What's the actual difference between an automation and an agent?
An automation follows rules you set. An agent makes decisions you didn't. That's the whole distinction, and getting it right will save you a pile of money and a bigger pile of mess.
An automation is a train on rails, when this happens, do that. A form comes in, so a row is added, so an email fires. It doesn't think. It doesn't need to. An agent is handed a goal instead of a script, and it works out the steps itself, choosing what to do next based on what it finds. One is predictable and cheap. The other is flexible and, if you're not careful, expensive and unpredictable.
Here's the bit most people get backwards: agents are the shiny thing right now, so everyone wants one. But most marketing problems don't need judgement, they need something reliable done the same way every time. So start with automation. Only reach for an agent when the job genuinely can't be written down as rules.
What does the spectrum actually look like?
It runs from "records a click" to "decides what to do about it", and most useful work sits nearer the boring end.
| Level | What it is | Example |
|---|---|---|
| Macro | A recorded set of clicks, replayed | Auto-formatting a spreadsheet export |
| Workflow automation | Fixed if-this-then-that logic across tools | New lead in a form → CRM → Slack ping |
| AI-assisted | Automation with a smart step inside it | A step that drafts a reply, a human sends it |
| Agentic | Given a goal, decides its own steps | An orchestrator that plans your year and reshuffles it |
The trick isn't climbing to the top. It's stopping at the lowest level that solves your problem. A workflow in n8n, Make or Zapier that never surprises you beats a clever agent you have to babysit, nine times out of ten.
When is an agent worth it (and when is it a liability)?
An agent is worth it when the job genuinely needs a decision made repeatedly, on changing information, that you can't reduce to fixed rules. Everywhere else, it's a liability.
If you can describe the job as "when X, do Y", it's an automation, build that, it'll be cheaper and it won't wander off. An agent earns its keep when the inputs keep shifting and someone has to weigh things up: what's most important this week, what moved, what to do about it. That's judgement, and judgement is what agents are actually for.
The liability shows up when people hand an agent a job that didn't need one, usually something customer-facing, and then trust it to run unsupervised. That's how you get an agent confidently emailing the wrong thing to the wrong list. Flexibility cuts both ways: the same freedom that makes an agent useful is what lets it make a mess at scale.
What have we actually shipped?
An orchestrator for our own marketing. Not a writer, not a designer, a planner that keeps every event in the business front of mind and tells us what to start, when.
Every year is full of things you have to work backwards from: events we're attending, products we're launching, new collabs, a partner's own event. Each one has a lead-up. Launch a product in four months and you want to start teasing it and gathering early feedback now, drop previews at three months out, open pre-orders at two, and turn the hype up at one. Multiply that across everything happening in a year and it's a mountain of sequencing, the kind that normally lives in Google Sheets and Asana and gets reshuffled on gut feel whenever a new date lands.
So we built an orchestrator that does it properly. It has access to our own data on what's actually worked, past launches, collabs, events, so it prioritises on evidence, not vibes. It pulls new events out of our meetings and emails, drops them onto the timeline, and keeps the whole plan project-planned and current every week. When a launch slips from September to October, the run-up reshuffles itself instead of rotting in a spreadsheet.
"We trust it a lot precisely because it's not communicating out to anyone. It only talks to us internally, 'you're launching this, so you need to start that this week.' Keeping it that tightly scoped is exactly why it works so well."
The other thing it does is tell the other agents when to start their jobs for each event. But it stays an orchestrator and nothing else, it doesn't try to write, design or publish. That tight scope, with us as the humans in the middle, is the whole reason we trust it. An orchestrator that also tried to be a copywriter is an orchestrator you couldn't trust.
What guardrails does an agent need?
Three, really, and they're not optional.
- A tight scope. One agent, one job. The more an agent is allowed to do, the harder it is to trust and the easier it is to break. Our orchestrator orchestrates, full stop.
- A human in the loop. The agent surfaces and sequences; a person makes the call on anything that matters. That's not a training-wheels phase you grow out of, it's the design.
- Inward before outward. Point the risky, brand-facing output at yourselves first. Internal instructions are safe to automate today. A customer-facing voice earns that trust slowly, if ever.
Get those three right and an agent stops being a gamble. Skip them and you've built a very confident way to embarrass yourself at scale.
What will agents realistically own in marketing within 12 months?
Single-purpose agents, the ones handed one job, will go mainstream. The big all-running orchestrators will start to appear, but they won't be everywhere yet.
It depends a bit on whether you're an AI-first business. But the pattern I'd bet on for most companies is agents given a single activity. A newsletter agent whose only job is to send one newsletter a month, it knows what happened in the company, what's on the calendar, and out it goes. Or a social agent you hand a topic to that builds a week of content around it.
Those are brilliant for anyone doing no marketing at all right now. A genuinely good social media manager still beats the platforms, but that single-purpose agent is dead easy to stand up and get running, which is exactly why it'll spread.
"Those single-purpose agents, run my newsletter, run my social from this topic, become mainstream in the next 12 months. Everyone will have them. The broader thing, an orchestrator running sub-agents per department with humans in the middle, starts appearing at some companies, but it won't be universal yet."
That broader setup, an orchestrator running sub-agents across every department, humans in the middle, like what we've built, will show up at some companies inside a year. It just won't be the norm. If you want a hand working out which of your jobs is an automation and which genuinely wants an agent, that's exactly what our Marketing AI work is for, request a quote and we'll map it with you.
Frequently asked questions
Is an AI agent just a chatbot?
No. A chatbot answers when you talk to it; an agent is given a goal and takes actions on its own to reach it, planning steps, calling tools, deciding what to do next. A chatbot responds; an agent does.
Are agents reliable enough in 2026?
Reliable enough for tightly-scoped, internal jobs with a human in the loop, yes. Reliable enough to run unsupervised, customer-facing work, not really. Keep the scope narrow and point risky output inward first, and they're genuinely dependable.
What can go wrong with an agent?
The same flexibility that makes an agent useful lets it make a mess at scale. Most often when it's handed a customer-facing job it didn't need and left to run unsupervised. Tight scope, a human on the call, and internal-first output prevent nearly all of it.
Do I need engineers to run agents?
Not usually to start. Most marketing wins are automations you can build in n8n, Make or Zapier without a dev. Agents take more care to scope and wire up safely, that's where a partner helps, but you don't need an engineering team to get real value.
