What is an AI agent?
An AI agent is a written instruction file. Not software, not an app, not a robot: a document. In practice it's a markdown file, plain text with headings, sitting in a folder inside Claude, ChatGPT or a similar tool. The file tells the model what its job is, how to do it, how to behave, what it may and may not do, and which tools it can use (search the web, read a folder, send an email, update a spreadsheet). Given a goal, the model reads the file and decides for itself what to do next until the job is done. Simon Willison, a developer who has written about this more clearly than any vendor, describes what that file makes the model do in eight words: it "runs tools in a loop to achieve a goal". That's the definition. Everything else is packaging.
Here's how I'd say it to a mate in the pub. Typing into ChatGPT is you saying "act like the marketing director of a small accountancy firm and write me an email". One go, one answer, you decide what happens next. An agent is different. It's a particular set of skills written down, in plain files you could open in Notepad: how you do certain things, how you want it to act, and rules on what it may and may not do. You give it the goal, it works through the skills, and it uses the tools it's been allowed to use. The trick, and the thing the vendors never mention, is that you want each one as narrow as possible. Take SEO. There are dozens of separate tasks in it, so you'd write each agent one task it knows inside out, then write one agent on top that talks to all of them. That's the difference between an agent and a prompt: the prompt is a question you type once, the agent is a small, well-briefed team you wrote down once and can run again tomorrow.
The companies that make the models mostly agree, and disagree in ways that matter to you. Anthropic draws the line at who's in charge of the steps: a "workflow" follows a path someone coded in advance, an "agent" decides its own path as it goes. OpenAI and Google both say a simple chatbot is not an agent, however clever it sounds. Gartner defines agents as software "granted rights by the organization to act on its behalf", which is the most useful definition of the four for an owner, because it's really a sentence about liability. You are granting rights. Remember that when we get to the airline. Notice, too, that they all say "software" or "system". They sell software. From where I sit, the thing you actually write and own is the file. The software is just the model reading it.
Agent vs assistant vs automation vs prompt
None of the ten pages that rank for this phrase in the UK will give you this table, because nine of them sell agents and the tenth is a Reddit thread. So here it is, with the version of each that runs on this site.
| Thing | What you give it | Who decides the next step | Can it act without you? | On this site |
|---|---|---|---|---|
| Prompt | One question or instruction | You, every time | No | "Rewrite this email so it's shorter." |
| Assistant | A conversation plus your context (a business brief, past work) | You, every time, but it remembers | No | Claude with my one-page brief open, drafting a letter I then edit line by line. |
| Automation | A fixed rule: if this happens, do that | Nobody. The rule is the decision | Yes, but only the one thing | Comment a keyword on one of my Instagram posts and a link lands in your DMs. No thinking involved. It's a rule. |
| Agent | A written file: the job, how to do it, the rules, which tools it may use. Then a goal | The model, step by step | Yes, if you let it | My research-then-draft pipeline, if I let it run with nobody between the steps. I don't. Keep reading. |
Two things fall out of that table. First, a lot of what's sold as an agent is an assistant or an automation with a new label. Gartner reckons only about 130 of the thousands of vendors claiming "agentic AI" actually offer it, and has a name for the rest: agent washing. And even the real ones are, underneath, a model reading an instruction file that you could have written yourself. Second, the test is simple. If it can't take an action you didn't specifically approve, it's an assistant with a marketing budget. That's not an insult. Assistants are brilliant. It's just a different thing, at a different price, with a different risk.
What an AI agent for marketing would actually do for a business with no marketer
The easiest way to explain it is the way I write articles, including this one. The keyword research is already done and sitting in a spreadsheet. What's left is a chain of jobs: search the phrase and save what's actually ranking, research the topic properly for the nuance the top pages miss, draft the article in the house voice from a folder of things that are true about the business, make the images, check it. In an agent setup, each of those is its own narrow agent, which means its own file. A research file. A drafting file. An image file. And one on top handing the work along and checking it against the brief. Four documents, written once, in a folder.
If you asked me which marketing jobs I'd hand to an agent for a five-person firm today, it's two. Research: what your customers are searching, what the top results say, what they miss, what your competitors are claiming. And first drafts: the initial layout of a page, the first version of an email sequence, the twelve captions from twelve photos. Both are jobs where a wrong answer is cheap, because a human reads the output before anyone else sees it.
