Research methods

Understand the current state before redesigning it. Lab-based usability testing, where you watch someone click through a prototype and ask what they think, misses what matters here. People describe their own workflows badly. They skip steps they run on autopilot, forget the workarounds, and report following the standard procedure while leaning on intuition.

The designer works as a workplace anthropologist, using immersive methods to surface the invisible work that AI can absorb. Two things get mapped from the start. Tacit knowledge: the judgment experts apply without writing it down. Decision rights: what each person is authorized to decide, what evidence they treat as enough, when they seek cover from a superior, and which approvals are real safeguards. Many workflows look like information systems and run as liability-distribution systems. Automate the information and leave the liability architecture untouched, and the result is timid AI: smart enough to recommend, never trusted enough to act.

Contextual inquiry and cognitive task analysis

Shadowing tends to record what a user does on screen: click A, open B, type C. Because AI automates cognition, the designer maps what the user is thinking. An experienced underwriter evaluating a policy is recognizing patterns in risk, drawing on years of cases, and cross-referencing sources in their head. The designer deconstructs that judgment: which memories, which data points, which edge cases cause a pause. Mapping the cognitive workflow turns human heuristics into system prompts and agentic reasoning, so the agents support the hard decisions.

Glue work and shadow IT

Some of the richest ground is a company's shadow IT: the unofficial systems people invent to route around their official software. The designer hunts for sticky notes on monitors, undocumented spreadsheet macros passed around by email, and chat channels where people ask each other policy questions. In legacy workflows, humans are the connective tissue between disconnected systems, exporting from one, reformatting, pasting into the next. Every such workaround marks a broken seam and points to where an agent could read one screen, interpret the data, and populate the other.

Shadow IT is also the company's real ontology. The unofficial spreadsheet shows the categories people actually use. The workaround shows the question the official knowledge base never answered. The sticky note shows the risk signal no database field captured. Mine these artifacts for meaning before replacing them, or the new system will automate the official process and lose the practical intelligence that kept the business running.

The exception landscape

Standard procedures cover the routine 80% of the work. People spend most of their time and attention on the 20% of exceptions and edge cases, where traditional automation gives out and generative AI does well. The designer watches how experts investigate anomalies, captures the intuition and outside research they use, and designs guardrails that let the system handle most future anomalies on its own.

Outcome mapping

Ask people what they want and they describe a faster version of the tool they have, the proverbial faster horse. The magic wand question gets past that. Ask: if you had an infinitely capable colleague who knew everything about this company, worked instantly, and never erred, how would you brief them, and what would you do while they worked? The answer separates intent, the business goal, from action, the steps taken today. That intent is the blueprint for what the system needs to achieve.