Designing for Trust in AI Interfaces
When an agent makes decisions on behalf of a user, the interface has to answer a question that traditional software never faced: why should I trust this output?
A database query returns facts. An agent returns judgments. The design challenge is communicating the difference.
The trust gap
Users interact with Agentkit agents through dashboards, chat interfaces, and automated reports. In every case, the agent has made decisions: which data to retrieve, how to interpret it, what to flag, what to ignore.
If the interface presents agent output the same way it presents database output — as plain facts in a table — users will either over-trust it (treating judgments as facts) or under-trust it (dismissing everything because they cannot verify it).
Neither is acceptable. The interface needs to communicate confidence, uncertainty, and provenance.
Confidence indicators
Every agent output in the Agentkit dashboard includes a confidence signal. This is not a percentage — users misinterpret percentages as probabilities. Instead, we use a three-level system:
High confidence — the agent's output is directly supported by tool results with no ambiguity. The number came from a database query. The document was retrieved verbatim.
Medium confidence — the agent combined multiple sources or made an inference. The anomaly was flagged based on statistical analysis, but the threshold is configurable.
Low confidence — the agent is extrapolating or the data was incomplete. The recommendation is based on partial information and should be verified.
These levels are derived from the trace. If every step in the chain used structured tool output, confidence is high. If any step involved free-form reasoning without tool grounding, confidence drops.
Provenance trails
Every claim in an agent's output links back to the trace step that produced it. Click on a flagged invoice amount and the interface shows you: this was flagged by step 3, which ran a statistical outlier detection script on 1,284 invoices retrieved from Stripe, using a threshold of 3 standard deviations.
This is not a footnote. It is a first-class UI element. The provenance trail is visible by default, not hidden behind an expand button.
Editable outputs
Agent output should be a starting point, not a final answer. Every output in the Agentkit dashboard is editable. Users can dismiss a flagged item, adjust a recommendation, or add context. These edits are logged and fed back into the agent's memory for future runs.
This creates a feedback loop. The agent learns from corrections. The user gains trust through control.
The principle
Trust is not a feature you add to the UI. It is the result of making the agent's reasoning visible, its confidence honest, and its output editable. Design for trust from the first wireframe.