AI agents are usually framed as workers: generating screens, writing code, sorting research. But one of their more interesting roles may be to remember what a team was unsure about.
An AI agent could earn its place in a design review by keeping a disputed assumption visible until someone checks it. That's a useful brief for a team that can already generate more screens than it has time to evaluate.
Sunil Pai gives that role a name in Every company needs a cassandra, published August 9, 2026. His proposed agent watches workplace conversations and raises evidence-backed objections when the stakes justify interrupting. Pai's premise is social: repeatedly challenging colleagues can damage a person's standing, making useful dissent hard to sustain.
Cassandra is a concept in an essay. The point is not to object constantly. Pai wants it to step in selectively, using evidence the team may have overlooked and keeping earlier assumptions from disappearing. He also warns that management could use automated criticism as an excuse to stop listening to people.
For design teams, I'd build the experiment around the review itself. Consider a hypothetical navigation redesign: the room agrees that fewer visible choices will make the product easier to use. The proposed interface looks calmer. Someone mentions that occasional users might struggle to find the hidden options. The review moves on.
A useful intervention would attach that concern to the decision: which users does the team expect to help, and what observation would show that the change failed? Those questions give the next research session a purpose. Generating another layout leaves the uncertainty exactly where it was.
What should an AI agent remember about a design decision?
Pai proposes keeping a record of assumptions and predictions so the organization can revisit what it believed before knowing the outcome. Applied to the navigation example, that record would preserve the expected benefit, the objection, and the evidence available at approval. I'd want the original research linked beside the agent's interpretation, with a clear way to correct that interpretation.
The distinction matters when the work returns for another review. A screenshot can show what changed. It can't tell a new designer whether the team hid those options because of usability findings or because the launch deadline ruled out a larger revision. Without that context, an old compromise can acquire the authority of a design principle.
That gives designers a concrete interface problem. Where does an unresolved assumption live after the meeting ends? A comment can disappear into a resolved thread; a useful decision record should keep the concern visible long enough for someone to check it.
Attention belongs in the brief too. For this experiment, I'd ask the agent to bring one unresolved, consequential assumption to a scheduled review, with its source and a question the team can answer. A permanent stream of objections would create another inbox to clear.
Then evaluate the intervention. Did anyone correct the agent's reading? Did the concern change the research plan? The team should be able to dismiss an objection with a reason and recover it later if new evidence arrives. A confident paragraph from a model shouldn't become an extra approval layer.
Before building Cassandra, try the smallest version at your next review: write down one expected user benefit, one unresolved concern, and a date to check both. Keep the original wording when the results come back.
Solving the decision problem
Discussion