If you lead operations or IT, someone has probably asked you this: should we let coding agents build our automations, or should we use a low-code tool? Coding agents now generate a lot of software very quickly, so it is a fair question.
For most business process automation, our answer is to start with low-code AI automation. Coding agents are powerful, but the code they produce is hard to maintain and hard to hand over. Low-code tools sit between flexibility and ease of use. Combined with today’s large language models (LLMs), they can automate transactional processes end to end, and you can validate an idea in minutes instead of months.
Below we explain why, which workflows fit, what it costs and what to do first.
In this video, Hassan Ghiassi and David Grimm from Vylos discuss the arguments in this article in about 15 minutes.
Should your business use coding agents or low-code tools?
We recently ran a webinar on Claude Code and how to get started coding in a terminal. It gave us a clear lesson about what business leaders actually need.
David Grimm does not argue that coding agents are weak. Using them in Claude Code or other harnesses such as pi or opencode is very powerful, and you can generate a lot of software.
The problem comes afterwards. That code is not easy to maintain, and it is not nice for someone else to maintain and take over. In business process automation this matters, because the workflow has to be understood and changed by people other than the one who built it.
Flexible or easy to use: why you rarely get both
David applies a basic rule from product management to this choice:
“Your tool can be flexible or it can be easy to use. You cannot have both usually and this is true for coding agents.”
David Grimm
Coding agents give you maximum flexibility. You pay for it with complexity, because you own and maintain the code.
Low-code tools sit in the middle. They are still very flexible, and a workflow can be handed over and understood instantly, without being a really good developer. For most operations teams, that is the right trade-off.
Why low-code AI automation has momentum right now
The market is moving in this direction. David points to n8n, which he says is valued at over five billion euros, “which is a lot for a European startup.”
The bigger shift is what the tools can do. For years, low-code platforms promised that you could simply describe a process and get an automation. Now you can tell them to set up a flow. Many have AI agents you can talk to that build the flow for you. As David says, that is “what we wanted to have five years ago.”
His short verdict: “I think they deliver now what they promised for the last years.”
Which workflows are a good fit for low-code automation tools?
Classify first, then route
One workflow architecture covers a surprising number of cases. First, classify the incoming process. Then route it to a subworkflow built for that specific process.
David is confident about this pattern: if you master this idea, he has not yet seen a flow it could not handle.
Integrations are the strength, and connected systems are a commodity
These tools started as integration platforms as a service. Their out-of-the-box integrations with the systems you already use are a big part of why they work so well.
Having APIs and connectable systems is now a commodity. It is usually not where the conversation should start. But if your systems are still not interconnected, you really should work on that first. Our AI automation implementation work covers exactly this kind of system sync.
Example: transaction checks in a finance team
Picture a CEO at a finance company whose independent accounting team manually checks transactions against another platform. It is repetitive, rule-based and time-consuming. In David’s view, this kind of task can now be automated end to end.
Can LLMs reliably automate document processing?
Yes. In our experience, the current generation of LLMs has largely solved document processing, meaning OCR and document understanding.
Before, you had to train your own engine with thousands of documents and deal with bounding boxes. Bounding boxes are the coordinates that tell an extraction engine where each field sits on the page, which older systems had to learn from labeled examples. This was complex and expensive.
Now any frontier model understands the document, and you can ask it to extract what you need: “give me all the payments on this purchase order,” or all the data from an invoice, such as tax ID and tax numbers.
Structured output makes it safe to automate
“The structured output which is … the killer feature for every large language model.”
David Grimm
Structured output means the model returns its answer in a fixed format you define, such as JSON with named fields, instead of free text, so the next step in the workflow can read it without a person in between.
Output like JSON can be handled automatically by the rest of your workflow. You always want to validate that output before it moves on.
What you no longer need: retraining, model deployment and drift measurement. As David puts it, “it’s just prompting now and setting up these workflows in one of these tools and like this you can have a big return on invest in no time.”
How long should an AI automation proof of concept take?
Much less than most buyers expect.
The old way meant training classifiers. You needed clean data labeled by specialists, a first model, testing in production, finding overlapping groups and at least one retraining. For non-trivial use cases, David says, “we are talking 6 months easily.”
Today you can test a prompt against a real document almost while you are still talking with the customer. In David’s words, “what took weeks and months now only takes minutes … to at least validate it.”
