Three different products are being sold to you under one word.
One of them has been around for a decade and runs for the price of a modest software subscription. One of them can cost more each month than a part-time member of staff. The demo looks about the same for all three, and the person doing the demo has no particular reason to explain the difference.
Gartner has a name for that blurring. It calls it agent washing: the rebranding of assistants, chatbots and robotic process automation as agents without the substance to back it up. The underlying shift is real enough, and Gartner expects 40% of enterprise applications to carry task-specific AI agents by 2026, up from under 5% the year before. But a lot of what wears the badge is software you were already paying for.
So here’s the plain version: what each of the three actually is, what it’s good at, where it falls over, and how to work out which one belongs on the process you had in mind. If you want the deeper read on agents specifically, AI agents for business is the hub for this one.
The one-line version
A chatbot talks. It answers questions in a conversation and the conversation is where it ends.
An automation follows your rules. You write the steps, it runs them the same way every time, and it has no opinion about anything.
An agent takes a goal. It works out the steps itself, uses your real systems to carry them out, and stops at a written line where a person takes over.
That’s the whole distinction. Everything below is the detail that decides which one you should be buying.
Chatbots: software that talks
A chatbot sits in a conversation. Somebody asks it something, it answers, and then it’s done. Modern ones answer very well, because the language models underneath them are good. That doesn’t change the shape of the job.
The giveaway is what happens at the end of a useful exchange. If the outcome is that the customer now knows something, it’s a chatbot. If the outcome is that something changed in your systems, it isn’t.
What it’s genuinely good at: answering the same 30 questions your team answers every week. Opening hours, delivery times, “where’s my order”, how a product works, what your returns policy says. Also internal use, which owners forget about: a chatbot that knows your own handbook and process documents saves your team asking each other.
Where it falls over: the moment the customer wants something done rather than explained. It can tell somebody their order is late. It can’t chase the supplier, issue the credit, or rebook the delivery.
There’s a second problem that gets underplayed, which is that people can tell. The customer service numbers have been moving in an awkward direction for a while: Gartner expects half of the organisations that planned significant customer service headcount cuts to abandon those plans by 2027, and its survey of 321 service leaders found only 20% had actually reduced staffing because of AI, with 55% holding staffing flat while handling more volume.
Read that second figure the right way round and it’s good news. The businesses getting value from a chatbot used it to take more work with the same team, not to shrink the team. The ones that promised themselves a headcount saving are the ones writing it off.
Automations: software that follows your rules
An automation is a set of steps you wrote down, running on a trigger. Form submitted, so create the record, send the email, add the task, post to the channel. Zapier, Make, Power Automate, the workflow builder inside the software you already pay for, or a few lines of code somebody wrote in an afternoon.
It’s the oldest of the three and by some distance the most underrated. Every week we look at a business that’s been quoted for an AI build when what it needed was 40 minutes in a workflow builder.
What it’s genuinely good at: anything where the rule never changes. Moving data between two systems, sending the same message when the same thing happens, creating the same records in the same order, generating the same weekly export. It’s cheap, it’s fast to set up, and best of all you can read it and understand exactly what it will do.
Where it falls over: anything with an exception in it. An automation does precisely what you told it, including when what you told it is now wrong. The supplier changes their invoice layout and the rule that reads the total keeps reading the wrong box. Nothing errors. The numbers are just quietly incorrect from Tuesday onwards.
That brittleness is why automations get abandoned, and it’s almost never a technology failure. It’s that nobody owns the rules as the business moves underneath them.
Worth knowing that the analysts think most businesses should be staying here for now. Forrester expects fewer than 15% of firms to turn on the agentic features in their automation platforms in 2026, with governance and unproven returns keeping the rest on plain deterministic automation. When the thing touches billing or a customer promise, predictable beats clever.
AI agents: software you give a goal to
An agent gets an outcome rather than a script. “Deal with the enquiries that came in overnight” instead of “if the form says X, send template Y”. It decides the order and the method each time, reads and writes in your real systems, and hands anything on the restricted list to a named person.
The last part is what separates a finished agent from a switched-on one. A written list of what it never decides alone: refunds, pricing, anything legal, anything you can’t take back.
What it’s genuinely good at: the work that’s 80% identical and 20% judgement, done often enough that the 20% still eats an afternoon a week. Reading and routing what comes in. Assembling quotes and proposals from details scattered across three systems. Chasing. Reconciling two systems that have never spoken properly and reporting what it changed.
Where it falls over: three places, and you can check all of them before you sign anything.
It fails on repeats. An agent that completes a task in a demo will not complete it every time. Sierra’s research team built a benchmark that runs the same customer service task at an agent repeatedly instead of once, and 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. Ask any supplier to run the same job 20 times on your data and show you all 20 results.
It fails on cost. An agent thinks about every case rather than following a rule, so it costs more per job than an automation doing something similar. Third-party running costs for a small business system typically land somewhere around £150 to £400 a month, and an agent pointed at a high-volume process sits at the top of that or above it. Fine when it’s replacing an afternoon a week of somebody’s time. Painful when nobody worked out the volume first. Escalating cost is the first reason Gartner gives for expecting over 40% of agentic AI projects to be cancelled by the end of 2027.
And it fails on ownership. Somebody has to read what the agent flagged and notice when the flags change shape. That’s 20 minutes a week, not a job, but it has to be somebody’s 20 minutes. It’s also why we sell support monthly with no lock-in: an agent nobody looks at gets quietly worse until the team goes back to doing the work by hand.
