Utbenat: Automation

Introduction

Sebastian:
Picture yourself lying on the sofa on a Tuesday evening. You've just started a film, but the evening sun is glaring right at the edge of the TV screen. Getting up, walking three metres and pulling down the blind takes about four seconds. But… it's just so incredibly boring.

So what do you do? Well, on Saturday you go out and buy motorised blinds, a smart hub and light sensors. You think: "Never again will I have to get up off this sofa."

You spend the whole weekend hooking it all up. You write little rules in an app: if it's after seven, and the TV is on, and the sun is shining… then the blind comes down. When it finally works, you feel like a genius!

But then a month goes by. The app gets an update. Suddenly the blind drops in the middle of the morning while you're on a video call. A sensor loses its network connection. You spend an hour here and two hours there troubleshooting, changing batteries and rewriting the rules. If you do the maths, you've spent twenty hours and a lot of money to avoid a movement that takes four seconds.

Today we're talking about those four seconds on the sofa and the weekend that disappeared. We're talking about automation. More specifically, how the arrival of AI and autonomous agents is changing what can actually be automated in our companies. But above all, we're going to break down why AI is only a complement to the automation we already have, why it's ridiculously easy to lose money on your automation, and who actually carries the responsibility when the machines drive into the ditch.

The maths that doesn't add up

Sebastian:
Because why do we automate in the first place? The basic idea is that we want to save time. We want to free up hands and minds from repetitive work so we can focus on better, more valuable activities.

But we humans have a weakness. We tend to want to automate things not because it makes business sense, but because the task is unstimulating. It's boring. And when something is boring, it's incredibly easy to convince ourselves that we should build it away.

The problem is that it's ridiculously easy to spend more time getting an automation to work than you'll ever earn back. And now that we're bringing AI and agents into the equation, the playing field changes, for better and for worse.

Suddenly we can automate tasks that used to require human judgement. Things that don't follow rigid "if A then B" rules. But it comes with a price tag that's easy to forget. Every time an AI agent spins up in a flow, it costs a little. Maybe it's cents or a few kronor per run, but at high volumes it adds up. And even more important: AI models are perishable goods. Vendors update them, tweak their behaviour a little, or retire old models altogether. An AI automation today is not "set it and forget it". It needs ongoing maintenance. And so the gain has to be big enough to cover the build, the running costs and the ongoing maintenance.

The automation we already have

Sebastian:
But let's rewind a bit and look at what reality looks like in our companies today. Because sometimes the industry talk makes it sound like automation was invented the same day we got large language models.

The truth is that every company already has loads of automation. Think about your integrations. When a customer clicks buy in the online store, and that order automatically flows into your ERP system, reserves an item in Josefin's warehouse and sends a pick list to Lasse… that's pure automation. When a supplier types product dimensions into a portal and it syncs into your PIM system. Automation.

This traditional, rule-based automation is the backbone. It's fast, it's practically free per run, and it does exactly the same thing every time. AI and agents aren't coming in to replace that. They're a complement. They step into the gaps where the rigid rules fall short, like reading an unstructured email from a customer, understanding what the customer actually wants and preparing the case before the traditional automation takes over.

The human in the loop

Sebastian:
But when we start building these AI-driven automations, something interesting happens to our need for control. Because an AI isn't one hundred percent predictable, we feel a little uneasy. We don't quite trust it not to drive over the tulips in the flowerbed.

So what do we do? We build what's called "human-in-the-loop".

The AI does the heavy lifting. Maybe it reads in the supplier's product sheet, translates the text and suggests a categorisation. But then the flow stops. The system sends a notification to Anna and asks for a sign-off. A manual approval before the data is allowed to move on.

And at first this feels great. It gives us a sense of control and ownership. But… if we think about it for a moment. Let's compare it with our regular integrations. Could we, in the long run, ask Johan and his integration team to sit and check every single order sent between the online store and the ERP system, and press "Approve" manually?

Of course not! The whole point of automating the flow would be gone. And yet that's exactly where many companies stop with their AI automations. You build a robot, but you still make an employee walk two metres behind it holding an emergency stop button. In the long run we can't work with human-in-the-loop for high-volume flows. Then we haven't saved any time, we've just turned our colleagues into full-time reviewers for the machines.

