Embedded AI Works Better Than Chat For Repeatable Tasks
Chat bots help when goals are ambiguous, but they introduce context switches.
Jul 05, 2025
TLDR;
Chat bots help when user goals are broad or ambiguous. They are weaker when the task is well-defined and the user is already inside a workflow. What has worked for one of the projects we are building is to start with conversation while the problem is still broad, then bake the top use-cases into in-app controls.
Why are we Over-Chatting?
It feels like every new app must include an "AI assistant," and most people immediately picture a chat window. I get why. Freeform conversation promises flexibility: ask anything, get something back.
But when I use a product to finish a specific task, opening a chat panel often breaks my flow. I have to stop, explain the context, write a prompt, read the answer, and translate it back into the interface I was already using.
While building a new platform from the ground up, I kept asking the same question: is conversation really the right interface for every AI feature, or have we overestimated chat? For many common use cases, assistance inside the existing workflow wins.
The Chat: brings flexibility
Chat bots offer a blank canvas. Users can pose any task, like draft an email, troubleshoot code, brainstorm ideas, in plain language. This openness is invaluable when goals are unclear or involve multi-turn dialogue, such as customer support or creative exploration. First-time users usually appreciate the “just tell me what you want” simplicity, and support bots can deflect common questions faster than human agents. Chats are immersive and capture a user's focus.
But that very flexibility demands a heavy lift from users. They must break away from their task, compose an effective text (prompt), and then parse the AI’s response often iterating multiple times. Each context switch costs minutes of lost focus, aggregated across the day. These detours can erode productivity rather than enhance it if used for hours. Think about tools, systems and products that require cognitive decision making, the chat appears more of a distraction.
In-flow AI: help where the work already is
Contrast a chat panel with Smart Compose in Gmail: as you write, inline suggestions offer to finish your sentence. No separate window, no prompt to craft. By 2022, Smart Compose had been accepted 118 billion times 1, saving users keystrokes while staying inside the email composer.
Similarly, GitHub Copilot puts code completions directly inside the IDE. In a reported enterprise study, Copilot users completed a coding task 55% faster2. The useful part is not only the model. It is the placement: the suggestion appears where the developer is already working.
These examples share the same shape. The AI lives in the same interface where work happens. Users accept a suggestion instead of translating their need into a prompt. The system already has the current document, codebase, or data, so there is less re-explaining.
When Chat Makes Sense (and When It Doesn’t)
Use chat when goals are open-ended or highly variable, such as brainstorming or exploratory Q&A. It also helps when multi-turn conversation is the work, like troubleshooting or tutoring, or when the product has a steep learning curve and the user needs help getting started.
Avoid chat when tasks are well-defined and frequent, like “optimise the query” or “add a trend line”. Avoid it when users need predictable, low-latency interactions through keyboard shortcuts or command palettes. Also avoid it when the chat interface pulls users out of the workflow where the real decision is being made.
A Hybrid Path Exists
Start with an initial chat interface to discover which requests users ask most. Then embed the top use-cases as dedicated UI controls: buttons, shortcuts, context menus, and other in-flow affordances. Keep chat around for outliers.
Notion is a useful example of this pattern. General chat is available, but many common actions also show up close to the writing surface: summarize, translate, rewrite, brainstorm. The product does not force every useful AI action through a blank text box.
Let AI fade into the background
Chat bots showed what is possible with AI, but they are not the right container for every task. For routine, repeatable work, put the help near the action. Keep chat for discovery, onboarding, and tasks where dialogue is actually the work.
What worked for us was simple: start broad with chat, learn what users repeat, then turn the useful repeatable actions into controls inside the product.