ai agent

AI agent or workflow automation: what your process needs

AA
Aurum Avis Labs Author
9 min read

AI agent or workflow automation: which one your process needs

Do you know the steps before the work arrives? A fixed workflow is usually enough: low cost per run, the same result, easy to check. If the path must be found anew for each case, an AI agent can help when a person reads the output. Most office work in a Swiss SME sits between those poles: a fixed workflow with one AI step where someone still reads and retypes.

This is for owners and office leads offered «AI agents» who want to know what their process needs. It compares three designs on the same criteria. What an agent is in the first place is explained in what is an AI agent.

The terms follow the line Anthropic draws. In a workflow, models and tools run through predefined code paths. An agent dynamically directs its own steps and tool use. Anthropic’s own guidance: workflows offer predictability and consistency for well-defined tasks, agents fit where flexibility and model-driven decisions are needed.

Three designs side by side

The three designs differ less in what they can do than in how they behave when something goes wrong. The table compares them on the same rows.

Fixed workflowWorkflow with one AI stepAI agent
PredictabilitySame input, same resultRules fixed, AI step usually rightPath and result can vary
Cost per runClose to zero beyond hostingHosting plus one model callSeveral model calls, count varies
Behaviour on errorsStops at the failed step and reportsFixed checks after the AI stepCan follow a plausible wrong path
Audit trailEvery step loggedRule steps clear, AI input and output loggedLog of every tool call needed
Data protectionMoves only the fields you definedAI service sees the document it readsAccess to every system it might consult
MaintenanceChanges when a system or rule changesPlus a check when inputs or model changePlus review when instructions or tools change
Typical examplesMonday report, filing, reminderInvoices in many layouts, shared inboxAnswer across systems, unusual drafts

Fixed workflow automation: when the path is known

A fixed workflow fits when you can describe every step before the work arrives. A trigger starts it: a file lands in a folder, a date is reached, a form is submitted. Then rules act: copy these columns, send this reminder, route to this approver. There is no model involved, so there is nothing to interpret.

That makes it the cheapest and most predictable option. A run costs almost nothing beyond hosting. When it fails, it fails at a named step, and a well-built workflow says so instead of carrying on. A year later you can still reconstruct why a file went where it went.

Typical cases are the Monday report assembled from the same exports every week, described in reporting from five sources, or the reminder that appears when a quote has had no reply for seven days, covered in quote follow-ups.

A rule cannot read a supplier invoice in a layout it has never seen, or an order written as a sentence in an email. You will then either keep a person in the middle or write rules until nobody understands them.

The hybrid: a workflow with one AI step

The hybrid is a fixed workflow in which exactly one step is done by a model, usually the step where someone used to read something and type it off. Everything before and after stays rule-based. The model reads. The rules decide.

Take incoming supplier invoices. The trigger is fixed: a PDF reaches the invoices mailbox. The AI step reads it and returns supplier, amount, due date and order number as structured fields. Then rules take over again: is the supplier known, does the order number exist, is the amount above the second-approval threshold? Anything the rules find odd goes to a person. What passes goes to approval as before.

You do not need a special product for this step. Microsoft Power Automate has a Run a prompt action whose output later actions in the same flow can use. n8n offers both shapes side by side: a Basic LLM Chain node that sets one prompt for the model, and an AI Agent node where the model decides which connected tools to call. The choice between hybrid and agent is a design decision, not a software purchase.

The hybrid fits when one step of an otherwise stable process involves unstructured input. It keeps cost per run low, because the model is called once. It keeps the audit trail readable, because you log one AI input and one output per run.

A QR-bill, a CSV export or a web form does not need a model to be read. Skip the hybrid too when the step that needs judgement is the decision itself, not the reading. Then a person belongs in that step.

AI agent: when the path has to be found

An agent fits when the steps depend on what earlier steps found, and a person reviews the result. An internal question such as «why has this customer not received the delivery yet?» can mean checking the order list, the shipping confirmation in the mailbox and a note in the CRM, in an order that differs each time. An agent can do that search and write one answer.

