The most common way organizations accidentally undo the thing they paid for
There is a pattern in AI deployments that I’ve watched repeat itself across industries, company sizes, and use cases. An organization decides to deploy conversational AI. They go through the evaluation process, select a platform, run the implementation, and launch. And then — quietly, without anyone explicitly choosing it — they build something that isn’t conversational at all.
It sounds like this: “To get started, could you tell me whether your issue is related to billing, a reservation, or something else?”
Or: “I can help you with that. First, can I get your account number?”
Or: “Are you calling about an existing order or a new order?”
These are not conversations. These are forms. The fact that they are being spoken rather than typed doesn’t change what they are. The human on the other end of the call is being routed through a decision tree, one branch at a time, in spoken form. This is the old IVR system — press 1 for billing, press 2 for reservations — dressed in a friendlier voice and a warmer tone. The underlying structure is identical.
And here’s what makes this particularly worth examining: the AI didn’t need to be built this way. The organization chose to build it this way, usually without realizing that’s what they were choosing.
What AI can actually do
A well-functioning conversational AI does not need a question like “is this about billing or reservations?” to understand what someone is calling about. It can listen to a person explain their situation in their own words and derive what kind of issue it is from the content of what they say.
More than that — it can tolerate the fact that a person often doesn’t know exactly what category their problem belongs to when they start speaking. The person who says “I think there might be something weird with my account, like, I saw something on my bank statement that I didn’t recognize” doesn’t need to know whether that’s a billing question or a fraud question before they start talking. The AI can hold that ambiguity, follow the explanation, and figure out what kind of help is actually needed.
This is not an advanced capability. It’s what conversational AI does by default when it hasn’t been constrained to do something else.
The constraint — the “please select billing, reservations, or something else” — is almost never added because someone thought it would make the experience better. It’s added because it makes the workflow feel more controllable. The designer can see the branches. The routing is explicit. The system is legible to the team that built it. These are understandable concerns. But the cost is the thing the organization actually bought, which is an AI that could meet the customer where they are instead of requiring the customer to meet the system where it needs them.
Why this happens
Most workflow design has been built on a foundational assumption: the human must conform to the system. The system is structured. The human must become structured enough to enter it. Every traditional design artifact — the intake form, the support ticket, the IVR menu, the dropdown category selector — exists to perform this conversion. The messy, uncertain, emotionally textured human problem gets converted into a data type the system can process.
That conversion was a reasonable trade-off when the alternative was no system at all. When you couldn’t process something unless it was structured, making it structured was the only option.
But conversational AI changes the trade-off. The AI doesn’t need the conversion to happen first. It can process unstructured input. It can work with the messy, uncertain, emotionally textured version of the problem. The conversion — from human expression to structured data — can happen after the AI understands the situation, not before the human is allowed to explain it.
When designers who were trained on old-interface thinking build AI workflows, they often rebuild the conversion layer by instinct. They prompt the AI to ask classification questions early. They add intake fields. They build routing logic before the human has had a chance to fully explain what’s happening. They are, in effect, wrapping a conversational system in the structure that conversational systems were supposed to replace.
The 60-second test
Here is a practical way to evaluate any AI voice or chat workflow. Find the first minute of a conversation — the first three to five exchanges between the customer and the AI. Ask: is the AI gathering information in order to classify and route? Or is it following where the human is going and allowing the full situation to emerge?
If the AI is asking classification questions in the first 60 seconds, the design was built for the system’s convenience, not the customer’s. The system is trying to sort the human into a category before understanding the human’s actual situation. This is not necessarily wrong in every case — there are contexts where rapid routing is genuinely the right call. But it is worth being conscious that it is a choice, and that the choice has costs.
The alternative is to let the AI do what it already knows how to do: stay with the person, follow the explanation, and reach an understanding of the situation before deciding what to do with it. This requires prompting the AI to resist the urge to categorize early, not to prompt it into premature classification. It is, in many ways, a discipline of subtraction — removing the constraints that convert a conversational system back into a spoken form.
What you get when you let it breathe
When an AI is designed to let a person fully explain before it routes or responds, something changes in the conversation. The customer doesn’t feel processed. They feel heard. And there is a practical benefit that goes beyond tone: the AI ends up with more accurate information about what’s actually going on.
The person who was going to say “I think there might be something weird with my account” was also about to say “but I’m not sure if it’s from last month or the month before, and I think one of them might be a recurring charge I forgot about.” If the AI interrupts after the first sentence to ask for an account number, it gets the account number. If it waits, it gets the account number and a much clearer picture of what’s actually being investigated.
The result is a workflow that reaches the right resolution faster, with less back-and-forth and fewer transfers, because the AI understood the situation before it started responding to it.
This is not magic. It is what happens when you design for the nature of conversational AI instead of against it.
Since February 5, 2026, I’ve had the opportunity to study feedback from more than 18,000 complicated AI integrations running in production across 43 projects.
That work has taught me something important: AI workflow design is not just automation. It is a new discipline.
Humans do not arrive as clean tickets, categories, forms, or process steps. They arrive unfinished. They speak, the problem emerges, and the workflow designer has to understand what solution the human is actually reaching for.
I wrote this textbook to help other workflow designers think more clearly about that process.
The framework is built around five movements:
Emergence gives you the solution. Hallucination gives you the ideal. As-is discovery gives you reality. To-be selling gives you transformation. Stewardship gives you renewal.
The textbook is free to download below.