There is a problem sitting underneath a lot of these agentic AI systems… and I don’t think enough people are talking about it.
We keep trying to automate customer-facing workflows before we have properly captured the reasoning that makes those workflows work in the first place. That is the mistake. Because in a call centre, a medical booking system, a customer service department, or any client-facing environment, the process is never just the process. There is always a human layer sitting underneath it.
That human layer includes tone, timing, judgement, experience, common sense, empathy, and the ability to read what is happening in the interaction. It is knowing when to slow down. It is knowing when the customer is confused. It is knowing when someone is anxious, angry, embarrassed, or overwhelmed. It is knowing when the script technically says one thing… but the situation in front of you is clearly telling you something else.
And this is where I think a lot of agentic AI and customer-facing automation systems are going to struggle. Not because AI cannot automate tasks. It can. The problem is that client relationships are not just tasks. They are interactions. And interactions require reasoning.
The Workflow Is Not the Work
I have worked in call centre environments before, and one thing you learn very quickly is that the script is not the job. The script is just the skeleton. Yes, there is a system. Yes, there are fields to complete. Yes, there are questions to ask. Yes, there is usually a workflow that tells you what should happen next. But anyone who has actually done the work knows that the real skill is in how you move through that workflow with a real person on the other end of the phone.
People do not call in like perfect little process diagrams. They call in stressed. They call in angry. They call in embarrassed. They call in anxious. They call in confused. They call in talking too fast. They call in not knowing what they need. Sometimes they think the problem is one thing, when it is actually something completely different.
That is where the human reasoning starts. A good call centre worker is not just reading from a screen. They are listening. They are adjusting. They are interpreting. They are making tiny decisions every few seconds. Do I slow down here? Do I explain this differently? Do I ask that question now, or later? Do I skip ahead because I already know where this is going? Do I stop and reassure them before moving on? Do I escalate this? Do I follow the workflow exactly, or do I slightly bend it because this situation clearly does not fit?
That is reasoning. And that is the part most automation projects are not capturing.
Agentic AI Will Struggle If It Only Learns the Script
This is where I think a lot of agentic AI systems are being built on shaky ground. They are being trained around workflows, rules, process maps, decision trees, and scripted pathways. All of that is useful, but it is not enough. A system can know the steps and still not understand the work.
A workflow might say: ask question A, then ask question B, then determine category C, then send the customer to outcome D. On paper, that looks clean. It looks efficient. It looks like something that can be automated. But real client interactions do not move that neatly. A customer’s answer to question A might reveal something more important than question B. Their voice might tell you they are not okay. Their hesitation might tell you they do not understand. Their frustration might tell you the process has already failed before you even begin.
An experienced person hears that. They pick up the signal. They know when the customer is not following the conversation. They know when the formal pathway is no longer the best pathway. They know when the “correct” process answer is not actually the right answer for that person at that moment.
That is the gap. And that is where customer-facing AI starts to fall apart.
A Practical Example: Medical Bookings and Real People
I saw this recently with my own mum. She was trying to deal with a doctor’s appointment, and the clinic had one of these automated assistant-style systems that was supposed to help manage bookings and workflows. In theory, I understand why they used it. Reduce admin. Direct people to the right appointment. Make the system faster. Take pressure off staff. All of that sounds good.
But in practice… it caused problems.
The issue was not that automation is automatically bad. That is not the point. The issue was that the system did not seem to understand the reasoning that a real receptionist or call centre worker would apply in that situation. A person may have picked up confusion. A person may have asked a slightly different follow-up question. A person may have realised that my mum was not being directed properly. A person may have paused and thought, “hang on, this does not sound right.”
That is the difference. The automated system may have followed the workflow, but the workflow was not enough. It lacked the judgement layer. It lacked the relationship layer. It lacked the reasoning that sits behind the way experienced people deal with real humans.
And this is the exact issue I think we are going to see everywhere as more organisations rush into agentic AI.
Adaptive Machine Learning Should Learn From the People Doing the Work
This is where adaptive machine learning becomes important. Before we replace people with agentic systems, we should be using machine learning to understand how experienced people actually handle client relationships. Not just what they click. Not just what field they complete. Not just whether they followed the script. But how they reasoned through the interaction.
Imagine a call centre with 200 experienced staff. Instead of immediately saying, “great, let’s automate this”, the smarter move would be to let machine learning observe the work first. Carefully. Ethically. With governance. With privacy protections. With transparency. Not as surveillance. Not as a way to punish staff. Not as some creepy productivity-monitoring tool. But as a way to capture organisational intelligence before it disappears.
Because every call centre worker is holding a piece of that intelligence. They know where customers get confused. They know which questions create frustration. They know when a script does not work. They know how to calm someone down. They know what the system says should happen… and what actually needs to happen.
That is the layer machine learning should be capturing.
The Real Question Is “Why Did They Do That?”
The mistake is thinking that the workflow tells the whole story. It doesn’t. A workflow tells you what happened. It does not always tell you why it happened. It does not tell you why the worker paused. It does not tell you why they asked an extra question. It does not tell you why they changed the order. It does not tell you why they escalated. It does not tell you why they used a softer tone. It does not tell you why they ignored the standard pathway and took a slightly different route.
