Managed automation is automation that the builder keeps monitoring, adjusting and accounting for after delivery; at Oase Creative, a person has recorded 83 approvals since 23 September for what our own system was allowed to send out.
What is managed automation?
Managed automation is automation that is not left to fend for itself after delivery. The party that builds the workflow or the AI agent also keeps it running, monitors it and adjusts it. A person sits between the system and the outside world before anything goes out, there is a log of what the system did, and there is a pause switch. So you are not only buying a built automation, you are buying the certainty that someone keeps checking whether it still does what it should.
The term leans on a concept every business owner already knows: managed IT services. There you don't rent a server, you rent someone who keeps the server running. Managed automation does the same for work processes. The newsletter that puts itself together, the incoming requests that get sorted, the data that moves from one system to another overnight.
Why that distinction matters shows in the numbers. Statistics Netherlands measures that among Dutch micro businesses, process automation stalls at 2.3%. The tools to automate processes are available everywhere. What is missing is someone who keeps them working after week five.
Why is building only half of AI automation?
Because automation that works today does something different a month from now. An external service changes its connection, a credit balance runs out, a language model gets updated, a customer fills in a form slightly differently than expected. The build is a moment. Running it is every day.
We don't know this from a white paper but from our own log. Since March 2026 we have been running an agent system that helps run our agency. Measured over 134 days, roughly three in four scheduled runs made it to the finish line, and on 54 of every 100 days something went wrong that needed attention. That is not a failing system. That is what a running system is.
Larger organisations show the same pattern in a different shape. Deloitte surveyed more than 3,000 executives and found that only 25% had moved at least 40% of their pilots into production. On top of that, 37% mostly use AI on the surface, without real changes to the underlying business processes.
The pilot proves something can work. Management proves it keeps working. Managed automation is the name for that second part, and it is the part where most AI automation runs aground.
How does managed automation differ from a tool subscription and an agentic workflow?
Three things get mixed up in conversations about AI automation: the tools, the technique and the management. All three are needed, but they are different purchases.
A tool subscription is tooling. Zapier, Make or n8n give you the building blocks. You build, you monitor, you step in. That works well for anyone with the time and knowledge. What a subscription costs, by the way, depends less on the number of workflows than on the number of tool calls an agent makes per task.
An agentic workflow is a technique. Anthropic, the company behind Claude, draws the line sharply: workflows follow predefined code paths, while agents direct their own next step and the tools they use. An agentic workflow is a work process in which such an AI agent makes part of the decisions.
Managed automation is a way of running things. It says nothing about the tools or technique underneath, and everything about who is responsible when something goes wrong.
| Tool subscription | One-off build | Managed automation | |
|---|---|---|---|
| Who builds | You | The builder | The builder |
| Who monitors after delivery | You | Nobody in particular | The builder |
| Who notices when it breaks | You, if you look | Often the customer | Alerts to whoever runs it |
| Person between system and outside world | If you build one in | If it was agreed | Always, on record |
| Log of what happened | Platform log | Varies | Standard part |
| Pause switch per task | Manual | Varies | Standard part |
So the difference is not in the build, but in everything after it.
What goes into managed automation?
Five parts turn a built automation into a managed one. We use all five in our own system, and the same five go into what we build for clients.
A person between the system and the outside world. Everything that goes out on behalf of your business, a publication, an email, a message to a prospect, first needs a person's approval. For us that is not a tick in a spreadsheet but a recorded approval, tied to an actual message from Louie. Since 23 September 2026, 83 of those approvals have been recorded. Without one, the system refuses to publish or send.
A log. Every risky action writes down what is about to happen before it happens. That way you can always look back at what the system did, and why.
A pause switch per task. Every task has its own switch. One file, and that one task stops while the rest keeps running.
Monitoring and alerts. A system that fails silently is more dangerous than one that fails loudly. In our case, for example, a separate check runs outside the AI system and sends an alert as soon as the AI service returns two errors in a row.
A quality gate. Output passes fixed checks before it is offered for approval. Of all failed runs in our log, almost a third was our own quality gate rejecting the output.
| Part | What it prevents | How it runs for us |
|---|---|---|
| Human approval | A wrong email or publication you can't take back | 83 recorded approvals since 23-09-2026 |
| Log | Not knowing what happened | Written ahead of every risky action |
| Pause switch | Stopping everything for one problem | One switch per task |
| Alerts | Silent failure | A separate check next to the AI system |
| Quality gate | Mediocre output going out anyway | Almost a third of failed runs was a rejection |
What does managed automation look like for a client?
