AI assistant features that actually shipped — not demos

The 2025 AI assistant announcements were ambitious; the 2026 ship reality is more modest. carmannews lists the features that landed in production, the ones still in beta, and the ones quietly dropped.

Every keynote season brings a wave of AI assistant demos that look like magic on stage. Some of those features arrive in your phone working roughly as shown. Some arrive months later in a watered-down form. And some never ship at all, quietly disappearing between the announcement and the update. For a buyer, the useful skill isn’t memorising the demos — it’s learning to tell a feature that genuinely works from one that’s still a promise. The pattern repeats every year, and it’s readable if you know what to look for.

Why the demo and the product diverge

A stage demo is a controlled performance: chosen prompts, strong connectivity, ample time, and often a pre-release build running on hardware that isn’t yours. Your phone is none of those things. It has a weaker connection sometimes, a smaller battery budget, and a thousand edge cases the demo never touched. So the gap between “look what it can do” and “look what it does for me” isn’t deception so much as the distance between a lab and the real world.

Two structural realities widen that gap. First, the most impressive features often need to send data to a remote server to run, which makes them slower, dependent on a connection, and sometimes restricted by region. Features that run entirely on the device are faster and more private but necessarily more limited. Second, anything that touches accuracy — summarising, answering factual questions, taking actions on your behalf — carries a risk of being confidently wrong, and companies tend to slow-roll those features behind warnings precisely because the failure mode is embarrassing.

Which categories tend to actually ship

Without naming specific products — which change every cycle — the categories sort into reliable buckets based on how hard they are to get right.

• Usually ships and works: writing and editing help. Drafting a message, rewriting a paragraph in a different tone, summarising a block of text you provide — these have matured into dependable everyday tools, because the stakes are low and you can see and fix the output yourself.
• Usually ships and works: photo and audio cleanup. Removing an object from a picture, sharpening a blurry shot, cleaning up background noise on a recording. These are well-suited to the technology and have become genuinely useful, not gimmicks.
• Ships, but uneven: voice assistants that take actions. Asking an assistant to do multi-step things across your apps — book this, message that, change a setting — demos beautifully and works inconsistently in the wild, where it meets apps and accounts it wasn’t tuned for. Treat these as promising rather than dependable.
• Often delayed or dropped: anything promising factual reliability at scale. Features that answer open-ended questions as if they’re always right are the ones most likely to slip, get hedged with disclaimers, or quietly vanish, because confident wrong answers are the hardest problem in the field.

The honest test for any AI feature

There’s a simple way to judge whether an assistant feature deserves your trust: ask what happens when it’s wrong. For writing help, you read the draft and fix it — low risk, use freely. For a factual answer or an action taken on your behalf, a mistake can cost you real time or money, so verify before you rely on it. The features worth adopting immediately are the ones where you stay in the loop and can catch errors; the ones to approach cautiously are those that act without showing their work.

What to check before you buy for the AI

If a device’s AI features are part of why you’re buying it, don’t take the announcement at face value. A few checks separate marketing from reality.

• Is the feature shipping now, or “coming soon”? Buying hardware today for a feature promised in a future update is a gamble. Promised features slip and occasionally never arrive. Buy for what the device does today; treat anything still on the roadmap as a bonus that may or may not land.
• Does it need a subscription? A growing share of the most capable AI features sit behind a recurring fee on top of the hardware. Factor that ongoing cost into the decision, not just the sticker price.
• Does it work offline or only online? Features that require a server connection won’t work everywhere and may carry usage limits or regional restrictions. If you need something to work reliably and privately, prefer features that run on the device.
• What do current, real-world reviews say — not the keynote? Reviews written after launch, by people using the shipping product on their own devices, are far more honest than the announcement. Weight them accordingly.

There’s also a privacy dimension worth a moment’s thought before you lean on these features. Many of the most capable ones send your words, photos, or files to a company’s servers to do their work, which is what makes them powerful and also what makes them worth a second’s pause for anything sensitive. Features that run entirely on your device keep your data local and are the safer choice for private material. It’s not a reason to avoid AI help — it’s a reason to know, for a given feature, whether what you hand it stays on your phone or travels somewhere else.

This field moves faster than almost any other in consumer tech, and the specifics — which assistant does what, what’s free versus paid, which features survived to release — change with every update. For the current state of any particular product’s AI features, check recent independent reviews and the maker’s own documentation rather than the launch promises.

The fair question for any AI feature is simple: what does it cost you when it’s wrong? Answer that and you’ll know which ones to trust today and which to wait on.

Kenji Tanaka, Tech Editor, carmannews