Beyond the Chat Box: Designing AI Experiences That Actually Fit How People Work

This is the third post in my series recapping takeaways from Reforge's AI Strategy course. Part 1 covered how AI accelerates PMF collapse; Part 2 introduced the AI Maker Matrix for user segmentation. This week, I'm diving into something that's been gnawing at me since the early days of ChatGPT: why do so many AI products feel like a novelty you try once and never return to?

The answer, according to Reforge (and I wholeheartedly agree), is that we've been building AI products around the technology rather than around how humans actually work.

The Chat Box Trap

When ChatGPT launched, it set a template: user types a prompt, AI responds, repeat. That interface worked brilliantly for introducing the world to generative AI's capabilities. But as a foundation for product design it has serious limitations.

Chat interfaces suffer from a few core problems. First, they're ambiguous about task boundaries. Unlike a document editor with a clear "start typing here" moment, chat leaves users wondering what to ask, whether they're done, or if they should start a new conversation. Second, chat forces users to translate visual or spatial intent into words—describing the image manipulation you want rather than just doing it. That translation burns mental energy that should go toward the actual work. I've seen this in my own work at Ontra, where we considered building in a chat interface to an AI document editing product. We ultimately decided against it, because if users have the time to type what change they'd like into a chat box, why not just make that change yourself directly in the document?

Finally, most AI chat products operate as walled gardens, disconnected from the tools and data where real work happens.

The result? That initial excitement about AI capabilities fades, and users drift back to traditional tools that better match how they actually think and work that mirror their mental models.

Digital-Native vs. Human-Native Interfaces

This brings us to a paradigm shift worth internalizing: the difference between digital-native and human-native interfaces.

Digital-native interfaces—the buttons, dropdowns, toggles, and form fields we've been designing for decades—mirror how computers think. They're built around binary states, enumerated choices, and structured fields that databases can easily ingest. Users have to learn the metaphors, build mental mappings from intent to operations, and manually perform tasks the computer can't anticipate.

Human-native interfaces flip that relationship. Instead of forcing users to think like computers, AI enables interfaces that interpret human intent directly. Natural language, context awareness, visual recognition, and semantic understanding let interactions align with how people naturally work.

Consider the transformations this enables:

• Video editing goes from precise timeline manipulation to "remove all the ums" and "trim the section where we talk about pricing."
• Photo editing goes from layers, masks, and complex parameters to "make the background blurrier" and "brighten the subject's face."
• Data analysis goes from SQL and pivot tables to "show me sales trends by region for the last three quarters with highest growth highlighted."

Descript is a great example of this shift. It reframes video editing around the transcript: editing text edits the video. You delete a word, that segment disappears from the video. You fix a verbal mistake by typing the correction. It maps perfectly to how people think about spoken content by using words, not frames and timestamps. I've used it myself when editing video segments for distribution across Ontra and it works like a charm!

Finding Friction: Where to Innovate

So how do you identify where human-native design can transform your product? Reforge recommends a systematic friction analysis across three dimensions:

Physical Load: Map the interactions required for common tasks. If publishing a blog post with an image takes eight clicks through your CMS, that's physical friction. Look for data entry overhead, deep navigation paths, duplicate entry across systems, approval workflows with multiple handoffs, and wait times that compound.

Cognitive Burden: Identify where users' mental energy gets drained. Context switching between apps or mindsets. Working memory demands—tracking configuration settings, task status, or complex syntax. Decision points where users lack the information to proceed confidently. Features that most users never discover because the learning curve is too steep.

User Workarounds: Watch for the unofficial systems users create to compensate for your product's gaps. External spreadsheets for tracking what should happen in-app. Scripts and macros for automation you don't provide. Templates and boilerplate they maintain themselves. Third-party tools replacing features you technically have but that don't work well enough.

These workarounds are a golden compass pointing you to exactly where you should be innovating. They reveal where users are investing effort to bridge the gap between how they want to work and how your product forces them to work.

Design Principles for AI Experiences

Once you've identified high-friction opportunities, a few principles guide effective AI experience design.

Start with users' natural workflow. Align AI features with how people already think about the task—don't make them learn a new mental model just to use your AI. The discovery questions here are: How do users describe this task in their own words? What mental shortcuts do they use? What feels unnatural or frustrating? What would an expert human assistant do?

Balance structure and flexibility. Too much structure limits AI's power; too little increases cognitive burden on your users. The right balance depends on your users. Domain experts (Prosumers, if you will) typically prefer flexibility. Occasional users (Delegators) need guidance. Complex tasks benefit from structured workflows with AI at decision points. Creative tasks favor open canvases with suggestions.

Build clear feedback loops. Make quality visible where appropriate—confidence scores, alternatives, sources. Collect explicit feedback (thumbs up/down) and pay attention to implicit signals (edits users make to AI output, which alternatives they choose, when they abandon the result entirely). Most importantly, close the loop: show users how their feedback improves results. This creates a collaborative feel rather than a black-box experience.

Plan for failures. AI will make mistakes. Design for graceful fallbacks that maintain user progress. Enable easy correction that lets users edit, refine, or reject without losing work. Set expectations through interface and messaging so users aren't surprised when AI errs.

Manage the cost/quality/speed triangle. Higher quality tends to increase cost and slow speed. Faster responses may reduce quality or raise cost. Lower cost often means lower quality or slower speed. Know which dimension matters most for your use case and test user thresholds.

All this to say, the chat box is not the ultimate enemy and should be applied where it makes sense. However, the chat box is not the silver bullet that will fit every product experience.

Bringing This to Your Teams

If you're building AI products, here's a simple exercise you can run with your team:

1. Pick a core workflow. Choose something users do frequently where you suspect friction exists.
2. Map the friction. Walk through the task counting clicks (physical load), noting context switches and decision points (cognitive burden), and inventorying workarounds users have created.
3. Ask the human-native question: If an expert human assistant were doing this task, what would they do? How would they know what the user wanted? What would they just handle automatically?
4. Design for amplification, not replacement. Unless you're certain your users want full delegation, design AI to enhance what users already do rather than taking over entirely.
5. Plan for the failure cases. What happens when AI gets it wrong? How do users recover without losing progress?

Wrapping It Up

I hope these posts have been helpful! If you've employed any of the tactics from this post or the last two with your teams, let me know! I found Reforge's course to be extremely helpful in reframing the way I think about building AI features with my team. If you want the full takeaways, I'd definitely encourage taking their course. These posts only scratch the surface of the wealth of knowledge they pass on.

If you've applied any of the learnings from this series, let me know in the comments below! I'm also happy to answer any questions that arise. T