At Familierestaurant Pluis the newsletter runs as a system. The owners drop a topic into a channel in our software. We write, design and send. A person approves every email before it goes out.
Two systems run around it. Every hour, new sign-ups from the website go to the mailing list, and anyone who ever unsubscribed never gets added back. Every month a report is built from Google Analytics, Search Console and the sending data, in the Pluis house style. We read that report before it goes any further.
Look at what does not happen there. An email never goes out automatically. A guest who unsubscribed never gets mailed again by accident. And a report never lands with the owners unread. In an ordinary automation, each of those three mistakes is one misconfigured field away. In managed automation they are caught, because someone stays responsible after the build.
This is also why we don't see managed automation as a technical service but as part of the brand. A newsletter that fits the rest of the restaurant is brand work. The automation makes sure it arrives on time every month, the management makes sure it keeps fitting.
Why does almost no automation ask a person first?
Because nobody notices it is missing while building. The first version works, the demo goes well, and the approval step feels like a delay. Until the first time the agent sends something out that is wrong.
Researchers analysed more than six thousand public n8n workflows with a language model inside. Only 2.78% of 6,003 workflows ask a person for approval before an external action. In 26.52% the model's output goes straight into an action without any check. The researchers point out themselves that these are shared designs, not exactly what companies run in production.
At the same time, autonomy is growing. In a vendor survey by Zapier and Centiment among 508 small US businesses, 30% say they already let AI act across several systems with little human input. And among larger organisations, according to Deloitte, only 21% have a mature way to govern those agents.
More autonomous systems, hardly any checkpoints. That gap is what managed automation fills: the human in the loop is not an option you tick, it is the starting point.
When do you not need managed automation?
Not everything that runs automatically needs to be managed. Anyone who claims otherwise is selling management where a script will do.
You can do without it when three things are true at once. The automation stays inside your own business. A mistake is easy to undo. And you notice quickly yourself when something goes wrong. A workflow that copies orders from your store into a spreadsheet, or a reminder your team gets every Monday, falls into that category. A tool subscription and a bit of discipline are enough there.
In our own system too, most scheduled tasks run without a language model. A script that does the same thing every night needs no AI and no daily oversight. It changes when three questions get a yes:
| Question | A yes means |
|---|---|
| Does something go out on behalf of your business? | A person has to see it first |
| Does an AI agent decide on a step by itself? | You want to see what it chose and why |
| Do you only hear about a mistake when a customer reports it? | You need alerts, not hope |
Two yeses is the moment to take management seriously. Three yeses without management is a risk, not a saving.
How do you start with managed automation?
Not with the tools, but with the work. You can take these five steps this week, even if you haven't built anything yet.
- Write down which work comes back every week. Who does it now, how much time it takes, and what goes out. That last column matters most.
- Mark every point where something leaves your business. An email, a publication, an invoice, a reply to a customer. Each of those points later gets a human approval.
- Agree who gives that approval. A name, not a job title. And agree what happens when that person is away: the system waits, it does not carry on by itself.
- Decide how you stop it. Per task, not everything at once. If you don't know how to switch off one part, you are not ready to switch it on.
- Plan the first review. After a month you look at what the system did, what went wrong and what needs to improve. That moment is the start of management, not the end of the project.
With those five steps on paper, you can have an automation built or build it yourself. Either way you then know what you need after the build. And you know what to test a provider on: if you are looking for an AI consultant, you ask not only what gets built, but also who looks after it once it is delivered.
How does Oase Creative approach managed automation?
We build AI automation and agentic workflows for recurring work, from content and email to admin. Our services page puts it like this: AI systems and automation, with your brand rules built in, a person who reviews everything that goes out, and a pause switch. After delivery we stay involved to check that it keeps working, and after the first month we review together what it does.
We work this way because we built it on ourselves first. Our own agency runs on an agent system with 61 background tasks. Every mistake that happened in it has become a rule: no publication without a recorded approval, no send without a log, no task without a pause switch. What we build in for clients is what has already gone wrong for us once.
Managed automation is also a service of its own: we build, monitor and improve the automation, with a fixed review and pricing on request. See how it works on the managed automation page.
We are a brand agency, and that is no coincidence in this story. An automation that speaks on behalf of a brand has to sound like that brand and must never contradict it. That is why the brand rules sit inside the system itself, and why a person keeps watching. Want to know which work at your company qualifies first? Tell us where things get stuck, and we will look together at where you can best begin.
