What It Actually Means
Scripted bots follow a decision tree someone drew. Press 1 for accounts. They handle exactly what was anticipated and collapse the moment a question is phrased unexpectedly. Most people’s dislike of chatbots was earned by these.
Model-based bots use a language model to interpret what was actually said. They cope with phrasing, follow context across a conversation, and answer from supplied material. Genuinely different, and worth reassessing if your opinion was formed years ago.
Why They Are Still Resented
The complaint is almost never that the bot was artificial. It is that it stood between the customer and a person who could help.
A bot that answers a question at eleven at night is a service. A bot that is the only route to support, that cannot escalate, that loops through the same suggestions while somebody grows angrier, is a cost-saving measure the customer is paying for in frustration.
Customers can tell the difference immediately, and it is a decision about how the bot is deployed rather than about how good it is. A brilliant bot with no exit is worse for the business than a modest one that hands over quickly.
Where They Genuinely Work
High-volume repeated questions. Where is my order, what are your hours, how do I reset this. Real, cheap, and unambiguously better for the customer than waiting.
Out of hours. Something useful at ten at night beats a form and a two-day wait.
Getting to the right place. Understanding what someone needs and routing them with the context already gathered, so nobody repeats themselves.
Internal use, which is consistently undersold. Staff asking about policies, process and where things live. Lower stakes, tolerant users, and it removes a surprising amount of interruption from the people who currently answer.
Where They Should Not Be
Complaints, anything involving money going wrong, cancellations, and any situation where the customer is already unhappy. Those need a person, quickly, and the bot’s only useful job is to get them there faster with the details already captured.
Anything giving advice with consequences: legal, medical, financial. A confident wrong answer becomes your liability.
What To Ask
- How does someone reach a person, and how obvious is it? If it is buried, the design has chosen deflection over service.
- Does it pass the conversation across? Making a frustrated customer repeat themselves undoes any goodwill the bot earned.
- What does it refuse? There must be a defined list, routed straight to a handoff.
- What is it answering from? Your documented material, or general training. The second will invent policy.
- What do we do with the transcripts? Every failed conversation names a gap in your content. This is where the compounding value is and it is usually ignored.
The Measure That Matters
Containment rate, the proportion of conversations the bot handles alone, is the metric most often reported and it is the wrong one. It rises when escalation is hidden, which is the failure mode.
The honest measures are whether the customer’s problem was solved, how fast they reached a person when they needed one, and whether satisfaction held up against the previous arrangement. A bot deflecting seventy percent of contacts while quietly costing you customers is not succeeding, and containment will not tell you.
More terms are in the glossary.
