Every piece of software you already pay for grew agents at some point in the last 18 months. Your CRM has agents. Your helpdesk has agents. The automation tool somebody set up in 2023 and left behind has agents now too.
Most of them are the same product with a new sticker on the box. Gartner reckons only around 130 of the thousands of vendors selling agentic AI are doing anything that earns the word, and has a name for what the rest are up to: agent washing, the rebranding of assistants, chatbots and robotic process automation without the substance underneath.
So this is the plain version. What an agent actually is, the jobs they genuinely do inside a business your size right now, where they fall over, and how to work out whether one belongs anywhere near your week. If you’re at the earlier question of what AI does in a company your size at all, start with the complete guide and come back to this one.
What an AI agent actually is
An agent is software you give a goal to rather than a set of steps. It works out the steps itself, uses your real systems to carry them out, and stops to ask a person when it reaches something it shouldn’t decide alone.
Three parts of that sentence do the work.
A goal, not a script. “Deal with the enquiries that came in overnight” instead of “if the form says X, send template Y”. The agent decides the order and the method each time.
Your real systems. It reads and writes in your CRM, your inbox, your accounts package, your job sheets. Software that can only talk is not an agent, it’s a chat window.
A stopping point. A defined list of things it never calls on its own, and a named person it passes those to. An agent without that list hasn’t been finished. It’s just been switched on.
Here’s the difference laid out against the two things people usually mean when they say agent:
| Chatbot | Automation | Agent | |
|---|---|---|---|
| Who decides the steps | You, in the script | You, in the rules | The software, each time |
| What it can touch | The conversation | Whatever you wired up | Your systems, with permission |
| When something is odd | Says it can’t help | Breaks or does the wrong thing | Flags it and asks a person |
| What it’s good at | Answering | Moving data | Finishing a piece of work |
Nothing on that table makes agents the best option. Plenty of jobs are better served by an automation costing £30 a month, and a good supplier will tell you so. AI agents vs chatbots vs automations takes that table apart properly, with the five questions that sort a process into the right column and what it costs you to pick the wrong one. Which processes to automate first then covers the order to do them in.
The jobs they actually do
Forget the demos with a robot answering the phone. Here’s the work agents are doing inside real small and medium businesses this year, with the line to a person marked in each one.
1. Reading what comes in and routing it
Enquiries, invoices, job requests, supplier emails, forms. The agent reads each one, pulls the relevant history, works out what it is and who it belongs to, files it, and drafts the reply.
The line: it sends the standard replies and holds anything unusual for a person. Most owners expect the split to land the other way round and are surprised by how much of their inbound turns out to be the same six situations.
2. Drafting the thing that’s nearly the same every time
Quotes, proposals, onboarding packs, chase emails, standard responses. The agent assembles the document with the right details pulled from your systems rather than from a person’s memory of where the details live.
The line: a person approves before it leaves the building. You keep the judgement and lose the typing, which is usually most of the hour.
3. Chasing
Nobody enjoys the follow-up sequence, so it gets done inconsistently and revenue leaks out of the gap. An agent chases every time, at the right interval, and stops the second somebody replies.
The line: it escalates to a person when the answer isn’t a yes or a no, and it never touches anything involving money owed beyond a polite reminder.
4. Keeping two systems telling the same story
The job that eats an afternoon a week in almost every business we look at: the same information re-entered in a second place because the two systems have never spoken. The agent reads one, works out what the other one needs, and writes it, including the parts that need a small judgement about formatting or matching.
The line: it reports what it changed. Silent data writing is how you end up with a mess nobody can unpick.
5. Putting the weekly numbers together
Pulling from three or four places, reconciling the differences, writing the summary, flagging what moved. The kind of work that takes somebody senior two hours every Friday because only they know where the numbers live.
The line: the agent produces the report, a person reads it before anybody acts on it.
Notice what those five have in common. They’re all high-volume, mostly identical, and already being done by somebody who’d rather be doing something else. That’s the shape of the job an agent suits, and it hasn’t changed much in two years.
Where the line has to sit
The most useful thing in any agent build is the written list of what it never decides alone. Refunds, pricing, hiring, anything with a legal or safety edge, anything you can’t take back.
This isn’t caution for its own sake. It’s the reason the thing survives contact with your business. Gartner expects at least 15% of day-to-day work decisions to be made autonomously by 2028, up from none in 2024. Read that the other way round and the overwhelming majority of decisions still sit with people, even on the optimistic view. The businesses getting value are the ones that worked out which slice was which before go-live rather than after an incident.
We drew that line hard on MidShift. The software handles assessment, matching and the repeatable guidance for more than 20,000 professionals. A human mentor still owns the conversations where being wrong would cost somebody a career move. That split is why it works at that volume.
If a supplier tells you the agent handles everything, they’ve either never run one past launch or they’re hoping you stop asking. How to choose an AI consultant covers the rest of the questions worth asking on that call.
Why so many get switched off
Agents fail for four reasons, and only one of them is technical.
They work once and not eight times
This is the one nobody warns you about. An agent that completes a task in the demo will not complete it every time, and the gap is wider than it looks.
Sierra’s research team built a benchmark called τ-bench that runs the same customer service task at an agent repeatedly instead of once. The best model tested solved a retail task 61% of the time on a single attempt and only about 25% of the time across eight consecutive attempts. Models have improved since. The shape of the problem hasn’t, and single-run success still tells you almost nothing about Monday morning.
What this means for you is simple. Ask any supplier to run the same job 20 times on your data and show you all 20 results. A demo that ran once proves the thing is possible, not that it’s reliable.
