Vibe automation means building a workflow by describing it in plain language and letting AI assemble the steps. In 2026 that part works. Running the workflow is a different job: real orders and invoices arrive messy, and the value sits in reading them, checking them against your systems and knowing when to stop and ask a person.
We know this from our own history. In July 2025 Lleverage launched our vibe automation feature: type what you want automated and get a working workflow in seconds. The interest was far beyond what we expected. What we learned in the months after is the reason for this post. At Lleverage we think the description was never the hard part. A description captures the path the work takes on a good day. A back office lives in the other days.
That is why we now build AI agents that take a whole job off the desk, such as order entry or invoice matching, and run it in the background with every exception handed to a person. If you want to see the difference on your own documents, book a demo.
What is vibe automation in 2026?
Vibe automation is the practice of creating an automated workflow from a description in plain language. You write what should happen, and an AI model chooses the steps, the connections and the logic. The term borrows from vibe coding, which applies the same idea to writing software.
Vibe coding spread fast. Collins Dictionary named it its word of the year for 2025 and defined it as "the use of artificial intelligence prompted by natural language to assist with the writing of computer code". Workflow builders followed the same route, and our own 2025 feature was one of them: type a sentence such as "when an order email arrives, extract the lines and create the order in the ERP" and get a draft flow back.
For a demo or a prototype, that is a real step forward. A first draft that used to mean dragging boxes around a canvas now arrives in seconds. The trouble starts on the first morning the flow meets the real inbox.
Why does a workflow built from a description break on real work?
It breaks because the description holds only the cases the writer thought of. Real input arrives in formats nobody planned for, with gaps and with customer habits that nobody wrote down. A workflow that follows a fixed path either stops on those cases or, worse, pushes a wrong record into the ERP.
Take order intake. The sentence "extract the lines and create the order" assumes each line carries an article number. At SIG Benelux, around 3% of incoming orders carry an article number and roughly 47% have only a description. The rest arrives as a WhatsApp message, a photo pasted into an email or a sentence that only makes sense if you know the customer.
Nobody can describe that in a prompt, because the knowledge is not in anyone's head as a rule. It sits in years of order history: this description, from this customer, means that article. A workflow assembled from a sentence has no access to it.
The second break is the system on the other end. Writing into an ERP is guarded for good reason, and an on-premise ERP often has no open door at all. Microtechniek's ERP, Ridder IQ, ran on-premise without the API module the cloud version has. That door had to be built before any order could go in, and according to Microtechniek's story 2 other automation projects at similar companies had already failed at exactly that point. A workflow builder cannot describe its way past that. Someone has to build the connection.
What is the difference between describing a workflow and running it?
Describing a workflow produces a plan for the expected case. Running it means handling every case that arrives, at any hour, against your live data, with a record of what was done and why. The table below sets the two side by side for a typical back office process.
| Describing a workflow | Running it in production | |
|---|---|---|
| Input | The example in the prompt | Every PDF, spreadsheet, email body and phone photo that arrives |
| Exceptions | Not covered unless someone thought of them | The daily work: unknown articles, missing references, odd formats |
| Knowledge | What the writer put in the sentence | Order history, customer habits and the corrections the team makes |
| Uncertainty | The flow continues or fails | A confidence on every field, held back below your threshold |
| Systems | A connector, if one exists | Read and write access scoped per system, including on-premise ERPs |
| Change | Edit the prompt and rebuild | Correct the agent once and it applies the correction next time |
| Accountability | Not designed in | Every action logged with what was read, decided and why |
The left column is what a vibe automation demo shows. The right column is what a finance director or an order desk manager will ask about before they let anything write into the ERP.
Why do so many AI agent projects stall after the demo?
Most stall because the demo proved the description and nobody priced the running. Gartner predicted in June 2025 that more than 40% of agentic AI projects will be cancelled by the end of 2027, and named escalating costs, unclear business value and inadequate risk controls as the reasons.
Our reading is that all 3 reasons trace back to the same gap. Costs escalate when every exception becomes a ticket for a developer to fix the flow. Business value stays unclear when the workflow handles the easy cases and the team still checks everything because they cannot tell which records are safe. Risk controls are inadequate when the flow has no notion of confidence and no audit trail.
The same Gartner forecast estimated that only about 130 of the thousands of vendors selling agentic AI are real, and called the relabelling of chatbots and robotic process automation as agents "agent washing". A useful test for any vendor is to ask what happens to the order the system is unsure about. If the answer is a log entry, you are looking at a workflow with a new name.
What does an AI agent do differently when it runs the work?
An AI agent that runs a process reads each document on its meaning, checks it against your ERP and business rules, scores its confidence per field, and acts only above the threshold you set. Below it, the work goes to a person with everything already gathered. Corrections go back into the company brain.
Three design choices separate running from describing.
