AI Can Wait. Go and Look.
By Jonas Astrup Beck · April 2026
Part 3 of a series on AI and the systems perspective. Part 1 · Part 2
I ended the last article with a question: do you actually know why people contact your organisation — and how much of that contact should never have happened?
This article is about what happens when someone actually goes and finds out. And the answer, in every system I have studied, is the same: the system is producing exactly what it was designed to produce.
The sceptic
Vanguard once worked with Hayley. She was a leader in a large insurance company. She had read one of John Seddon’s books and attended a training course on the systems perspective, but she was deeply sceptical. Hayley had built her career as a top-performing salesperson. When she became a leader, she coached and trained her own team in the same methods that had made her successful. She received recognition for that too. She was convinced — genuinely convinced — that her work with her salespeople made a real difference.
So when she was presented with Deming’s idea that 95% of performance is attributable to the system and only 5% to the individual, it felt like a personal challenge. Everything she believed about her own career told her otherwise.
Hayley left the training determined to prove the idea wrong. She would demonstrate that it was the salesperson who made the difference — that individual skill determined whether the customer bought the insurance or not.
She decided to listen to sales calls. Ten. Then 20. Then 30, 40. After 50 calls, she had to face what she had heard: in only 2 could she honestly say it was the salesperson who made the difference.
The customers were calling to buy. The decision had already been made before they picked up the phone. In Hayley’s sample, the individual made the difference in 4% of cases. Not even 5%.
She had set out to disprove a rational argument and ended up disproving her own assumptions instead. That is what happens when you go and look.
This happens more often than not. So go and have a look at your own system. You might find it works brilliantly from the customer’s perspective. Or you might not. Either way, then you know.
What a million calls actually tell you
Hayley’s experience is one person’s revelation. But what happens when you study the whole system?
In the same industry — private insurance — a company receiving one million customer calls a year decided to study the nature of their demand. Not opinion. Not management reports. The actual calls, classified one by one.
Management insisted on a large sample — over four thousand demands classified one by one. Far more than you technically need. The pattern was clear after a few hundred. But sometimes people need more to be convinced.
The management logic until that point had been straightforward: get people to take more calls and sell more. Reduce average handling time. Standardise the process. Send difficult work to back offices so the salespeople can focus on selling. Explore outsourcing. Use IT to replace, support, or control employees. In the last article, I described this as command and control logic. This is what it looks like in practice.
What the study found turned that logic upside down.
70% of all demand was failure demand — people calling because something had not been done or had not been done right, from the customer’s perspective. Only 30% was value demand. And within that value demand, just 0.56% were actual sales — situations where the employee created a need and persuaded the customer to buy.
The purpose of the system, from the customer’s perspective, was simple: give me peace of mind. What matters: understand my need, give me the right product, make sure I understand it, charge me a fair price.
What the system was designed to produce was something else entirely.
The system was doing exactly what it was designed to do
The salespeople who earned the highest bonuses were doing exactly what the system rewarded them for: push products, apply maximum discounts, close the sale, move on. Skip the complicated stuff — the failure demand, the confused customers, the mismatched policies. Those calls took too long and did not count toward the bonus.
The person who actually listened — who understood the need, who took the time to get it right first time — was ranked lower. Penalised by the measures. Doing the right thing, and the system telling them they were underperforming.
This is not a story about bad people. This is a story about system conditions.
Sales targets ranked employees against each other. Bonuses steered behaviour toward volume, not value. Quality was checked retrospectively, per employee, per transaction — it could not see the systemic pattern. The organisation was built around functionally specialised units. The IT system was designed to handle transactions. The budget focused on unit costs per transaction per department. Even the interpretation of legislation had led to terms and conditions that nobody could understand.
Every one of these conditions was a management decision. And together they produced 70% failure demand, an estimated 90% failure work, and a system that penalised the very behaviour the customer needed most.
This is what Deming meant. 95% of performance is attributable to the system. Not theory. An observation you can make in your own organisation — if you are willing to go and look.
A different system, the same lesson
You might think this is an insurance problem. Or a sales problem. It is not.
In adult social care — home care for elderly people — the traditional system works like this: someone other than the care worker plans what should happen in each visit. The plan specifies tasks and times. When the care worker arrives at someone’s home, they follow the plan.
The problem is that the plan rarely matches the actual need. The person did not sleep last night because they are worried about an upcoming hospital appointment. They do not need help cleaning the bathroom right now — they need someone to sit with them over breakfast and talk about what is frightening them. But the plan says: clean bathroom. Prepare breakfast. Move on.
The care worker who follows the plan — ticks the boxes, moves on — looks good on the route schedule. On time. Efficient. The care worker who sits down, prepares breakfast properly, listens to what is actually going on, and maybe helps with something unplanned — that person runs over time. In the system’s eyes, they look like a bad performer.
And here is the opposite, which is equally true: quite often, the care worker arrives and there is no real need for what is planned. The person is fine today. But the plan says they should be there, so they stay — or they leave and wait in the car until the next scheduled visit. The irony is sharp: a planning system whose stated purpose is to optimise time and the public purse ends up wasting both.
Nobody’s fault. The system promotes both behaviours.
The purpose of the visit is bigger than the planned task, and the planned task rarely fits the actual need — in either direction.
A municipality in Sweden — Mörbylånga — decided to study their home care system and redesign it from the citizen’s perspective. They drew on the Vanguard Method: studying demand and individual value, understanding what mattered to the people they served, and redesigning the work around purpose rather than tasks. They eliminated central planning. They let care workers plan their own rounds based on actual need. They adjusted thermostatically — more help when it was needed, less when it was not — rather than delivering the same standardised package regardless of what they found when they walked through the door.
