AI Assistant Capabilities: What They Can and Can't Do

An AI assistant in 2026 can do far more than answer questions. The big chat assistants can write, research, read your files and photos, generate images, remember context between conversations, and, when you connect them, take actions in other apps. What they still can't do is guarantee they're right, know anything you haven't shared, or take responsibility for a decision.

Here's the short version of AI assistant capabilities today:

The rest of this guide covers what each capability is good for, where assistants still break, and the characteristics that separate an assistant you can rely on from one you have to babysit.

What AI assistants can do today

Conversation and writing

This is the core skill, and it's genuinely strong. An assistant can turn rough notes into a clean email, explain a concept at whatever level you ask for, rewrite something in a different tone, or argue the other side of a decision you're about to make.

The trick is context. "Write a project update" gets you something generic. "Write a three-sentence project update for my manager; we're two days late because the vendor missed a delivery, and here's the new date" gets you something you can send.

Search, research, and summaries

Most major assistants can now search the web and come back with a summary and links. Some go further. Gemini, for example, lists Deep Research among its features, which runs a longer multi-step research task and returns a report.

The same skill works on documents you provide. Paste in a 40-page report and ask for the three decisions it asks you to make, and a good assistant will find them faster than you would. Treat the output as a map, not the territory: click through to the sources for anything that matters.

Reading files, photos, and screens

Assistants are now multimodal: they understand images and documents, not just typed text. You can photograph a receipt, a whiteboard, or an error message and ask about it. You can upload a spreadsheet and ask which month looks off.

This is often the most useful capability for non-technical users, because it removes the step of describing the problem. Show it the problem instead.

Creating images, audio, and video

Image generation is standard across the big assistants, and some now go beyond it. Google's Gemini app lists image, video, and music generation as separate features. That's handy for mockups, slides, and social posts. It's weaker for anything that must be exact, like a real product photo, a legible diagram, or a real person's likeness.

Memory across conversations

Assistants increasingly carry context forward. Claude, for instance, can search your past chats and use memory to build on earlier conversations, so you don't have to re-explain your project every time.

Memory is convenient and worth managing. Know where your assistant's memory settings live, what it has saved, and how to turn it off for a sensitive conversation.

Taking actions through connected apps

This is the newest and fastest-moving capability. Connect an assistant to your email, calendar, or documents and it can find things and draft replies in place. Some can go further and operate a computer directly. Anthropic's computer use tool, for example, gives Claude screenshot, mouse, and keyboard control of a desktop environment.

When an assistant plans and completes multi-step tasks on its own like this, it starts to behave like what people call an AI agent. The line between "assistant" and "agent" is mostly about how much it does without checking back with you.

Voice assistants vs chat assistants

For years, "AI assistant" meant a virtual assistant like Siri or Alexa: a voice interface for short commands such as setting a timer, sending a message, or turning off the lights. Apple's guide to using Siri on iPhone is still mostly that kind of task.

Chat assistants such as ChatGPT, Claude, Gemini, and Microsoft Copilot come from the other direction: long, open-ended conversations and complex requests. The two categories are converging. Chat assistants add voice modes, and voice assistants add language-model features.

If your phone is where you use AI most, our guide to using AI on your phone covers what's built into iPhone and Android.

What AI assistants still can't do reliably

Capabilities grew fast. Several limits didn't move much.

They can be confidently wrong. Language models sometimes produce plausible statements that aren't true, a failure widely called AI hallucination. Google's own Gemini overview has a section on limitations for exactly this reason. Anything factual that you'll act on (a figure, a legal point, a medication detail) needs checking against a primary source.

They only know what you give them. An assistant can't see your company's context, your client history, or last week's meeting unless it's connected to them or you paste them in. Most "bad answers" are really missing-context answers.

