"When I see that wall of text, I just turn to the escalation desk."
That's a contact center agent describing what happens when she opens her company's chatbot mid-call. She didn't say the chatbot gave her the wrong answer. She said she couldn't use the answer it gave her.
There's a difference, and most companies rolling out AI tools for frontline employees haven't yet caught it.
She kept talking. "I can't follow that in the moment. And a lot of times I'm escalating because I really want to cover my butt in case there's a mistake."
Read that twice. She wasn't escalating because she couldn't find information. She was escalating because she needed someone else to be responsible for the answer. The chatbot handed her summarized procedures. What she needed was clarity and confidence.
The moment of performance is the exact window where an employee has to apply knowledge—they have to answer a question, perform a task, or solve a problem. In that moment, they have to rely on either memory or guidance.
This isn’t the same as training. In training, the audience is controlled, the distractions are shut out, and the consequences are minor. But in the moment of the performance, the audience (the customer) is impatient, the distractions are everywhere, and the consequences can be significant.
In a bank or credit union, many procedures are too complex for employees to memorize. That is why chatbots have become so popular. They allow employees to reference information right when they need it.
But in the moment of performance at a financial institution, with some procedures, summarized text is not enough.
Two variables decide what an employee needs in that window:
Most training and knowledge tools are built as if accuracy is the only thing that matters. They ignore the gap between knowledge presentation and knowledge application.
There are two separate barriers, and companies usually only build for one.
The first is findability — can the employee locate the right information fast enough. This is the problem chatbots were built to solve, and for a lot of tasks, they solve it well.
The second is followability under pressure — once they've found it, can they actually scan and apply it live, mid-conversation, without the customer noticing a long silence and without the employee's own confidence cracking. This is the problem almost nobody is building for.
Confidence is the variable that gets ignored. An employee who isn't sure they've got it right will not commit to the action — it doesn't matter whether the information was technically available to them. Being able to find an answer and being able to act on it under pressure are not the same skill, and a lot of knowledge tools are optimized only for the first one.
When the tool doesn't deliver that confidence, there are only two things an employee can do: fall back on memory, or escalate.
Falling back on memory is a compliance and accuracy risk, exactly the thing a regulated business can't absorb.
Escalating is fine in small doses. It's what escalation desks are for. But when it becomes the default response to any procedure with real complexity, it stops being a safety valve and starts being a symptom.
"When I see that wall of text, I just turn to the escalation desk."
This isn't an argument against chatbots. It's an argument about fit. Three tiers, three different tools.
In financial services, "what's the current wire transfer cutoff time?" One question, one correct answer, no branching. A chatbot is exactly the right tool here.
Ten steps, linear, no real decision points. A chatbot can still walk someone through this live — the structure is simple enough to read and follow in real time.
Dozens of steps, spanning multiple systems, with decision branches that change the path based on the situation the employee faces. This is where the chatbot breaks down.
The instinct is to make the chatbot explain Tier 3 better — tighter summaries, clearer language, better prompts. That's polishing word soup. The underlying issue doesn't move: a wall of generated text, however well-written, is the wrong format for something that has to be followed live, under pressure, with branching decisions.
This isn't a knowledge problem. It's a format problem.
What that employee needs instead is a decision tree — not a replacement for the chatbot, but something the chatbot should point them to.
The decision tree wins in two ways that a paragraph never will:
It's human-approved. The employee isn't trusting an AI-generated summary of a policy; they're following a path someone with authority already signed off on. That's where the confidence problem actually gets solved — not with better writing, but with a structure the employee knows is correct before they start.
It's designed for performance, not for reading. Visual cues, explicit decision points, a clear next step at every branch. A musician doesn't perform from a paragraph describing the song. They perform from sheet music — built specifically for real-time execution, not comprehension after the fact. A complex procedure deserves the same design standard.
Most companies measure the success of their chatbots by adoption (how many people are using the tool). That number tells you almost nothing about whether the moment of performance problem is solved.
The metric that matters is internal escalation volume, specifically on your Tier 3 procedures. If a decision tree is working, escalations on that procedure go down, because employees have a path they trust enough to follow without backup.
That number is worth watching closely, because escalations aren't just a support cost. They're the visible symptom of a moment-of-performance failure happening upstream and they show up everywhere else in the business:
The goal was never "give employees a chatbot." It was give employees the confidence to act correctly under pressure. A chatbot delivers that for two of the three tiers. For the third, it hands someone a wall of text and calls it done.
Which of your Tier 3 procedures are still living in a chatbot instead of a decision tree — and what is that costing you in escalations right now?