a penny, or a pound or two. Run it 50 times a day and the monthly bill lands somewhere between a few pounds and several hundred. Estimate from your own transcripts before scaling, rather than from a vendor example.
Custom work is a development project. An agent that reads supplier emails and updates a stock system typically costs £3,000 to £15,000 to build reliably, plus ongoing model costs. Off-the-shelf automation platforms with AI steps start around £20 to £100 a month.
How do you know whether an agent is worth keeping?
Run one task for four weeks and record three numbers: how often it completes without a human correction, the cost per completed task, and the human minutes saved each week.
Set the bar at around 80% unaided completion for a short task. Below that, someone is redoing the work and the agent has added a step rather than removed one.
If it saves less than about 30 minutes a week, stop. Maintenance, prompt fixes and checking take time, and a task that small will not repay the attention.
Should a small business build an agent or buy one?
Buy first. A subscription to an assistant with agent features costs a few hundred pounds a year and will tell you inside a month whether your task is a good fit.
Custom development makes sense only once a task is already working, runs often, and joins up two systems you own. That order matters. Building before measuring is how projects reach five figures and then quietly stop being used.
Start with back office work rather than customer facing work. Reports, internal summaries and data tidying get reviewed anyway, so a mistake costs a correction rather than an apology.
Keep reading
What is a website accessibility audit, and do you need one?
A website accessibility audit checks whether disabled customers can use your site. What it covers, what it costs in 2026, and when it is worth paying for.
Read the articleHow much does a small business pay for a WordPress website in 2026?
Real 2026 prices for a WordPress website, what changes the cost, and when a cheap or expensive build is the right choice for a small business.
Read the articleFind out how visible you actually are
A free written audit of how search engines and AI assistants currently read your website. No obligation, and no sales call required to receive it.
Prefer to talk first? Email contact@luiinteractive.com or message +44 7349 961542 on WhatsApp.