- Held back, not guessed. A confidence score for each field, not one for the document. Below your line the agent waits and does not write its best attempt into the ERP.
- Handed over whole. The exception arrives in Teams, Slack or the inbox with the document, the record and the reason attached, so the person decides rather than starts again.
- Corrected, not reprogrammed. When someone fixes a match, the agent keeps that as a rule your team can read in the company brain. There is no prompt to rewrite.
The results show where the effort goes. At Topa Bathroom Products, more than 90% of incoming orders are now created in Business Central without manual input, and some customers receive their order confirmation within 30 seconds. The agent at Topa also handles the exception logic: when a commission number is missing it drafts the reply to the customer, and when an article code is unrecognised it flags it for review.
At SIG Benelux, 73% of order emails reach Dynamics 365 as a draft order on their own, and 88% of lines match the right article first time. That matching comes from the order history, which no description could have supplied. Ditte Smit, Marketing Director at SIG Benelux, described the surprise:
"I knew the technology made this possible. My biggest question was the quality of our data. It was a positive surprise to see the order history was enough of a base to get results this fast." Ditte Smit, Marketing Director, SIG Benelux
At Microtechniek, around 60 purchase orders a week land in Ridder IQ without anyone typing them, through evenings and weekends. Anything below the confidence line goes to a person instead of being guessed at, so the admin's day shifts from typing every order to looking at the few that need a human.
Where does vibe automation still make sense?
Vibe automation is a good fit where a mistake is cheap and the input is predictable: an internal notification, a report pulled together each Monday, a prototype to test an idea before anyone commits budget. It is a poor fit where a wrong record costs a return, a credit note or an unhappy customer.
We would keep the good part of the idea. Describing what you want in your own words is the right way to start. Our agents are briefed the same way, in your words, with your exceptions and your tone. The difference is what happens after the brief: the agent is then corrected on real work until the team trusts it, and it keeps every correction.
If you are choosing between categories of product, our comparison of AI workflow automation tools for operations sets out where general connectors, AI builders and operations agents each fit.
How do you move from a described workflow to one that runs in production?
Start with one process that has volume, a clear endpoint in the ERP and rules two people can explain out loud. Run the agent alongside the team on real documents until its output is dull, set the threshold together, and let it write on its own only when you choose.
- Pick one process and one channel. Order intake from the central mailbox is a common first choice. SIG Benelux started with one location, one process and one mailbox.
- Collect real documents, not examples. A few weeks of the actual inbox shows the formats and exceptions a prompt never would.
- Shadow run. Let the agent process work next to the team, so everyone sees where it fails before a customer does.
- Set the line. Decide per process which confidence lets the agent create the record and which waits for a person. Master data changes and purchase orders can stay with a person for good.
- Move it when the numbers say so. After a run of clean weeks, lower the share that needs review. One setting returns the process to full supervision if you need it.
For order intake this is what our order management agents do, from the mailbox to a sales order in the ERP. For the wider picture of how ERP work gets done without retyping, read ERP automation in 2026.
Frequently asked questions
What is vibe automation?
Vibe automation is building an automated workflow by describing it in plain language and letting AI assemble the steps, connections and logic. The term comes from vibe coding, Collins Dictionary's word of the year for 2025. It works well for prototypes and simple flows, and less well for back office work with messy input and many exceptions.
Is vibe automation the same as an AI agent?
No. Vibe automation is a way to build a workflow: the description produces a fixed set of steps. An AI agent runs a job. It reads each document on its meaning, checks it against your systems, scores its confidence per field and hands the uncertain cases to a person. Many agents start from a plain-language brief.
Can vibe automation handle order entry into an ERP?
It can handle the clean orders that match the example in the prompt. Real order entry includes lines without article numbers, new formats and customer habits nobody wrote down. That needs order history, business rules and a confidence threshold. Without those, the workflow either stops or writes a wrong sales order into the ERP.
Why do automation projects fail after a good demo?
A demo proves that the expected case works. Production brings every other case, and someone has to handle them. Gartner names escalating costs, unclear business value and inadequate risk controls as the reasons agentic AI projects get cancelled. In our experience all 3 come from exceptions that were never designed for.
How long does it take to get an AI agent running in production?
It depends on the process and on how the ERP can be reached. SIG Benelux went from interviews to a shadow run within a single 4-week project on one mailbox. A cloud ERP with a standard interface is quicker to connect than an on-premise one, which our engineers connect during delivery.
What to do next
Write down the one process your team retypes most, and count how many of last month's documents for it would have matched a single description. The gap between that count and the total is the work an agent has to do, and it is where the hours sit.
We build AI agents that take that work on, connected to the ERP you already run, with your team setting the line on what goes through on its own. Our forward deployed engineers set up the first process with you. Book a demo and bring a week of real orders or invoices: we will show you what the agent does with them.