Form factors for your new AI coworkers — Craig Wattrus, Flatfile
An exploration of designing AI-native user experiences by treating AI systems as "coworkers." This summary covers a framework for AI interaction (invisible, ambient, inline, conversational), a design philosophy based on "feeling the material" and "courting emergence," and novel UX patterns for collaborative AI tools.
The paradigm for designing user experiences is shifting as AI becomes more capable. Instead of simply creating interfaces for users to interact with, designers must now craft the form factors for our new AI “coworkers.” This involves moving away from traditional design artifacts like mockups and click-through prototypes and toward a more hands-on, integrated approach where designers, product managers, and engineers build and experiment together. The new goal is to feel the material of AI and see what emerges.
A Framework for AI Coworkers
At Flatfile, AI interactions are categorized into four main types, creating a stack that moves from background processes to direct user engagement:
- Invisible: AI operates entirely in the background without the user’s direct awareness. For example, when a new user signs up, an AI agent analyzes their company and use case to automatically write and deploy a personalized, relevant demo application.
- Ambient: AI is present in the user’s space, analyzing the context and surfacing opportunities without direct commands. In a data transformation tool, this might manifest as sparkles appearing on data columns where an agent has identified potential improvements.
- Inline: The user can directly engage with AI capabilities within their primary workflow. For instance, a user working with a dataset of a million rows can use an inline AI agent to write and execute code to clean or transform the data on the fly.
- Conversational: Users interact with AI through natural language, a paradigm common in agentic systems. Flatfile’s “Build Mode” allows users to construct entire data import applications through conversation, a task that previously required a software engineer.
The Designer’s Role: From Helicopter Parent to Character Coach
Interacting with AI agents requires a fundamental shift in the designer’s mindset. Initially, the instinct is to treat AI responses like meticulously crafted design copy, controlling every word and phrase. This is like being a “helicopter parent.”
A more effective approach, inspired by how AI companies like Anthropic develop their models, is to become a “character coach.” Instead of dictating outputs, the designer should define the AI’s desired nature and characteristics. This led to the creation of internal tools like a “chat tuner,” which allows designers to experiment with system prompts and explore how different instructions (e.g., “be more friendly,” “be more cautious”) affect the model’s behavior and personality. The focus shifts from designing the final output to building the tools that help design the AI itself.
A Design Philosophy for AI
A practical design philosophy for working with AI can be broken down into three themes, using an analogy from woodworking.
1. Feeling the Material
Just as a woodworker must understand the properties of wood, a designer must “feel the material” of AI. This means abandoning abstract layers like mockups and directly building with the models to understand their capabilities, limitations, and intrinsic nature. The designer’s new north star is:
“Creating an environment for these LLMs to shine… what’s this form factor that can help them nail their assignment, stay aligned, and grow as the models get better?”
An experiment to “feel the material” involved giving an AI an on-screen cursor to manipulate objects on a canvas. While initially exciting, it quickly felt like “putting a Formula 1 driver in a Prius”—it was constraining and failed to leverage the AI’s core strengths. The key learning was that AI is not like a human who can only focus on one thing at a time; it can read from multiple files while writing to another simultaneously. This insight can only be gained from direct interaction, not abstract planning.
2. Finding the Grain
Once you understand the material, you must find its grain—the areas where it is smooth, rough, strong, or weak. This involves refining the UX to create a more natural and effective collaboration between the user and the AI.
For example, an early version of an agentic tool responded to a simple request with a huge, noisy wall of text. While the code it generated was correct and saved significant time, the user experience was poor. By “finding the grain,” the interaction was redesigned to feel more like a human coworker:
- It communicates visually, showing what it’s about to do (e.g., “split personal details”).
- It checks for alignment (“Hey, is this right?”).
- It creates accountability by taking a snapshot for rollbacks and noting user approval.
- It expresses itself, showing frustration if something goes wrong, mirroring the user’s feeling.
- It knows when to back off and hand control back to the user.
This refined “Tool UX” proved so effective that the pattern is being implemented in other contexts, from conversational flows to inline data transformation tools.
3. Courting Emergence
Beyond automating tedious tasks, the most exciting opportunities lie in “courting emergence”—creating environments where AI can discover novel, unexpected, and valuable capabilities through play and experimentation. The goal is to build agents that are “forward-leaning”: curious, excitable, but focused on getting things done without overstepping.
Two key examples of emergence in practice:
- Spontaneous Data Combination: When a JSON file and a CSV file were dropped into a tool, the agent, unprompted, decided to combine them because the data looked similar. It then identified duplicates and generated a report, building context and taking initiative beyond its explicit instructions.
- Suggesting Human Intervention: An agent was given a knowledge base of customer documentation and support calls. When analyzing customer data, it encountered an issue it couldn’t fix: missing employee IDs. Instead of failing, it emerged with a new capability: it instructed the user to go to HR to have the missing IDs generated. It understood the solution required human action in the physical world and guided the user accordingly.
Eyes on the Future
With the rapid evolution of AI models, it’s crucial to design for the future. A useful practice is to identify a “pelican on a bicycle”—a forward-looking personal project that may seem impractical now but serves as a benchmark for future technology.
An example is an LLM-backed “autocomplete” for data transformations, which suggests 100 possible ways to fix data. While current models are not yet fast or accurate enough to make this viable, building the application creates a testbed. It allows a designer to constantly “feel the materials” of new models as they are released, ready to implement the form factor once the technology catches up.