And yet… that is often where the real value is.
Because that is the reasoning. That is the part that explains how experienced people protect the customer relationship while still getting the work done. That is the part that explains how the organisation actually operates in the real world, not just how it says it operates in a procedure manual.
If agentic AI cannot learn that, then it is not really learning the work. It is just learning the visible process. And there is a big difference between those two things.
The IT Mindset Often Gets This Wrong
This is probably where I will stir the pot a bit… but I think a lot of IT-led automation gets this wrong. There is often this belief that if you map the workflow, you have captured the work (I will admit, I am guilty of this over the last 40 years of my programming work). But you haven’t. You have captured the neat version of the work.
You have captured the version that looks good in a diagram. The version that fits into a system build. The version that can be turned into rules, steps, conditions, and outputs. But real work is messier than that, especially when people are involved.
A customer-facing process is not just a technical workflow. It is a relationship workflow. And relationship workflows depend on reasoning. What is the person really asking? What are they not saying? Are they confused? Are they distressed? Are they misunderstanding the question? Is the system sending them down the wrong path? Does this need human intervention? Do we need to slow down? Do we need to explain this differently? Do we need to stop the automated pathway entirely?
That is not just process logic. That is human judgement. And it needs to be captured before we start handing these interactions over to agentic AI.
Cutting Staff Before Capturing Their Reasoning Is Organisational Memory Loss
This is the part that really concerns me. A lot of companies are now talking about removing thousands of staff because AI can supposedly do the work. Four thousand people here. Six thousand people there. Whole support teams cut down. Whole knowledge layers removed. And on a spreadsheet, it probably looks like efficiency.
But I think there is another way to look at it.
You may not have just removed cost. You may have removed the people who knew how the work actually worked. You may have removed the people who knew the edge cases. You may have removed the people who knew where the system failed. You may have removed the people who knew how to calm customers down. You may have removed the people who could tell when the script was technically right… but practically wrong.
That is not just labour. That is organisational memory.
If you cut that out before capturing the reasoning, then the AI system will have to relearn it later. Usually through customer complaints. Usually through failed interactions. Usually through escalations. Usually through reputational damage. Usually through the business realising too late that the human layer was doing far more than anyone measured.
That is not efficiency. That is memory loss dressed up as transformation.
Better AI Starts By Learning the Human Reasoning First
The better approach is not “don’t use AI”. That is not what I am saying. I am not anti-AI. I am saying we are doing the order wrong.
Before automating the client relationship, we need to capture the reasoning inside the client relationship. Before building the agentic workflow, we need to understand how experienced people actually apply the workflow. Before replacing the call centre worker, we need to learn what that worker knows. Before assuming the process map is enough, we need to understand the judgement that sits behind the process.
That is where machine learning should be used. It should observe patterns. It should identify where staff adjust scripts. It should find where customers commonly get confused. It should detect where the formal process does not match the real process. It should help us understand why experienced people get better outcomes.
That is a much more intelligent use of AI than simply throwing a chatbot in front of a customer and hoping the workflow holds up.
Agentic AI Needs Reasoning Trails
If an AI system is going to act on behalf of an organisation, it should not just complete tasks. It should be able to explain its reasoning. Why did it classify the customer that way? Why did it ask that question? Why did it escalate? Why did it stop the workflow? Why did it decide the person needed a human? Why did it treat this interaction differently from the last one?
Without that reasoning trail, you do not really have intelligent automation. You have a fast process with a weak memory.
And that is risky. Because when the decision is challenged, when the customer complains, when the regulator asks questions, or when the organisation needs to understand what went wrong, the output will not be enough. You need the reasoning. That is what makes the system defensible. That is what makes the system auditable. That is what allows the system to improve over time.
The Real Opportunity Is Not Replacing the Call Centre
The real opportunity is not just replacing call centre workers with bots. That is the shallow version of the opportunity. The real opportunity is to capture the intelligence that already exists inside the organisation.
The way experienced staff handle difficult conversations. The way they adjust to different customers. The way they make judgement calls. The way they protect the relationship while still moving the process forward. That is valuable. That is not just “soft skills”. That is operational intelligence. That is customer intelligence. That is organisational reasoning.
If we can capture that properly, then AI systems become far more powerful. Not because they are replacing people blindly, but because they are learning from the people who understand the work.
That is the difference.
Final Thought
I think this is where agentic AI either succeeds or fails.
If we keep building systems that only understand the workflow, they will keep breaking when the customer does not behave like the process diagram. But if we build systems that capture the reasoning behind client relationships, then we have a much better chance of creating AI that actually works in the real world.
Customer-facing work is not just about moving someone from step one to step two. It is about understanding what is happening in the interaction. It is about knowing when the script is not enough. It is about knowing when the workflow needs human judgement. And it is about preserving that judgement before it disappears.
That is the part we should be capturing.
Not just the call summary. Not just the completed form. Not just the workflow outcome.
The reasoning.
Because if agentic AI cannot understand why experienced people do what they do… then it is not ready to take over the relationship.