Now the honest bit. I'm not running any agents autonomously at the moment. I've got the skills written and implemented, and the process is documented well enough that I could let it run. I don't, because I always have a human check before anything goes live, and at work, even after my own check, it goes to my boss for a read before it's published. That's not caution for the sake of it. It's the80/20 rule from the AI guide: the machine gets you most of the way, and the last part, the bit where you know something the machine doesn't, is the whole value. The thinking stays human. The typing gets delegated.
"You want each agent as narrow as possible. One job it knows inside out, and one agent on top."
Where AI agents go wrong
An assistant that hallucinates wastes ten minutes of your evening. An agent that hallucinates does it at speed and then acts on it. Four things worth knowing before you give one any rope.
- It's confidently wrong, and it doesn't slow down. The AI guide has the example from my day job: reports produced with AI in the loop that padded sections nobody asked for and got a statistic flat wrong. A human caught it because a human reads every number. An agent with permission to send would have sent it.
- It has an edge to its competence, and neither of you can see it.The best study on this is from Harvard Business School and BCG, who ran a pre-registered experiment on 758 consultants in 2023. On 18 tasks inside the AI's capability, people using it finished 12.2% more work, 25.1% faster, at around 40% higher quality. On the one task deliberately placed just outside its capability, the AI users were 19 percentage points less likely to get the right answer than the people with no AI at all. The researchers called it the "jagged frontier". You can't see it from the outside, and the agent can't either.
- The customer doesn't care that a bot said it. In February 2024 a tribunal in Canada ordered Air Canada to pay C$812 to a passenger whose website chatbot had invented a bereavement refund policy. The airline argued the chatbot was "a separate legal entity that is responsible for its own actions". The tribunal's reply: "It makes no difference whether the information comes from a static page or a chatbot." A month earlier, here in the UK, DPD's customer service bot swore at a customer and wrote a poem about how useless DPD was, after a system update. They switched it off the same day. Small money, both of them. The point is who carried the can.
- The rules already apply. The ASA's position is that the advertising codes have no AI-specific rules because they don't need any: the existing rules apply however content is made, and you must hold evidence for any objective claim before it runs. The ICO has no agent-specific instrument either; the normal data protection and direct-marketing rules cover it, including consent for email and the do-not-call registers for phones. An agent that emails a lead you never had permission to email has broken the same law you would have. This isn't legal advice, it's the plain version of what the regulators have published.
None of this is an argument against agents. It's an argument for what I do now: let it draft, keep a person between the draft and anything a customer sees.
What AI marketing agents cost (checked 11 September 2026)
Prices below are what the vendors published on the day I checked, in whatever currency they publish in, and they move fast. Salesforce alone has changed its pricing model three times in eighteen months. Treat these as the shape of the cost, not a quote.
| Option | Cost | The catch |
|---|---|---|
| A paid chat assistant (Claude, ChatGPT, Gemini) | £15 to £20/mo | Enough for most firms, and enough to run Claude Code, which is where my agent files live. I pay for a bigger plan because of how much I use it. You don't need to. |
| Relevance AI (build-your-own agents) | Free for 200 actions/mo; Pro $19 to $29/mo | "Actions" and model credits are billed separately, so the bill has two dials. |
| Zapier Agents | Free for 400 activities/mo; Pro about $33/mo | Zapier is moving agents onto per-step metering with multipliers for bigger models, so a busy agent costs more than the plan price suggests. |
| ManyChat "AI" replies (Instagram, WhatsApp) | Free to $69/mo, plus $29/mo for the AI step | Reviewers describe the AI step as keyword-matching, not a conversation. The comment-to-DM flow on my account is the free, rule-based bit. |
| HubSpot Breeze agents | Professional seat from $90/mo, one-off onboarding $1,500 to $3,500, then $0.50 per resolved conversation | The agent doesn't exist on the Starter plan at all. The content agent's cost per piece is more than a Starter plan's whole monthly allowance. |
| Salesforce Agentforce | $2 per conversation, or $125 per user/mo, on top of an existing Salesforce licence | You're buying a floor before you buy the agent. |
| Your checking hour | Nobody publishes this | I looked. There is no independent estimate of set-up or supervision time for a small firm. Budget it the way you'd budget a junior's work, because that's what it is: someone else's draft you're responsible for. |
The last row is the one that matters. A £20 assistant plus twenty minutes of your evening does the research-and-first-draft job for almost every firm reading this. The costs above only make sense once the volume is real and the process is boring. The basics, what to pay for and what not to, are in theAI marketing guide's cost table.
