Automation · AI agents
Workflow or AI agent: when should you follow a procedure, and when should you investigate?
Published7 min read
A supplier confirms a maintenance visit. Should you simply put it in the schedule, or check who will step in if the tills fail to restart? Here is how I would divide the work between a workflow, an AI agent and a human decision in this situation.
The article in 30 seconds
- A workflow is a good fit when the steps are known, even if AI helps read incoming requests.
- An agent becomes useful when each document it consults determines what to check next.
- Measure the cost of handling a case correctly, including human review and maintenance.

“Maintenance confirmed for Saturday at 6 pm.” The supplier has replied. The slot is booked. Everything looks ready.
But the shop closes at 8 pm. If the tills fail to restart, who will answer the phone?
This fictional scenario captures the choice between a workflow and an AI agent. Logging the appointment follows a procedure. Checking that the work can go ahead without leaving the shop stranded means putting several pieces of information together.
A procedure is enough to log the date
With n8n, I can build a simple workflow: receive the email, find the shop and ticket number, check the date and update the schedule. If any information is missing, the case goes back to a person.
This is a workflow : the steps and the conditions for moving between them are defined in advance. AI can help interpret a free-form message. It is unnecessary when a form already provides all the required fields.
What if the supplier writes “late on Saturday”? The workflow asks for a specific time. It handles a known variation without having to work out what to do next.
Finding the risk means opening the case file
I would give an agent like Hermes a different assignment: “Check what could prevent the shop from resuming operations after this maintenance visit.”
In our example, its investigation might go like this:
- The ticket says the tills are due for maintenance. The agent checks the shop’s opening hours.
- The shop is open until 8 pm. It looks up who provides support during those two hours.
- The contract says on-call support has to be booked. It looks for a confirmation in the case file and cannot find one.
A useful response takes only a few lines:
“The maintenance slot is confirmed, but I couldn’t find a booking for on-call support. The shop will still be open. Can you confirm who will handle a till failure between 6 pm and 8 pm, and what number to call?”
Each finding determined the next check. This is where an agent adds value. It gathers the evidence and drafts the outstanding question. The person responsible decides whether the work can go ahead.
The distinction matters: failing to find a booking does not prove there is none. A good report should say so.
The last three cases test
Before choosing a tool, I would ask the person doing the work to walk me through their last three difficult cases.
- “I always check the same things.” I would start with a workflow, adding an AI step if requests come in as free text.
- “The first document tells me what I need to look for next.” I would try using an agent to prepare that investigation.
The distinction is about how work is organised. n8n also supports agents and can require human approval for certain tool calls. [1] Choose based on how you need the work to flow.
What do we actually gain across 50 cases?
Set aside subscriptions and thousands of tickets for a moment. Consider just 50 difficult cases to review each month, like the maintenance example. The figures below are fictional and show what to measure in a trial.
Without the agent: 6 hours 40 minutes of work
A person takes 8 minutes per case to find the documents, read them and make a decision. Across 50 cases, 50 × 8 minutes = 400 minutes, or 6 hours 40 minutes a month.
With the agent: 1 hour 40 minutes of review
The agent gathers the documents and flags what is missing. Suppose the person now needs only 2 minutes per case to check the agent’s work and make a decision. Across 50 cases, 50 × 2 minutes = 100 minutes, or 1 hour 40 minutes a month.
The difference is five hours freed up each month. The example hinges on one assumption: two minutes must be enough to produce a result just as reliable as before.
What are those five hours worth in euros?
If we value staff time at €45 an hour for this example, five hours are worth €225. If the agent’s investigation costs €0.18 per case, 50 investigations add €9 in AI usage.
€225 in staff time − €9 in AI usage = an estimated benefit of €216 per month. This excludes design, hosting, subscriptions and maintenance. Costs that are identical under both options do not affect the difference.
The €216 is an estimate of the value of the time recovered. If salaries stay the same, the team mainly has more time for other work. Cash savings arise only when an expense is actually avoided, such as overtime.
What if correcting the agent’s work takes as long as the time it saves? The benefit disappears. I would measure the time spent searching, reviewing and correcting, as well as the speed of the AI response.
Give the agent permission to investigate and keep decisions with people
For Saturday’s maintenance, three rules can define the agent’s remit:
- Read the relevant documents. Use authorised tickets, opening hours and contracts, with their dates. If two documents conflict, present both.
- Prepare a decision. Explain the risk, cite the documents and suggest the question to ask. Sending a message, booking on-call support or rescheduling the visit requires specific approval.
- Know when to stop. If the evidence remains unavailable or the investigation reaches its budget, report what you found and flag what is missing.
Tool permissions and approval checks must enforce these limits. An instruction written in an email must be unable to expand them. If contract excerpts go to an external model, the company must also authorise that transfer.
Should the AI run every Friday?
Only when there is a reason to investigate. Hermes lets a script first check whether the agent needs to run. [2] In our example, this check could flag a change to a case or an approaching deadline. If there is nothing to investigate, it avoids an unnecessary model call.
In my setup, Hindsight stores memories and Activepieces provides access to applications. A previous decision can help you find where to look. Authorising the Saturday maintenance still requires a confirmation for this specific case.
Start with twenty cases before expanding to the whole department
I would take twenty closed cases and withhold their resolutions from the agent. I would include routine cases, a missing document and conflicting evidence to see whether the agent actually helps the person making the decision.
I would look at three things: did it spot the risks, how long did the human review take, and did it stay within its permissions? I would repeat the trial on difficult cases. A single good answer is insufficient.
For a team handling many requests, the two approaches can work together. The workflow receives and sorts requests, the agent examines selected exceptions, and a person approves the action. Meta describes a similar split between an orchestrator responsible for fixed steps and agents that handle domain-specific reasoning. [3]
I would keep the workflow for clear procedures and trial an agent where investigation is needed. What if the agent finds the same on-call support problem every week? I would add that check to the procedure. The agent’s work would then have made the next case easier to handle.