Evolving Search for Intelligence

The User: Trained to Distrust

My users were intelligence analysts. Their profession trains them to distrust information, the consequences of being wrong are asymmetric and severe, and they have zero tolerance for a tool that obscures where an answer came from. A search experience for them had to earn trust at every step rather than assume it. That single constraint, design for the professionally skeptical, drove every decision that followed.

That distrust showed up as two user types I had to serve at once:

Yet underneath their different roles, both wanted the same thing: to find every piece of relevant information and answer their question comprehensively.

Two Opposing Search Styles

In practice the analysts split into precision seekers, who used rigid, syntax-heavy Boolean queries, and simplicity seekers, who wanted fast but less predictable semantic search. The challenge was to serve both camps without alienating either, and to never ask a skeptical analyst to trust a result they could not trace.

Semantic Foundation

I designed a single search entry point with primary and secondary pre-search filters, then proved semantic search worked on the customer’s data.

Result Comprehension

I built result views that kept the query editable, led with an AI summary, and mapped filters to the established user goals.

Boolean Power, Hidden by Default

I kept semantic as the default and added an opt-in Boolean mode to support librarians and legacy users’ existing saved-query workflows.

The Outcome

The MVP validated semantic search on the customer’s data and secured the contract. More importantly, early testing surfaced the signal that drove the rest of the work.

Unified Bar, No Mode Decision

I introduced a unified search bar. By moving execution options to after the user types, I removed the upfront cognitive load.

What This Didn’t Solve

This relocated the friction rather than eliminating it. The decision moved later in the flow but did not disappear.

Scoping the AI Bet

An AI-assisted conversational UX had been deferred earlier. ML Engineering favored step-by-step development, citing broad evaluation scopes and complex bug fixes. Rather than override that caution, I scoped the bet to de-risk it: I aligned with Product and ML Engineering on a single end-to-end vision we could evaluate as a whole, on a deadline none of us could hit incrementally. We agreed to build the full vision first and evaluate later.

Landing the Leap: AI-Driven Intent Inference

Taking that leap meant trusting the system to do the heavy lifting. I introduced an AI-powered search that translated plain language directly into execution. Instead of asking the analyst to choose, the system inferred intent, determined the method, composed the query, and set the filters automatically.

Earning Trust in a Non-Deterministic System

Non-determinism was the real risk, so I designed three mechanisms that prioritized trust alongside speed:

The Through-Line

I never eliminated control. Across four phases, I relocated it from a strict prerequisite to an optional override, letting the system assume the heavy lifting while keeping the analyst in the loop.

The Reusable Pattern

Underneath the components was a portable standard for any AI-mediated surface: set expectations, show the interpretation, preserve a fallback. I treated it as a reference, not a one-off for search.

The Agentic Evolution

As queries become tasks, the interface shifts from tool to collaborator. The next phase extends this foundation into agentic workflows, where the system runs the research itself, synthesizing sources and surfacing insight. That is only possible because the first phase was anchored in explainability and trust. I designed for where the models were going, not only where they were.