To be honest about the limits: validation is fast, but non-trivial cases still need a lot of setup before they run in production. Still, you should see a solution working early. Hassan Ghiassi gives buyers direct advice here:
“If people are telling you it’s going to take six months or a year to deploy something it might not be the company that you want to work with even if it’s … not us.”
Hassan Ghiassi
Do AI automations get stuck in the pilot phase?
This is one of the most common objections we hear. David’s experience is different:
“I hear a lot that people say yeah it looks nice in your … pilot and when you deploy it in production it breaks. This is actually not what I experienced.”
David Grimm
The approach matters. Start iteratively, get into production fast, then add documents and groups step by step. The tools themselves are battle tested and in production globally.
Changes are prompt updates, not projects
When a new edge case, document type or class appears, you update the system prompt. The change applies almost instantly. There is no retraining, no redeployment, no model versioning and no fallback to manage.
One rule: every time you update a prompt, measure and grade the results so you don’t make the automation worse.
How do you pick your first workflow and measure AI automation ROI?
Start by knowing every step of the workflow. This sounds obvious, but the knowledge is often scattered across several people.
Then look for transactional processes: clearly structured and defined step by step. For example, “the document is incoming … the solution is an approved document.”
These processes are also the easiest to justify. You measure the time people spent before, then compare it against the token or model costs after. If you need help mapping the steps and building the business case, that is what our AI consulting is for.
Do you need the most expensive AI models for document automation?
No. You don’t need top models such as Mythos, Opus or Gemini for this work. According to David, you can get away with smaller models such as nano, Flash, Sonnet or Haiku, which have very low token costs.
He also sees low vendor lock-in risk: “nearly all models can perform this now and you could even use open-sourced models and host it yourself for these tasks.”
Should you use autonomous agents in business processes?
Usually not. Agentic fleets are, in David’s words, “usually … not what you want” in business process automation. You want defined workflows.
That does not rule out agents entirely. You can give an agent more agency in some steps, as long as you always validate the outcome.
David draws a clear line. For personal use, you can give an agent access to your inbox or Gmail account. In a business process, “that’s absolutely impossible.” Low-code workflows work because “the setup is quite strict which you need in a business process context.”
Where should operations leaders start?
After our Claude Code webinar, Hassan’s takeaway was simple: “we should actually simplify things even a little bit more because many leaders are not ready to code or things like that.”
So start where the effort is low and the payoff is visible:
- Pick a low-code tool such as n8n, Zapier, Make or Workato.
- Focus on manual, repetitive tasks.
- Map every step of the workflow before you automate it.
- Choose a transactional process so you can measure ROI.
- Use structured output and validate every result.
- Go to production early, then expand step by step.
If your team wants hands-on practice first, our AI workshops and courses are a good starting point.
FAQ
Should my business use coding agents or low-code tools to automate processes?
For most business process automation, we recommend low-code tools. Coding agents in Claude Code or similar harnesses are very powerful and can generate a lot of software, but that code is hard to maintain and hard for someone else to take over. Low-code tools such as n8n are still very flexible, and a colleague can understand a workflow instantly without being a strong developer.
How long should an AI automation proof of concept take today?
Initial validation can take minutes. With a frontier model you can test a prompt against a real document almost while talking with the customer. Before LLMs, classifier projects easily took six months. Non-trivial cases still need real setup work, but if a vendor quotes six months or a year to deploy something, Hassan Ghiassi suggests it may not be the right partner.
Do I need the most expensive AI models for document automation?
No. For document automation you can usually get away with smaller models such as nano, Flash, Sonnet or Haiku, which have very low token costs. David Grimm sees low vendor lock-in risk, because nearly all models can do this work now. You could even self-host open-source models. Whichever model you choose, always validate the output.
How do I measure ROI on an AI automation?
Pick a transactional process that is defined step by step, for example an incoming document that ends as an approved document. Measure the time people spent on it before. After automation, compare that against the token or model costs. Because the steps are clear, the before-and-after comparison is simple to make and easy to defend.
Should I use autonomous agents in business processes?
Usually not. Business process automation needs defined workflows with a strict setup, not autonomous, non-deterministic agents. You can give an agent more agency in some steps, but you always validate the outcome. Giving an agent access to your own Gmail for personal use is fine. In a business process, David Grimm calls that “absolutely impossible.”
Find your first workflow to automate
Not sure which of your processes is the right first candidate? Book a free AI automation audit and we will look at your workflows with you: book your free AI automation audit.