EY and Aaru are now formal partners. On September 30, 2026, EY announced an alliance between Aaru, an AI behavioral simulation company, and Ernst & Young LLP (EY US) to help clients test decisions on simulated populations before acting (EY newsroom). The alliance follows a blinded test in which Aaru recreated EY's 2025 Global Wealth Research Report, a six-month survey of 3,600 affluent investors in more than 30 markets, in a single day, reaching a median Spearman correlation of 0.90 across 53 single-choice questions (EY insight). Facts below were checked on EY and Aaru pages in October 2026.
For research buyers, the useful questions are what was actually measured, what the alliance changes, and how to evaluate a managed simulation partner against self-serve tools. This page separates the published facts from the marketing language. Minds is a synthetic research platform and an alternative to Aaru; we say where Aaru's model fits better.
In August 2025, EY's Wealth & Asset Management team challenged Aaru to recreate the 2025 EY Global Wealth Research Report without seeing its results. The original study surveyed 3,600 affluent investors across more than 30 markets and took six months of fieldwork; Aaru's case study describes the fieldwork cost as seven figures (Aaru case study).
Aaru built simulated investors from demographic, behavioral and sentiment data drawn from national censuses, financial institutions and social media, and the run took one day. EY then compared the two datasets on 53 matched single-select questions using three metrics:
| Metric | Published result | What it tells you |
|---|
| Median Spearman correlation | 0.90 | Answer options were ranked in a very similar order |
| Average difference per question (RMSE) | 7.1 percentage points | Typical gap between simulated and surveyed answer shares |
| Average Euclidean distance | 12.9 | Distance across multi-option answer patterns |
| Strongest agreement | Trust, value perception, personalization | Topics where simulation tracked the survey |
| Largest divergence | Inheritance planning (−37.82% correlation), recommendation likelihood, portfolio review frequency | Topics where simulation and survey disagreed |
Source: EY insight, October 2025.
EY and Aaru argue the divergences reflect the gap between what people say and what they do. Asked whether they would keep their parents' advisor, 82% of heirs in the survey said yes; the simulation predicted 43%, and industry studies cited by EY put actual retention at 20% to 30%. Likewise, 69% of surveyed investors said they prefer a single financial provider, the simulation predicted 37%, and EY cites 33% of high-net-worth individuals actually using one advisor.
How to read these results
- The test was designed and published by the two companies involved, not peer-reviewed or independently replicated.
- It covered single-choice questions only; open-ended reasoning, pricing trade-offs and new-product concepts were not part of the comparison.
- A median rank correlation of 0.90 means the order of answers usually matched; the 7.1-point average gap is the better guide to how far individual percentages can be off.
- The "simulation was closer to real behavior" argument rests on a few questions and on third-party industry figures, so treat it as a promising signal rather than proof.
The alliance announced on September 30, 2026 pairs Aaru's simulation engine with EY US's sector and functional teams. According to EY, Aaru generates populations of AI agents that represent a defined audience, and EY US helps clients turn the results into strategy. The first named focus is financial services, where the two firms are testing marketing messages, product and pricing offers, sales and service strategy, and the impact of mergers and acquisitions (EY newsroom).
For buyers, the practical change is the delivery route: enterprises can now reach Aaru through an EY US consulting engagement, with EY's industry data and advisory teams around the simulation, in addition to working with Aaru directly.
Aaru describes itself as a behavioral simulation platform trained on real-world behavioral outcomes. In the EY study, it built agents with behavioral architectures that model decision-making traits, placed them in simulated environments, and asked them the survey's questions. Aaru says the same simulated population can be rerun after an event such as a rate change or a competitor's move, projecting shifts in sentiment within 24 hours (Aaru case study).
Aaru does not publish self-serve pricing; access runs through enterprise sales or, now, the EY US alliance. For a side-by-side view of Aaru and self-serve options, see Minds vs Aaru and the wider list of Aaru alternatives.
To evaluate any synthetic research platform objectively, teams must understand the statistical and methodological boundaries of synthetic populations. Synthetic outputs are directional. They do not establish representativeness, causal proof, forecast demand, exact willingness to pay, or replace recruited participants for final high-stakes validation.
Simulated agents generate responses from patterns in their training data, demographic priors, and the context they are given. Run across several scenario options, they highlight relative trade-offs, potential messaging friction, and qualitative rationales. These directional signals help narrow a large set of ideas before committing primary research budgets.
However, synthetic simulations cannot replace empirical reality in the following areas:
- Representativeness: Language model personas do not reproduce the exact demographic and psychographic distribution of a target population in a statistically representative way.
- Causal proof: Simulating a response to an isolated stimulus does not isolate causal mechanisms or account for real-world confounding variables.
- Demand forecasting: Stated simulated preferences cannot accurately forecast actual adoption rates, market penetration, or purchasing volume.