Actions carry real risk. Anthropic's computer use documentation warns that the model may follow instructions it finds in webpages or images, even when they conflict with yours. It recommends running it in an isolated environment, keeping it away from sensitive data, and having a human confirm consequential decisions. That advice applies to any assistant you let loose on your accounts.

They don't own the decision. An assistant can lay out options and trade-offs well. The judgment, and the accountability, stays with you. That's why the jobs worth delegating tend to be the prep work, a split our look at the AI executive assistant breaks down task by task.

Characteristics of a friendly, trustworthy AI assistant

"Friendly" in an AI assistant shouldn't mean flattering. It means easy to work with and safe to rely on. The U.S. National Institute of Standards and Technology's AI Risk Management Framework lists the characteristics of trustworthy AI systems: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed.

Translated into what you'll notice day to day, a good assistant is:

• Honest about uncertainty. It says "I'm not sure" or "check this" instead of bluffing. Anthropic has written about the traits it trains into Claude's character, including curiosity, open-mindedness, and being honest about its views even when you disagree.
• Clear about its rules. It tells you when it won't do something and why. OpenAI publishes a Model Spec describing how its models are meant to behave, which is the kind of transparency worth looking for.
• Careful before acting. It asks before sending, buying, deleting, or accepting anything on your behalf.
• Transparent about memory. You can see what it remembers and switch memory off.
• Easy to correct. When you point out a mistake, it fixes the output rather than defending it.
• Consistent. The same request gets a comparably good answer tomorrow.

A useful test: give an assistant a question you already know the answer to, including one where the honest answer is "it depends." How it handles that tells you more than any feature list.

Getting more from the capabilities you already have

Most people use a fraction of what their assistant can do. Three habits close the gap:

1. Give it the material. Attach the file, paste the thread, or share the screenshot instead of summarising it yourself.
2. Ask for the check. "List the claims in your answer I should verify" is a prompt that turns a confident paragraph into a to-do list.
3. Save what works. When a prompt or setup reliably solves a recurring job, keep it rather than rebuilding it every time.

If you're earlier in the journey, start with our practical guide to using AI. And when the thing you want is an assistant setup someone else already built, Taku is an AI-native desktop workspace where you can mirror a working AI app, agent, or workflow, run it, and keep it, instead of recreating it from scratch. Browse the free app library to see what's there. Taku is in Beta, and the Mac app is available now.

Key Points

• Six core capabilities: writing, research, reading files and images, creating media, memory, and taking actions in connected apps.
• Context is the multiplier. Most weak answers come from missing information, not missing ability.
• Accuracy isn't guaranteed. Verify facts you'll act on against primary sources.
• Actions need guardrails. Keep assistants away from sensitive data and confirm consequential steps yourself.
• Friendly means trustworthy: honest about uncertainty, clear about its rules, careful before acting, and easy to correct.

FAQ

What can an AI assistant do?

Modern AI assistants can write and edit, answer questions, search the web and summarise sources, read documents and images, generate images and other media, remember context across chats, and take actions through apps you connect. Voice assistants like Siri focus more on quick commands and device control.

What are the characteristics of a friendly AI assistant?

A friendly assistant is honest about what it doesn't know, explains when and why it won't do something, asks before taking consequential actions, lets you see and control what it remembers, and accepts corrections. NIST's trustworthy-AI characteristics, including reliability, safety, transparency, and privacy, are a good formal checklist.

Can AI assistants take actions for me?

Yes, increasingly. Connected to your apps, they can find information and draft replies, and some can operate a browser or desktop. Vendors recommend isolating these tools from sensitive data and confirming important actions yourself.

Do AI assistants remember what I tell them?

Many now do, through memory features and search across past chats. Check your assistant's settings to see what it stores and to turn memory off when you need to.

What's the difference between an AI assistant and an AI agent?

An assistant mostly responds to what you ask, one step at a time. An agent plans and carries out multi-step tasks with more independence. Many assistants now include agent-like features, so the difference is increasingly a matter of how much you let it do without checking in.