In 2018 we spent our days writing sample sentences so Alexa could understand people. Today, a language model understands almost any way of asking for something without anyone giving it examples. In between, almost everything about how you build an assistant has changed. What hasn’t changed is what makes an assistant work for people, and now it’s even more obvious when it’s missing.
How it was done in 2018
An assistant was built piece by piece. First came language understanding: each request was classified into an intent (“book a table,” “check opening hours”) and the important details were extracted, such as the date or the number of people. In Alexa, you defined this with intents, sample utterances and placeholders for the data, which Amazon calls slots (Amazon). Next, a dialog manager decided what to do based on the state of the conversation. Finally, the response was generated, almost always from hand-written templates.
That architecture came from decades of research on spoken dialog systems, well documented in the handbook on conversational interfaces by Michael McTear, Zoraida Callejas and David Griol (McTear et al., 2016). It worked, but it had well-known limits. It only understood what someone had anticipated: if a person asked for something using words that weren’t in the examples, or something there was no intent for, the answer was “sorry, I didn’t understand.” Every new intent required lots of sample sentences, and maintaining hundreds of them was a huge job. At Monoceros we built many Alexa apps this way, and some of what we learned is in eight tips for building good Alexa skills in Spanish.
What has changed
With language models, those three pieces can be handled by one. The model understands what it’s asked even if it’s phrased a thousand ways, keeps track of the conversation’s context and writes the response. You can put together a prototype that really converses in days.
The platforms have gone the same way. Microsoft fully retired its LUIS language understanding service on March 31, 2026, and recommends migrating to its new service, Conversational Language Understanding (CLU) (Microsoft Learn). In February 2025, Amazon introduced Alexa+, the new generation of Alexa built on generative language models (Amazon, 2025). And tools like Rasa have moved to a hybrid approach, where the model interprets what the user wants and the business logic stays deterministic (Rasa).
What hasn’t changed
You still have to design the conversation
Just because a model can talk doesn’t mean it knows what to say in your service. You have to decide what the assistant does and doesn’t do, how it introduces itself, how it asks for information, how it responds when it can’t help and what personality it has. We used to write it response by response. Now it goes into the instructions and the examples, but the work of thinking through the conversation is still there.
You still have to know the limits
A 2018 assistant failed visibly: it didn’t understand and it said so. Today’s can fail without anyone noticing, confidently answering something false or promising something it can’t deliver. Knowing what the technology doesn’t do well, and designing for it, matters more than ever.
You still have to test it
You used to be able to walk through every path in the flow. Now the paths are endless and the model never answers the same way twice. Evaluation has gone from a final phase to an ongoing part of the work, as we explain in evaluating an agent before your customers use it.
You still have to know when to hand over to a person
Some conversations an assistant shouldn’t handle on its own: a serious complaint, a sensitive situation, a case that doesn’t fit. Designing that handoff well mattered in 2018 and it still does.
Context of use still rules
The noise of a kitchen, the hurry of someone calling from the car or the patient speaking from a hospital bed still shape what works. We explain it in what to ask yourself before building an assistant.
What I’d keep from 2018
There are situations where the risk of an error is high, because it affects people or customers, and the business too: a payment, a medical appointment, a legal detail. For those parts I still prefer closed flows and texts, reviewed by people. The language model can understand the request and carry the conversation, but the decision and the text that commits the company must stay under control. That’s the idea behind hybrid systems, and it’s also the lesson of many years building assistants that people actually use.
Technology has taken a lot of repetitive work off our hands. The work of design, of evaluation and of getting to know the people who’ll use the assistant is still ours, and it’s what we offer in conversation and agent design.
References