Do you actually need one? Five questions before you spend anything
When a five-person firm with no marketer asks me whether they should get an AI agent, my answer is the same every time: there's no point unless you have processes in place, written down to a T, that you now want to automate. So before the free trial:
- Is the job written down so simply that someone else could follow it?Every step, in order. If you can't write it down, you don't understand it yet, and neither will the machine.
- Have you done it by hand at least three times? Automate nothing until it bores you. Boredom is the signal that the judgement has gone out of it and only the typing is left.
- Does it have a source of truth to read from? One folder: who you are, who you serve, what you charge, what you never say. The agent refers back to it with everything it does. Without it you get fluent guesses.
- What happens when it's wrong, and who sees it before a customer does?If the answer is "nobody", you don't have an agent, you have a liability with a subscription.
- Would a £20 assistant and twenty minutes do the same job?Usually yes. Start there, and move to an agent when the twenty minutes has become two hours a week of the same boring steps.
If any answer is no, you're not ready, and that's fine. Most firms aren't, and the DSIT survey of 3,500 UK businesses in 2025 found only 16% of firms with five or more staff using any AI at all. The path is: prompt, then assistant with your context, then a fixed automation for the bits that never change, then, maybe, one narrow agent for one boring job with you reading every output. In that order, and no skipping.
The stats you'll be sold
The ad copy for this search term says "agents that work while you sleep". The numbers underneath it are worse. A quick burial:
- "10x productivity." No source, anywhere. The most rigorous study on the subject, the Harvard and BCG experiment above, measured 12% more work and 25% faster, inside the AI's competence, and worse results outside it. Twelve percent is a good number. It isn't ten times.
- "80% of enterprises will be using AI agents by 2026." Nobody can trace it to a report. Gartner's own 2026 survey of chief information officers found 17% had deployed agents; Deloitte found 11% had anything production-ready.
- "40% of agent projects will be cancelled by 2027." This one is real: a Gartner prediction from June 2025, citing rising costs, unclear value and weak risk controls. It's also a prediction whose methodology Gartner never published, so quote it as an opinion from a firm that sells opinions, not as a fact.
- "Multi-agent systems outperform single agents by 90.2%." That's Anthropic's own internal research evaluation of its own system, and the version that gets pasted around drops the next line: the multi-agent version used about 15 times the tokens. It was better because it was allowed to spend more.
- "55% of UK small businesses now use AI." From a survey of 904 FSB members in November 2025. The government's survey of 3,500 firms with five or more staff, earlier that year, found 16%. The ONS, in June 2026, found around 35% of firms with ten or more staff. All three are honest. They're measuring different businesses at different times. There is no single UK number, and anyone who quotes one without a date and a sample is selling you something.
The honest verdict
AI agents are real, they're a text file and a model, they're useful for two jobs, and they're mostly being sold to people who haven't written their process down yet. The ads say agents that work while you sleep. Mine don't, because I'd rather sleep than wake up to an email that went out under my name with a number in it nobody checked. Do the job by hand. Write it down. Hand over the boring bits one at a time. Keep the judgement, and keep a person between the draft and the world. If you want to start where I'd start, theAI marketing guide is the assistant stage done properly, and the free prompt pack is the twenty minutes.
Sources: Anthropic, "Building effective agents" (December 2024) and "Measuring agent autonomy" (quoting Simon Willison); OpenAI, business leader's guide to agents; Google Cloud, "What are AI agents" (February 2026); Gartner research note on agentic AI and Gartner press release, 25 June 2025 (the 40% prediction, the ~130 genuine vendors and "agent washing"); Dell'Acqua et al., "Navigating the Jagged Technological Frontier", Harvard Business School Working Paper 24-013 (758 BCG consultants, 2023); Moffatt v. Air Canada, 2024 BCCRT 149; DPD statement, January 2024; ASA/CAP guidance on AI in advertising (2025 to 2026); ICO, "Regulating AI: the ICO's strategic approach"; DSIT, AI Adoption Research (3,500 businesses, fieldwork February to May 2025); ONS Business Insights and Conditions Survey wave 159 (June 2026); FSB member survey (November 2025, 904 owners); Gartner 2026 CIO Survey and Deloitte via secondary reporting; vendor pricing pages for Relevance AI, Zapier, ManyChat, HubSpot and Salesforce as published on 11 September 2026, in the vendors' own currencies. Named as unreliable and not used: "10x productivity", "80% of enterprises by 2026", "90.2% better" without its token cost, "748% SEO ROI", and any single UK adoption figure. Jasper, Copy.ai, Lindy and Make prices were not independently verified and so aren't printed. My own workflow figures are recollection. Nothing here is legal advice.