> Site index: [llms.txt](https://ortelian.com/llms.txt) with all pages and descriptions. # Do you have to replace your workflow automations when you add AI agents? Source: https://ortelian.com/notes/keep-workflow-automations-when-adding-ai-agents/ Markdown: https://ortelian.com/notes/keep-workflow-automations-when-adding-ai-agents.md Author: Laurens Nys Published: 2026-11-05 No. Keep the automations that run fixed steps reliably. Add AI agents only where a step needs judgement, and replace a flow only when it keeps breaking. No. Automations that run the same fixed steps every time, on a schedule or a trigger, should usually stay as they are. Add an AI agent only at the step where a person still has to read, compare or decide. Replace a flow only when it keeps breaking on cases it was never built for. The work is to map which steps are fixed and which need judgement, not to start over. ## Why the question comes up Teams that already run their own automation flows ask this early. On calls with one client, a consultancy, the people who built the flows wanted to keep them, and the question was whether an AI project would throw that work away. The same client asked whether the process map they had already drawn was the starting point. Both worries are fair. Rebuilding something that works costs time and trust, and the people who built it know the exceptions best. ## Fixed steps and judgement steps A workflow automation follows the path someone wrote down: when this happens, do that. It is cheap, predictable and easy to check. An AI agent gets a goal and limits and chooses the next step itself. That is useful where the path depends on context, and wasteful where it doesn't. Anthropic's guide to building agents makes the same split: workflows follow predefined paths, agents direct their own steps, and the advice is to use the simplest option that works and add agents only when the task needs that flexibility ([Anthropic, 2024](https://www.anthropic.com/research/building-effective-agents)). Our note on [AI agents vs RPA](https://ortelian.com/notes/ai-agents-vs-rpa/) draws the same line for scripted bots. | Step in the flow | Keep it as automation | Give it to an AI agent | | --- | --- | --- | | Input | Same shape every time (a form, a webhook, a nightly export) | Free text, attachments, or records that don't line up | | Decision | A fixed rule anyone can write down | Depends on context across several systems | | Exceptions | Rare, or easy to park for a person | Frequent, and each needs reading | | When it fails | Fails loudly, and you fix the rule | Needs evidence a person can check | | Example | Copy a signed order into the ERP | Work out which customer and order an unclear email is about | ## How to add AI agents without breaking what works 1. **List the flows you run today.** For each one, write down the trigger, the steps and where a person still steps in. The places where someone reads, compares or decides are the candidates for an agent. 2. **Leave the fixed steps alone.** If a step has a fixed input and a fixed output and rarely fails, it stays an automation. An agent there adds cost and makes the result less predictable. 3. **Put the agent at the judgement step only.** The automation hands over the case, the agent does the reading or matching, and the result goes back to the fixed steps or to a person. Write down what a good result looks like first, using real cases, as in [agree the standard before you build](https://ortelian.com/notes/agree-the-standard-before-you-build/) . 4. **Keep approvals where the flow leaves the company.** Sending an email or an invoice should wait for a person until you trust the step. How to set that up is in [approvals and control for AI agents](https://ortelian.com/notes/approvals-and-control-for-ai-agents/) . 5. **Replace a flow only when it keeps breaking.** A flow that keeps growing new branches for exceptions, or that people work around by hand, is a sign the step needs judgement. Rebuild that one, and only after you have looked at whether the work itself should change first ( [redesign the work before you automate it](https://ortelian.com/notes/redesign-the-work-before-you-automate-it/) ). ## What tends to go wrong - **Starting over.** Rebuilding every flow as an agent throws away tested logic and the knowledge of the people who built it. - **An agent on a fixed step.** A step that was cheap and predictable becomes slower, more costly and harder to check. - **Automating the workaround.** When people already handle exceptions by hand, wrapping the workaround in an agent hides the problem instead of fixing the process. Lisanne Bainbridge called this kind of trap an irony of automation: the hard cases are left to people, who now see them less often ( [Bainbridge, 1983](https://doi.org/10.1016/0005-1098%2883%2990046-8) ). - **No record of what the agent did.** If the agent's step is not recorded with its sources, nobody can tell whether a wrong result came from the flow or the agent. ## How Ortelian approaches it We spend time on site observing the work and talking it through with the people who do it, and we read the records in your systems and your written procedures. That shows which steps are fixed and which need judgement. Then we find the constraint and redesign the process before anything is built ([how we work](https://ortelian.com/how-we-work/)). On [the platform](https://ortelian.com/platform/), the two kinds of work are kept apart. Workflows handle the fixed steps that run the same way every time, on a schedule or a trigger. Agents take on ongoing work that needs judgement, act in your systems and ask a person where the work needs approval. Each agent gets only the tools its job needs. Actions that leave the company, such as sending an email or an invoice, wait for a person by default, and every action is recorded with the sources behind it. We build with your team, test on real examples and deploy what meets the standard, in the tools your team already uses. After launch, your team runs it or we keep running it for you. Either way, we host and operate the platform. ## Questions people ask ### Do AI agents replace tools like workflow automation? No. They do different jobs. Automation runs a known path reliably. An agent handles the steps where the path depends on context. Most companies need both. ### Where should the first AI agent go in an existing flow? At the step where a person currently reads, compares or decides before the flow can continue, and where that happens often enough to matter. ### Is it cheaper to rebuild everything as one AI agent? Usually not. Fixed steps are cheap and predictable as automation. Running them through an agent adds cost and makes the result harder to check. ### When should we retire an existing automation? When it keeps breaking on exceptions, keeps getting new branches, or people routinely work around it by hand. Those are signs the step needs judgement or the process needs to change. ## Sources - Anthropic, [Building effective agents](https://www.anthropic.com/research/building-effective-agents) , December 2024. - Lisanne Bainbridge, [Ironies of automation](https://doi.org/10.1016/0005-1098%2883%2990046-8) , Automatica 19(6), 1983. - Ortelian, [Platform](https://ortelian.com/platform/) , [How we work](https://ortelian.com/how-we-work/) and [Trust](https://ortelian.com/trust/) .