Who owns the outcome?

Sebastian:
So why do we get stuck in this loop where a human always has to press the button? Often it comes down to an uncomfortable question nobody has really dared to answer: once we've fully automated something… who owns the outcome?

As long as Anna presses "Approve" at the end of the flow, it's simple. If something goes wrong in the product catalogue, or a customer gets a strange answer, well, Anna let it through. She owns the outcome.

But the second we remove the manual step and let the AI agent send the data on by itself, a vacuum appears. If the AI sets the wrong shipping class on a hundred products and the margin gets eaten up by shipping costs… whose fault is that? Is it Johan in IT who connected the API? Is it the consultant who wrote the prompt six months ago? Or is it the product manager?

If it isn't crystal clear who in the business is responsible for what the automation produces, nobody will dare remove the manual approval either. The fear of being left holding the hot potato means we keep the human as an expensive rubber stamp in the flow.

From in the loop to on the loop

Sebastian:
So how do we move forward? How do we get the real time savings from AI automation without losing control?

The answer is that we have to move from human-in-the-loop to human-on-the-loop. The human on the loop, not in it.

The difference is crucial. Instead of having a person review every single run, we set clear frameworks, rules and thresholds around our automation. If the AI agent handles an incoming case or cleans up a product text, and a separate validation shows that everything looks complete, all mandatory values are reasonable and the confidence margin is high, the flow simply carries on by itself. No human needs to look at it.

But… if we don't meet our thresholds. If a figure deviates, if the supplier's input is ambiguous, or if the AI flags that it's unsure, then the case is escalated to a human. The human steps in when it's hard, not when it's routine.

And for this to work in practice, it takes exactly the same thing as our classic integrations: clear ownership. Someone in the business, not in IT, has to be the system or process owner for the automation. Someone who owns the thresholds, follows up on quality over time and has the mandate to say: "This margin of error is acceptable for the time we save."

Cruise control on the motorway

Sebastian:
Think of the difference as driving a new or an old car on a long road trip.

Working fully manually is sitting there accelerating and braking yourself for five hundred kilometres. It works, but it's pretty tiring.

Human-in-the-loop is like switching on an old, non-adaptive cruise control. Every time you approach another car, you have to adjust manually. Sure, it gives your right foot a bit of a rest, but you're constantly fiddling with it.

Human-on-the-loop is how adaptive cruise control is meant to work. You set the speed and the distance to the car in front, your thresholds. Then the car handles the accelerator and the brake entirely by itself ninety-nine percent of the time. You're still responsible for the journey. You're in the driver's seat, keeping an eye on the road and ready to take over if roadworks appear. The car does the heavy lifting within the limits and escalates to you when reality deviates.

Three questions for Monday

Sebastian:
So here's what you do on Monday. Take a look at the automations or AI initiatives you're building, or planning to start. And ask three straight questions.

First: do the numbers add up? Are we automating this because it genuinely saves a lot of time and money, even after factoring in the cost of every AI run and the ongoing maintenance? Or are we doing it just because the task happens to be a bit boring?

Second: how do we get out of the loop? If you've designed the flow so that a colleague always has to sit and click "Approve" at the end, what would it take for you to dare let go? What thresholds or checks do you need to put in place so that only the exceptions are escalated?

And third: who's in the driver's seat? Decide, in black and white, who in the business owns the outcome of whatever the automation churns out. Because without a clear owner, nobody will ever dare make the cruise control adaptive, and genuinely helpful.

Summary

Sebastian:
To sum up: AI and agents open entirely new doors for what can be automated, but they're a complement to your existing flows, not a free ticket. Beware of spending a hundred hours automating away four seconds on the sofa, and don't fall into the trap where people act as rubber stamps for the machines. Set clear thresholds, clarify who owns the outcome, and move the human from sitting in the middle of the loop to overseeing it.

Thanks for listening to today's episode of Utbenat! If you want to talk more about today's topic, or just talk data, get in touch. We're on LinkedIn. Or contact us through our website, Fiwe.se, spelled with a W. Talk soon, and take care out there!

You've been listening to Syntra, a podcast from Fiwe.