Why is Alexa so bad? An engineer's explanation

Alexa feels bad because it is built to make one fast guess. Alexa+ uses large language models, like ChatGPT and Claude, but it runs them inside a different harness. A voice harness prepares the context, skills and tools before the model runs and gives the model one try, so the answer comes in a second or two. ChatGPT and Claude run in an agent harness that lets the model think, look things up and try again, which takes far longer. The same kind of model behaves very differently in each.

We have built production voice assistants and the Port OS device platform. This post explains why Alexa still misunderstands simple requests after the move to language models, and what you can do about it.

Alexa+ uses LLMs, so why is it still bad?

Amazon announced Alexa+ in February 2025, built on large language models. Many people expected it to behave like ChatGPT with a speaker. It does not, and the reason is the architecture around the model.

The model is one part of a voice assistant. Around it sits a harness: the code that decides what the model sees, which tools it may call and how long it has. Our post on how Alexa works covers the two common designs in detail:

• A streaming voice harness starts work as soon as you speak. It fetches context about the device you are talking to, classifies the request when you stop, then fetches the relevant devices, skills and tools in parallel. The model gets one focused call to choose the action. The design is optimized for zero-shot inference: right on the first try, in about two seconds.
• An agent harness, the kind ChatGPT and Claude use, keeps devices, skills and tools outside the core loop. The model decides what to search for, reads the result, and decides the next step. It is optimized for long inference: right eventually, after several model calls of seconds each.

So Alexa+ can understand more phrasings than old Alexa, and it still fails in the same places: when the pipeline hands the model the wrong context, the model has no way to find the right one.

Five reasons Alexa gets it wrong

1. The pipeline picks the context before the model sees the request

The classifier and the context fetches decide which devices, skills and tools the model gets. If the request is classified as music when you meant the TV, or the search misses the lamp you meant, the model answers from the wrong facts. In an agent harness the model would notice and search again. In a zero-shot harness, the first guess is the answer.

2. The speed budget leaves no time to think

A voice reply has to start within a second or two, or it feels broken. Classic Alexa turned off a light in about 1.3 seconds from the end of speech in our measurements. Every extra model call would add seconds, so the harness allows one or two. That leaves no room for the checking that makes chat agents accurate.

3. Names in your home collide

Most home failures are naming problems. A light group called "Living room", a switch called "Living room switch" and lamps called "Lamp 1 living room" all match the same words. Larger homes make this worse quickly, and a single guess cannot ask the home which one you meant.

4. Cost at scale means smaller models

Alexa answers millions of requests a day. Each large model call costs money, so a voice assistant leans on small, fast models and fixed routes for common requests. Smaller models are quicker and cheaper, and they are less forgiving of unusual phrasing.

5. When Alexa does run an agent, it is slow

Alexa+ also splits work between a voice model that talks with you and a background agent that does longer tasks. The conversation feels quick, and the task runs at agent speed. In our experience, a task handed to an agent like this can take 30 seconds to a minute, which feels worse than an honest "I can't do that".

Why ChatGPT and Claude seem smarter

ChatGPT, Claude and Gemini in chat run the model in an agent harness. They read the whole request, list devices, search, call a tool and check the result. When we timed AI assistants controlling a home, a single light took 8 to 11 seconds, and an ambiguous request took up to two minutes.

That trade is fine in a chat window and too slow for a speaker. Neither harness is better at everything. Alexa gives up flexibility for speed, and chat agents give up speed for flexibility. Our Gemini vs Google Assistant comparison shows the same pattern on Android.

How to make Alexa work better today

• Give every device a unique name. Do not reuse room names for devices or groups.
• Put devices in rooms and groups. Then "turn off the kitchen lights" has one clear target, and the context the pipeline fetches is right.
• Use routines for long requests. A routine turns a complex job into one fixed phrase Alexa can match.
• Talk to the device in the room. Alexa uses the room of the device that heard you, so "turn off the light" works best near the light.
• Keep phrasing short. A zero-shot harness handles common phrasings best.

For jobs that need the agent's patience, such as "check the house and turn off whatever was left on", a chat agent is the better fit. Our post on the AI smart home covers those jobs. Port OS connects assistants like ChatGPT and Claude to the devices in your home, including those linked to Alexa. Read how to connect an AI assistant to start.

Frequently asked questions

Why is Alexa so bad lately?

Alexa+ moved Alexa to large language models, which understand more phrasings but add time to each request. The voice harness still gives the model one fast try with context chosen in advance, so requests that the pipeline misroutes still fail, and tasks handed to a background agent take longer than before.

Is Alexa+ the same as ChatGPT?

Both use large language models. Alexa+ runs them in a streaming voice harness built for fast zero-shot answers, while ChatGPT runs its model in an agent harness that can think and call tools for many seconds. That is why they behave differently on the same request.

Why does Alexa not understand me?

The pipeline classifies your request and fetches context before the model sees it. If your phrasing lands in the wrong category, or the device name matches something else, the model answers from the wrong context and cannot go back to look again.

Why does Alexa do the wrong thing in my smart home?

The most common cause is a naming collision: a device, a room and a group with the same or similar names. Give each one a unique name, and put devices in rooms and groups.

All posts