Designing trust into enterprise AI.

01Context

When AI-generated information becomes a business decision.

Organizations process large volumes of contracts, invoices and reports whose contents have to be extracted, checked and transferred into business systems. Machine learning can automate much of that. But automation brings a new problem: an extracted value can look right without being right — and in an enterprise, a wrong value travels into downstream processes and decisions.

Human verification stays essential. The person who reviews and approves the values is the product’s real user — and their confidence is its real feature. The design challenge was to make that verification efficient and understandable.

02The problem

The bottleneck wasn’t extracting information. It was verifying it.

The system could find the relevant information and predict values. Users didn’t distrust it because it made mistakes — they distrusted it because they couldn’t quickly check whether a prediction was correct. So even high-confidence predictions were checked against the original by hand, in the same loop, every session:

That context switching became the slowest part of the workflow. The problem wasn’t the interface’s appearance, and it wasn’t accuracy. It was the effort required to verify AI-generated information.

03My role

Designing the human–AI interaction.

As a senior UX designer working with product managers and machine-learning engineers, I was responsible for the review workflow — translating what the model produces into an experience that supports human judgment. My work covered:

• Evidence linking — connecting every value to its place in the source.
• Uncertainty — how model confidence is communicated, and what it asks of a reviewer.
• Review prioritization and the correction interaction.
• User research and usability testing with the people doing the reviews.
• Functional prototypes the team could test and build from.

Working this closely with the machine-learning engineers meant designing with the model’s real outputs — its confidence scores, its failure modes — rather than an idealized version of them. Alongside the product work, I mentored junior designers on the team.

04Design principle

Every prediction should be explainable.

The design process centred on one question. If a reviewer can answer it quickly, they can decide with confidence — and the design goal was to reduce the effort of answering it, and nothing else:

“Why did the AI produce this value?”

That question led to three interaction decisions, each one shown below in a working reconstruction.

05Decision one

Connect every prediction to its evidence.

Instead of separating extracted values from their source, the interface keeps both in view. Selecting a value highlights the passage it came from, and selecting a passage opens the value it produced — so reviewers verify instead of hunt. The demo carries all three decisions; try clearing the queue.

06Decision two

Turn uncertainty into priorities people can act on.

Machine-learning systems express uncertainty as confidence scores. But a percentage on its own doesn’t tell a reviewer what to do. So the interface translates confidence into review priority — surfacing the predictions where human judgment adds the most, and letting the ones the model and document agree on clear with less interaction, within the product’s verification rules.

The aim isn’t to hide uncertainty. It’s to make it understandable and useful: instead of asking someone to interpret a number, the interface says this one deserves your attention.

07Decision three

Make correction part of the normal workflow.

A wrong prediction shouldn’t send anyone into a separate debugging process. Reviewers correct a value in place while its link to the original evidence is preserved, and the correction becomes feedback for future training — without interrupting the flow. Correcting AI should feel like reviewing a document, not like debugging software.

08The experience

A review workflow designed around human judgment.

Together, the decisions connect four elements — and turn AI output into information a person can evaluate and act on.

• Prediction

  What the system extracted.

• Evidence

  Where in the document it came from.

• Uncertainty

  How much attention it needs.

• Action

  What the reviewer can confirm or correct.

09Outcomes

From opaque predictions to verifiable information.

By connecting every prediction to its source, reducing the context switching around it, and making uncertainty actionable, the review workflow became faster and easier to understand — better suited to an enterprise setting where accuracy is a liability question, not a preference. Verification stopped being the slowest part of the workflow.

• Before: Search the document for every value.→After: Every value shows its source.
• Before: A confidence score to interpret.→After: A priority to act on.
• Before: Corrections as a separate fix.→After: Corrections inside the review, fed back to training.

Specific figures are confidential. The review model on this page — evidence, priority, correction — is the part that travels.

10Reflection

Trust is something an interface has to enable.

I used to think trust in AI was a model-quality problem. It isn’t, or not only. People don’t need an AI system to appear certain. They need to know when its output can be relied on, when it needs their attention, and how to step in.

For enterprise AI, transparency isn’t a secondary feature. It is part of the interaction model.