Agentic AI needs UCD smarts
It’s been a while since anyone has uttered the words ‘Large Action Model’. If you’re a normie, you possibly last heard it in the same breath as Rabbit R1, the magic orange box that by most accounts was a flop.
However, Motorola has come out of the stocks at CES 2025 with a LAM that can order coffees and taxis on your smartphone, no other device needed. (It looks like my prediction about Android and iOS winning out could become true.) Teams are investing time in developing agents backed by large action models that can interact with the real world, and Motorola’s hoping it’ll be a key value proposition for them (presumably).
But when you look at how to develop agentic AI, as suggested in this paper from Microsoft, the online evaluation stage is called out as the most critical part of development.
A final but crucial step in the development of LAMs is evaluation (Xie et al., 2024). Before deploying a LAM for real-world applications, it is imperative to rigorously assess its reliability, robustness, and safety. Unlike LLMs, which may be limited to generating text-based outputs, LAMs have the capacity to directly affect their environment through actions. This introduces new risks, as incorrect or inappropriate actions could have significant consequences. Therefore, thorough evaluation processes are essential to ensure that both the LAM and its accompanying agent are capable of making reliable decisions while minimizing potential risks.
User testing is a key component of good product development and user-centred design practice. It assesses a design in real-world scenarios, highlighting where a design may need to change or iterate for its intended environment and user base.
Agent testing will require similar methods to ensure the agent responds as intended in the environment. Instead of observing and assessing human-computer interaction, teams will adapt user-centred methods for computer-computer interaction.
Product and design professionals who have worked in government or regulatory environments may be well suited to the job. Ethical risks are a big concern for agentic AI:
As these models gain the ability to interact with real-world environments, questions about accountability, transparency, and fairness come to the forefront (Ferrara, 2024; Liesenfeld et al., 2023; Li et al., 2023a).
I think that’s an interesting prospect for product and design people, who will be needed to join the ranks of AI teams as AI tools become more prevalent. To understand not only agents as users but also the users of agents.