Nobody costed the before
If you never counted what the process cost by hand, you can’t tell whether the agent helped. That’s most of why pilots never ship: everyone agreed it felt faster, nobody could put a number on it, and the renewal conversation went badly.
McKinsey’s 2026 global survey found 37% of organisations reporting that AI contributed to earnings, flat on the year before, even though far more of them were scaling the technology than a year earlier. Broad deployment without anybody redesigning the actual work produces thin returns, and thin returns get cancelled.
Count first. The cost of manual work has the arithmetic, and the manual work cost calculator will give you a starting figure in about five minutes.
The costs run away
Agents cost more to run than automations because they think about every case rather than following a rule. Model usage, hosting, the tools underneath. Typical third-party running costs for a small business system land somewhere around £150 to £400 a month, and an agent pointed at a high-volume process sits at the top of that range or above it.
That’s fine when it’s replacing an afternoon a week of somebody’s time. It’s not fine when nobody worked out the volume in advance. Escalating cost is the first reason Gartner gives for expecting over 40% of agentic AI projects to be cancelled by the end of 2027.
Nobody owns it
Work with nobody’s name against it doesn’t happen. An agent needs one person who reads what it flagged, notices when the flags change shape, and says when something needs fixing. That’s 20 minutes a week, not a job, but it has to be somebody’s 20 minutes.
Agents rot
This is the part suppliers skip, so we’ll be blunt about it.
Your business changes. You add a service, drop a supplier, change a price, hire someone who does the job differently. Every one of those quietly moves the ground under an agent that was trained on how things worked in March.
Nothing breaks loudly. The agent keeps running and gets slowly worse: more cases flagged for a person, more replies that read a bit off, more small corrections nobody mentions. Six months on, half the team has gone back to doing it by hand and nobody remembers deciding to.
What stops that is dull and cheap. Somebody reads a monthly report on what the agent handled and what it escalated, and the thing gets retrained when the business moves. Teaching an agent how your business actually works is the same job as keeping it taught, so it’s worth reading how that’s done before you start. That’s the whole trick, and it’s why we sell support monthly with no lock-in rather than shipping a build and disappearing.
The agents that pay for themselves are the ones somebody kept an eye on. Founderise took 12 hours a week off its founder and lifted margin 3.5x with an MVP live in 9 weeks, and it still gets looked at every month.
Should you build one at all
Probably not yet, if we’re being honest about where most businesses are.
The UK numbers say as much. The British Chambers of Commerce found 54% of UK SMEs actively adopting AI in 2026, up from 35% a year earlier, and yet the government’s own AI adoption research, covering 3,500 UK businesses, put agentic AI at 7%, far behind every other kind of AI it measured. Nearly everybody is at the assistant layer. Very few have built anything that finishes a piece of work.
That gap is an opportunity rather than a warning, but only if you pick correctly. A process is worth an agent when it’s high volume, mostly the same every time, already settled, and expensive in hours. Take one away and the answer is usually an automation, a form, or a tool you already pay for that nobody configured.
Four questions to run your candidate through:
- How many times a week does it happen, and how long does each one take?
- How much of it is identical, and how much needs somebody to weigh something up?
- Has the process changed in the last six months, and will it change in the next six?
- Who reads what the agent flags, and do they have 20 minutes a week?
If you want to work that through across your whole business rather than one process, the systemisation scorecard ranks them for you and will tell you when the answer is to leave something alone.
Where agents are heading
Worth knowing, even though none of it changes what you should do this quarter.
Customer service is the front line. Gartner expects agentic AI to autonomously resolve 80% of common customer service issues by 2029, taking around 30% out of operational costs. Whether that lands on schedule matters less than the direction: the repeatable half of contact work is going, and the businesses that already counted their volumes will move first.
The practical read for an owner is unglamorous. Nothing about that prediction requires you to buy anything this year. It does mean the process you’re planning to hire for is worth costing before you post the job.
The short version
An AI agent takes a goal, works out the steps, uses your systems, and stops at a written line where a person takes over. Most things sold as agents are chatbots or automations wearing the word.
They do genuinely useful work: reading and routing what comes in, drafting the near-identical documents, chasing, keeping two systems honest, assembling the weekly numbers. They fail when nobody costed the process first, when nobody owns the escalations, when the running costs were never worked out, and when nobody retrains them as the business moves.
Pick one process. Count it. Ask to see the same job run 20 times before you believe anything. That order beats every agent strategy you’ll be sold.
Related reads
- How to train an AI agent on how your business actually works
- Shadow AI: what your team is already doing with company data
- AI agents vs chatbots vs automations: what’s the difference
- AI for small business: the complete guide
- What AI can actually do for a small business (and what it can’t)
- Which processes to automate first (and how to rank them)
- The cost of manual work: what repetitive admin really costs
- How much does it cost to add AI to a small business?
- Why most small business AI pilots never ship
- What an AI audit is and what you should get from one
- How to choose an AI consultant (and the red flags)
- How AI agents handle client onboarding while you sleep
- Case study: how MidShift built an AI career guidance engine for 20,000+ professionals
- Case study: how Founderise automated delivery
Got a process in mind and no idea whether it’s agent-shaped?
Run it through the systemisation scorecard first, then bring us the worst one. We’ll tell you what we’d build, what we’d wire together for a fraction of the cost, and what we’d leave alone.
Whatever we do end up building, the price is fixed and agreed in writing before anything starts. No day rates, no paid discovery, and a full refund if you don’t approve the prototype.
Book a free call, or read what the AI audit covers before you commit to anything.