AI stopped being a destination. Now it lives inside the work

For two years the AI product was a chat box you went to. You opened a tab, pasted context, and carried the answer back to where you actually work.

That round trip is disappearing. The tools that shipped in August prove it, and I've argued that AI products fail at the interface, not the model — where you put AI now matters more than which model you picked.

Embedded AI beats standalone chat because it removes the trip: it remembers across surfaces, acts inside the file, and asks for approval where the decision lives — not in a separate tab.

Why did standalone chat stop being enough?

Standalone chat taxes the user on every turn.

You explain the project in one place, then re-explain it in another. You copy an answer from chat into Figma, then fix the drift by hand. Each switch breaks flow. At daily use it drives churn.

McKinsey's 2025 State of AI report found 88% of organizations now use AI in at least one function — it has to behave where work happens. When AI is everywhere, a product that makes you go elsewhere feels slow, even if the model is fast.

What does embedded AI actually look like in 2026?

Embedded AI looks boring — it shows up as memory or an action inside a tool you already have open.

Three examples shipped within a week. Anthropic's announcement that Claude's memory now works everywhere unified chat and Cowork into one layer. You tell Claude who your manager is in chat, and Cowork already knows it when it drafts the update. Topics update as you talk and live in Settings > Memory where you can read, edit, or delete them.

Figma's announcement that the canvas is now open to agents did the same for design. Agents work through MCP directly on your canvas, using your components and variables instead of inventing new ones. The file became the agent's workspace.

The third is quieter: writing tools that draft inline, summarize in place, or act in the inbox instead of opening a chat panel. They just save the trip.

How do you design for AI that lives inside the workflow?

Design it like a teammate next to the user, not instead of them.

Put the action where the decision is. If the user is editing a doc, the AI drafts in the doc. If they're on the canvas, it edits the canvas. A Cowork task that needs files should start with those files attached.

Make the work visible and reversible in place. Show what changed and offer undo where the output landed. I wrote that your second session decides if it ships — embedded AI fails that test if a returning user can't see what it remembered and why it changed the answer.

Keep approval human and close. One decision — accept, edit, or reject — right next to the change. No separate queue.

Scope memory tightly. Unified is useful until it leaks a client voice into the wrong project. Keep Topics scoped and easy to prune. A list you can scan in ten seconds beats a hidden store.

Embedded doesn't mean invisible. It means placed correctly.

When should you still make AI a destination?

When the job is thinking, not doing.

Exploration needs space. Brainstorming a direction, researching a category, or wrestling with a first draft benefits from a separate surface where you can be messy without touching production. Chat is a whiteboard for that.

Execution needs proximity. Filing a doc, updating a component, triaging an inbox, or generating the tenth variant should happen inside the tool where the result lives.

A simple rule: if the user already knows where the output goes, embed it. If they don't yet know what the output should be, give them a destination to figure it out.

Most solo products need both, but they put too much in chat because chat is easy to ship. Ship chat for learning, embedded actions for retention. The products that feel fast in 2026 rarely make you open the chat box.