Market + NewsHuman composes the order
→Product research case study · 2026
Strata
How far should AI go
in expert equity trading?
An interactive research system testing how assistance—from evidence delivery to an editable order draft—changes expert judgment and trading workflows.
3AI assistance levels
2Market scenarios
102v3 analyzed sessions
17Expert personas
01 · Framing the problem
The question was not the feature.
It was the depth of intervention.
When generative AI moves from summarizing evidence to drafting action, it changes both the decision anchor and the boundary of responsibility.
The goal was not to decide whether AI belonged in the product. We needed evidence for which level—raw evidence, synthesized insight, or an editable order draft—should become the product default.
The real question was how far AI could intervene while preserving expert judgment.
Accuracy only explains the output.
To understand product value, the study also had to capture what experts reviewed and edited, when they rejected or deferred, and how they perceived final responsibility.
01Model
Check output accuracy, consistency, and format compliance.
02User
Measure expert comprehension, trust, control, and pressure.
03Behavior
Record time, evidence views, edits, acceptance, and deferral.
04Product
Assess adoption intent, workflow value, and required safeguards.
02 · Choosing the context
Choose a context where
judgment leaves a trace.
Expert trading is a high-stakes decision made with incomplete information: compare evidence, define whether and how to trade, then own the outcome.
The process reveals more than a single answer. It shows which evidence was reviewed, how a draft changed, and whether responsibility stayed with the expert—making anchoring, overconfidence, and responsibility transfer observable.
01Anchoring · AI-proposed price and quantity can become the reference point for later judgment.
02Overconfidence · Polished explanations can make uncertain evidence feel definitive.
03Responsibility transfer · Experts may shorten review or attribute the final choice to the system.
Initial contextLoan trading
The expert pool was narrow, and the task was difficult for outside readers to understand.
Research pivotU.S. equity trading
Preserved the core constructs while improving recruitment, comprehension, and realism.
03 · Research design
A 3 × 2 experiment that changed
only the depth of AI.
Across two market scenarios, the protocol compared how assistance depth changed decision speed, trading behavior, comprehension, reliance, and perceived control.
01Market & News
Experts review market data and news, then compose the order without AI output.
02AI Insight
AI summarizes the strongest evidence and uncertainty; the expert composes the order.
03AI Draft
AI prepares the order list; the expert edits and makes the final execution decision.
Manipulation checks separately measured usefulness, influence on judgment, recognition of AI-generated output, and product realism so template artifacts would not be mistaken for treatment effects.
EXPERIMENT ARCHITECTUREHold the evidence constant. Change only the assistance.
3 × 2NVDA+2.8%
- Price movement
- News context
- Portfolio position
AI InsightAI summarizes evidence
→AI DraftAI prepares an editable order
→SideBUY
Quantity20
04 · Hypotheses & measures
Translate hypotheses into
interactions and logs.
The study connected preference with decision time, evidence views, order input and edits, approval, rejection, deferral, control, and reliance within a single session.
H1Speed
AI Draft reduces decision time versus Insight and no AI.
H2Reliance
AI Draft increases draft acceptance and system reliance.
H3Ambiguity
Under ambiguity, Draft can reduce review or invite misplaced reliance.
H4Control
AI Insight preserves more perceived control than Draft.
H5Bias
No AI requires more manual work but reduces automation bias.
01Behavior · Decision time, evidence views, input, edits, final choice, and execution.
02Decision experience · Information quality, comprehension, confidence, uncertainty, and pressure.
03Human agency · Perceived control, responsibility, and excessive reliance.
04Product value · AI usefulness, judgment impact, adoption intent, and product realism.
05 · Building the research system
Turn a demo into
an experiment system.
Real equities and news, researcher-approved AI snapshots, condition-specific surveys, centralized collection, and a research console became one validation system.
- 01Realistic inputs
Used listed equities, completed daily market data, and news metadata from established publishers.
- 02Approved AI artifacts
Researchers reviewed insights and order drafts before exposing them as fixed snapshots.
- 03Behavioral instrumentation
Linked decision time, evidence views, input, edits, confidence, and final action to one session ID.
- 04Central QA
Reviewed responses and events in one console and exported analysis-ready raw data.
102 / 102Completed sessions + surveys
17 × 6Personas across experiment cells
0Approval-label errors outside Draft
99 / 102Unique response patterns
These 102 sessions were not a human-subjects study. Seventeen synthetic expert personas were repeated across six experiment cells to validate branching, surveys, event logs, and the analysis pipeline.
06 · Pilot evidence
Insight supported evaluation.
Draft reduced the work.
These descriptive results test whether the instrumentation can distinguish conditions. They prioritize hypotheses for a human study; they are not claims about human behavior.
84%Lower input workload
Average input and edit actions fell from 4.76 to 0.76 with Draft.
5.06Highest information quality
Insight outperformed the no-AI condition at 3.96.
6.50–6.57Control remained stable
Perceived control barely moved as automation depth increased.
01AI lifted information ratings: Insight scored 5.06 and Draft 4.91, versus 3.96 without AI.
02Draft reduced average input and edit actions from 4.76 to 0.76, the largest drop in composition effort.
03Less work did not always mean a faster decision. Evidence-review time and input reduction need separate analysis.
04Execution stayed between 38.2–47.1% and control between 6.50–6.57, while reliance rose modestly in AI conditions.
05Participants recognized AI-generated output, but product realism remained moderate at 3.82–3.94.
DIRECTIONAL INSTRUMENTATION PILOTAssistance improved evaluation without removing the final human decision.
n=102INFORMATION QUALITY/ 7
+1.10 Insight vs. no AI
INPUT + EDIT ACTIONSAVG.
−84%4.76No AI→0.76AI Draft
Draft removed repetitive order-entry work.
HUMAN AGENCYRANGE
6.50—6.57
Perceived control remained stable across conditions.
Execution rate38.2–47.1%
07 · Product decision
Separate default support
from optional automation.
Make Insight the default decision-support layer and offer Draft only when order composition creates meaningful workload. Both paths preserve evidence review and the final human decision.
01Insight
The default layer: higher information quality and adoption intent at a lower automation cost.
02Draft
An opt-in execution tool that reduces repetitive input and keeps every change visible.
03Guardrail
Explicit approval, a defer option, evidence traceability, and data timestamps remain mandatory.
011EvidenceMarket data + source context
022Insight by defaultSummary with uncertainty
033Draft on demandEditable order preparation
044Human approvalExplicit final action
Next validation sequence
01Compare approved AI snapshots with participant-facing market and news data field by field.
02Run cognitive interviews with 6–12 experts to test task comprehension and product realism.
03Use an expert-scoring rubric to separate acceptance from objective decision quality.
04Pre-register sample size, exclusions, analysis, and interaction tests before the human study.
05Test Insight as the default and Draft as an opt-in tool with human participants.
08 · Outcome & reflection
Before the feature,
build a system that can test it.
Strata’s core output was not the AI order feature itself. It was a research architecture that can reveal which level of assistance supports human judgment—and where additional safeguards become necessary.