The three side by side
| Chatbot | Automation | Agent | |
|---|---|---|---|
| What you give it | A question | A rule | A goal |
| Who decides the steps | You, in the script | You, in the rules | The software, each time |
| What it touches | The conversation | Whatever you wired up | Your systems, with permission |
| When something is unusual | Says it can’t help | Does the wrong thing quietly | Flags it and asks a person |
| Cost to run | Low | Lowest | Highest, and varies with volume |
| Cost to set up | Low | Low | Highest |
| Breaks when | The question is off-script | The process changes | Nobody retrains it |
| You can predict the output | Mostly | Exactly | Not exactly |
| Good for | Answering | Moving data | Finishing a piece of work |
The row people skip is the second from the bottom. An automation will do the same thing every single time, and for some processes that certainty is worth more than the cleverness. Payroll, invoicing, anything a regulator might ask about later. Take the boring option there and spend the budget somewhere it earns more.
Five questions that tell you which one you need
Run the process you have in mind through these. The answers usually point at one column on their own.
1. At the end of a good outcome, has anything changed in your systems? No means a chatbot is enough. Yes means it isn’t.
2. Can you write the rule down completely, including the exceptions? If you can fill a page and genuinely cover it, that’s an automation and you should stop reading. If the page ends with “and then you use your judgement”, that’s agent territory. Getting those exceptions out of your team’s heads is most of the work of training an agent.
3. How often does it happen, and how long does each one take? Multiply them. If the answer is under an hour a week, nothing here will pay for itself, whatever it costs.
4. Has the process changed in the last six months? If it’s still moving, automate it and you’ll be rebuilding it. Settle it first. Which processes to automate first covers how to tell when a process has actually settled.
5. Who owns it after launch? If there’s no name, buy the automation. It’ll degrade more gracefully than an agent nobody is watching.
If you’d rather work through your whole business than one process at a time, the systemisation scorecard ranks them for you and will tell you when the honest answer is to leave something alone.
Most real systems are all three
This is the part the comparison articles miss. In practice you don’t pick one. You end up with layers, and the good builds are mostly the cheap layer.
Take a business handling 200 enquiries a week. A chatbot on the website answers the 30 standard questions before anybody has to. An automation creates the record, assigns the owner, and sends the acknowledgement, because those steps never vary. An agent reads the 40 that didn’t fit the script, pulls the customer’s history, works out what’s actually being asked, drafts the reply, and flags the 6 that need a person to make a call.
Three technologies, one process, and only the last slice is expensive. That’s what a sensible build looks like, and build vs buy for AI covers how to work out which slice is worth owning and which you should keep renting. If a supplier’s proposal is agents top to bottom, they’ve either not looked closely at your process or they’re pricing the fun version.
The reverse is just as common, mind. Businesses that wire everything together with rules and then wonder why somebody still spends Thursday afternoon fixing what the rules got wrong. That afternoon is the bit that needed judgement, and it’s been sitting there the whole time.
What buying the wrong one costs you
Two ways to get this wrong, and they cost differently.
Buying an agent for an automation job. You pay more to build it, more every month to run it, and you get output you can’t fully predict for a task where predictable was the whole point. The tell is a process where nobody has ever had to use their judgement. Cheapest mistake to fix, most expensive to keep.
Buying an automation for an agent job. Cheaper on the invoice and worse over time. The rules cover 70% of cases, the other 30% land on somebody’s desk as exceptions, and because those exceptions arrive scattered through the week nobody adds them up. You’ve automated the easy part and left the expensive part exactly where it was. This is the more common of the two, and it’s most of why businesses tell us automation “didn’t really do much”.
There’s a third mistake that’s worse than either: buying any of them for a process nobody costed first. That’s the main reason AI pilots never ship. Everyone agreed it felt quicker, nobody could put a number on it, and the renewal conversation went badly. The manual work cost calculator will give you a starting figure in about five minutes, and the cost of manual work has the arithmetic behind it.
Where most UK businesses actually are
Useful context before anybody talks you into the expensive column.
The British Chambers of Commerce found 54% of UK SMEs actively using AI in 2026, up from 35% a year earlier, with 95% reporting no change to workforce size. Set that next to Forrester’s expectation that fewer than 15% of firms will switch on agentic features this year and the picture is clear enough. Almost everybody is at the chatbot and automation layer. Very few have built anything that finishes a piece of work on its own.
That’s an opening rather than a warning. The businesses pulling ahead aren’t the ones that bought the most advanced thing. They’re the ones that put the cheap layer on the boring 80% and spent the real money on the one process where judgement was eating a day a week.
The short version
A chatbot answers. An automation follows the rules you wrote. An agent takes a goal, works out the steps, and stops at a line where a person decides.
Most things sold as agents are one of the other two wearing the word. Ask what changes in your systems when it works, and whether you could write the rule down completely. Those two questions sort almost every process into the right column.
Then count what the process costs you by hand before you buy anything at all. Whichever column you land in, that number is the only thing that tells you afterwards whether it worked.
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 for business: what they actually do
- 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)
- Build vs buy for AI: when off-the-shelf is genuinely fine
- Case study: how MidShift built an AI career guidance engine for 20,000+ professionals
Been quoted for an agent and not sure you need one?
Send us the process. We’ll tell you which of the three it is, and we’ll say so plainly when the answer is a cheap automation in a tool you already pay for.
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 first.