The price is higher cost and less predictability. Anthropic says plainly that agentic systems often trade latency and cost for better task performance, and that the autonomy of agents brings the potential for compounding errors. One wrong lookup early on can lead to a confident, wrong answer at the end.

For that reason an agent in an SME office should mostly read and prepare. It may search, summarise and draft. Payments, bookings and replies that commit the company go through a person first. The reasons, including the risk of instructions hidden in emails, are in what is an AI agent.

One order inbox, three designs

A made-up example: a wholesaler with around 15 people receives orders by email. Some come as a PDF from the customer’s system, some as a sentence in the email body, some as a photo of a handwritten list.

A fixed workflow can handle the PDFs from customers with a stable layout. Everything else stays manual, so the office still reads most of the mail.

A workflow with one AI step reads all three kinds and returns customer, article and quantity. Rules then check the article numbers against the product list and send anything uncertain to a review queue. The office confirms, and the order is entered. The model never decides what gets delivered.

An agent could go further: look up stock, choose a delivery date and reply to the customer. That last step is a binding reply. For this office it is exactly the step that should stay with a person. The hybrid is the better fit here.

KitchenCrew, a product we built for Swiss restaurants, follows the same principle: the team types a shopping list in a chat, and what arrives in the inbox are order drafts per supplier, not orders already sent.

Data protection and audit trail in Switzerland

Swiss data protection law applies the same way whether a rule or a model processes personal data. The FDPIC states that the Federal Act on Data Protection is written to apply to all types of technology and is therefore directly applicable to AI-supported processing. It also says that users have a legal right to know whether they are speaking or corresponding with a machine.

Agents and AI steps raise one further point. Under Art. 21 of the Data Protection Act, you must inform a person about a decision based exclusively on automated processing that has a legal consequence for them or a considerable adverse effect on them. On request, they may state their point of view and ask for a natural person to review the decision. In practice: let the agent or the AI step prepare, and let a named person decide anything that matters to a customer, an applicant or an employee.

Two practical rules follow for the audit trail. Give an agent access only to the systems it needs, preferably read-only. And log every tool call, input and output, so you can answer «why did this happen?» later. For a fixed workflow, the automation tool log is often enough. For one AI step, also keep the prompt and structured response. Without a tool log, an agent run is hard to reconstruct when a customer or internal audit asks.

Which customer data may go into which AI tool at all is covered in ChatGPT and customer data. This is practical guidance, not legal advice. For decisions about people, ask your data-protection adviser.

Which design to skip

  • If you can draw the path on one page

    Build a fixed workflow. An agent would only add cost and variation.

  • If one step is «someone reads it and types it off»

    Build the workflow and put one AI step there. Let rules check what the model returns.

  • If every case needs a different path and a person reads the result

    An agent can help. Start with drafts and lookups, and keep decisions with a person.

Two limits on the agent remain. At hundreds of similar runs a week, an agent per case is not worth it: a fixed workflow or one AI step is. And if nobody reads the output, the agent must not trigger binding steps.

Before any of the three, the process itself has to pass the filter in which process to automate first: it repeats, it hurts, it has a clear start and end, and the data already exists.

When the Process Check fits

The Process Check fits when you have one recurring manual task and are unsure which of the three designs it needs. In 30 minutes on Microsoft Teams we look at how it runs today. You leave with a concrete case, a price range and a rough duration for the implementation. The answer may well be that a fixed rule is enough and no model is needed.

The check is free. What the half hour includes is described in what the Process Check is.

ai agent workflow automation process automation
AA

Written by

Aurum Avis Labs

Builds and ships at Aurum Avis Labs. Writes here about what we learn working with founders and SMEs in the DACH region.

Customize your cookie preferences. Essential cookies cannot be disabled as they are required for the website to function properly.

Essential Cookies

Required for basic website functionality, security, user authentication, and error tracking.

Always active

Analytics Cookies

Help us understand how visitors interact with our website to improve user experience. Includes Google Analytics and Microsoft Clarity session recordings.

Marketing Cookies

Used to track visitors across websites to display relevant and engaging advertisements.