So what happens when you build the organisation around purpose? The number of different care workers visiting each person per month went from 22 to 12. Their ranking in Sweden’s national home care index went from 104th to 9th. Overtime costs halved — from 1.97 million kronor in 2022 to 753,000 in 2024. Total home care costs declined, despite wage and price increases. Staff morale improved. And the number of alarms from residents fell — because when you know the person helping you, and they know you, trust replaces anxiety.
Something is working in Mörbylånga. And they know why.
Two completely different sectors. Same discovery. The system conditions — measures, targets, functional design, the separation of planning from doing — were producing the very behaviour everyone was complaining about. In both cases, the people inside the system were doing the rational thing given the constraints they were in.
The technology that did not help — and the one that might
Here is the part most people do not expect.
Mörbylånga did not digitise their way to these results. They stopped using their IT planning system. They turned it off and saved money doing it. They replaced it with something as simple as Word documents and team-based planning in groups.
Why? Because the IT system was built on the same logic as the management thinking it served. Minute-level control. Standardised tasks. Functional specialisation encoded in software. The technology reflected the thinking. Automating it just made the wrong thing more efficient — which is exactly what I warned about in the first article.
This is not an argument against technology. It is an argument about sequence. You have to understand what the work actually needs before you know what technology to use. Mörbylånga understood — through studying the work, through listening to the people they served, through experimenting — that they needed flexibility, continuity, and the ability to respond to the variety in individual need. None of the existing IT systems could support that, because they were all built on the assumption that the work can be pre-specified and controlled from outside. So the team went to the simplest tool that let them plan from need. And it worked.
Now here is where AI becomes genuinely interesting. Traditional IT systems require rigid, pre-specified logic — which is why they tend to encode command and control thinking whether they intend to or not. AI is fundamentally different. It can absorb variety. It can support flexibility and intuitiveness in a way that standardised systems cannot.
But — and this is the critical point — only if you have first understood what the work actually needs. AI used to automate a planning system built on the wrong logic will just produce wrong plans faster. But AI used to help care workers understand what actually matters to the people they serve — and how that changes over time — that could be genuinely transformative.
The technology that helps is the one that serves a purpose you have already understood. Not the one that automates a logic nobody questioned.
I have experienced this myself. I have tried numerous note-taking applications — digital tools designed to organise and categorise and retrieve my thinking. None of them stuck. I kept coming back to a physical notebook. Paper. Tactile. Flexible. It works. Paper is an amazing technology.
But then I started using AI in my own consulting work. And something different happened. AI did not try to replace the notebook with a better filing system. It helped me do what I had always been trying to do — reflect on my work, access years of accumulated knowledge, keep track of what I was learning and how my thinking was developing. It supported the purpose. It fitted the way I actually work rather than imposing a structure I had to conform to.
I still have my notebook, though. Because putting a computer or a phone in between people affects relationships. When I sit across from a client or a team, paper keeps me present in a way a screen does not. The right technology for understanding the work is sometimes no technology at all.
Your turn
Here is something you can do tomorrow morning. You do not need a consultant. You do not need a project. You need curiosity and ten minutes.
Look at your next ten emails. For each one, ask a single question: is this a consequence of something that has not been done or has not been done right?
A customer chasing a response they should have had last week. A colleague escalating something because nobody could resolve it at the first point of contact. A report that exists because nobody trusts the process it is supposed to verify. An internal request for information that should have been available from the start.
If you recognise these patterns — you are looking at failure demand from customers or failure work inside your own system. And you are seeing, for the first time, what your system actually produces.
But here is what matters most. Do not blame the sender.
That is the lesson from the insurance company, from the care workers, from Deming’s 95/5. The people in the system are responding rationally to the conditions they work in. Ask the follow-up questions instead. How easy was it for them to get it right? What are they being measured on? Did they have everything they needed?
Those questions will tell you more about your organisation than any AI dashboard, any benchmarking report, or any best-practice framework.
And here is a question worth sitting with, once you have looked: based on what you are seeing — would AI remove the causes? Or would it just handle the waste faster?
Now it is your turn.
In the first article, I said: understand the problem before you automate. In the second: the system conditions determine everything, and AI amplifies whatever you are already doing. In this one, I am saying: go and look. The answers are in the work — and that should inform your data, not the other way around. The people inside the system are not the problem. They are responding rationally to the conditions they work in.
That raises a harder question. If the system shapes the behaviour — and the people inside it are doing their best within the constraints — what is the system doing to them? That is where the next article begins.
If this resonates — if you are a leader who is willing to go and look before you automate — I would love to hear from you.
jonas@vanguard-consult.dk · LinkedIn · vanguardmetoden.dk
This is Part 3 of a series on AI and the systems perspective by Jonas Astrup Beck. Read Part 1: AI Won’t Save Your Organisation and Part 2: AI Does Not Care What It Amplifies on Substack.
This article was written with AI assistance — which, given the argument about understanding before automating, felt like the right thing to acknowledge.
Sources
Hayley’s story is retold from Riv Servicefabrikkerne ned (Jonas Astrup Beck and Kristian Astrup Nielsen).
W. Edwards Deming’s observation on system vs individual performance appears in Out of the Crisis (MIT Press, 1986). He used varying ratios over his career — 85/15, 94/6, 95/5 — and later estimated the system proportion was even higher.
Mörbylångamodellen: mörbylånga.se and dela.skr.se (Sveriges Kommuner och Regioner).