- Willingness to pay: Synthetic agents do not face real financial constraints, so pricing simulations cannot determine absolute price elasticity.
- Final high-stakes verification: Critical strategic moves, major capital allocations, and public disclosures still require verification with recruited human participants.
Treating synthetic outputs as iterative, exploratory inputs rather than definitive forecasts protects teams from over-relying on synthetic data in executive decisions.
When introducing synthetic research, enterprise buyers generally choose between two operating models: managed consulting engagements and self-serve simulation platforms.
Managed consulting engagements, such as the EY US and Aaru alliance, involve dedicated external teams who configure custom agent populations, ingest proprietary enterprise data, calibrate the simulation, and deliver curated strategic reports. This approach suits broad corporate transformations, scenario planning, and complex strategic initiatives where an organization prefers an advisory partner to manage the modeling.
Self-serve software platforms put simulation directly in the hands of internal product managers, UX researchers, and brand marketers. Internal teams configure their own Audiences, test concepts on demand, and run structured research without waiting for consulting cycles. This model prioritizes speed, autonomy, and continuous testing throughout the product lifecycle.
The right choice depends on team structure. Organizations with internal researchers who iterate daily on messaging or features benefit from self-serve access. Organizations seeking an outsourced strategic engagement often prefer managed advisory work.
| Dimension | Managed consulting | Self-serve platform |
|---|
| Primary user | C-suite sponsors, external strategy consultants | Embedded market researchers, product managers |
| Execution cadence | Project-based, multi-week strategic cycles | Continuous, on-demand research iterations |
| Workflow control | Outsourced setup and analysis | Direct researcher control over Audiences and Studies |
| Ideal use case | Scenario analysis, enterprise transformation | Concept screening, message testing, priority ranking |
Minds is the end-to-end platform for commercial synthetic research. It gives internal product and UX, service design, market research, Voice of Customer, marketing, innovation, and agency teams direct control from audience creation and study planning through stimulus testing, qualitative and supported quantitative methods, comparison, analysis, and export. Stimuli can include Figma inputs where enabled, websites and app flows, images, video, copy, decks, questionnaires, and concepts.
Within Minds, teams build reusable Audiences of AI personas, called Minds, and run Studies with open-ended, single choice, multiselect and scale questions. Researchers can interview an individual Mind or gather reactions from a whole Audience at once, keeping context across sessions.
For structured quantitative testing, Studies run MaxDiff for relative priority, conjoint for configured trade-offs, Kano for feature categories, and Van Westendorp and Gabor-Granger for pricing. Like the EY test, Minds compares simulated answers with real survey results: Audience Validation scores each Audience against published surveys or your own survey files, with a score out of 100, a 95% range and the questions it left out.
Minds does not claim representative output. Human observation, physical or sensory testing, regulated evidence, and final high-stakes validation can supplement that workflow when required; they do not make Minds a point solution. Minds is self-serve, with Pay as you go at $0.12 per response and Pro at $199 per user per month; Aaru is the better fit when you want a managed, consulting-led simulation for a large strategic decision.
When evaluating synthetic research software or managed simulation vendors, procurement and insights teams should apply a framework across four operational pillars.
| Evaluation pillar | Key verification requirements |
|---|
| 1. Managed consulting vs self-serve platform | Define internal skills, the need for external services, and turnaround expectations |
| 2. Methodology and study configuration | Confirm MaxDiff and conjoint, structured study design, and separation of chat from method runs |
| 3. Evidence inspection and auditability | Inspect agent prompts, full response distributions, and versioned persona definitions |
| 4. Human validation path | Set recruitment for final checks and gate high-stakes decisions on live data |
Pillar 1: Delivery Model Alignment
Buyers must decide whether they need strategic advisory support or ongoing research enablement. Managed services suit episodic, high-level assessments where an executive sponsor wants an end-to-end deliverable. Self-serve platforms suit embedded teams that need continuous concept screening, narrative iteration, and persona interaction.
Pillar 2: Structured Research Methods
Exploratory chat alone is not enough for quantitative trade-off decisions. Check whether a platform runs formal designs such as MaxDiff for ranking feature importance and conjoint for multi-attribute packages, and whether it keeps informal querying separate from experimental designs.
Pillar 3: Evidence Inspection and Auditability
Enterprise governance needs visibility into how synthetic conclusions are produced. Verify whether you can inspect persona prompts, demographic assumptions, and raw transcripts, and whether the vendor publishes how its accuracy metric is defined, as EY did with Spearman correlation, RMSE and Euclidean distance.
Pillar 4: Human Validation Protocols
Treat synthetic simulation as a top-of-funnel filter, not a replacement for fieldwork. Document when a finding can proceed on directional synthetic data and when it must be validated with human panels. High-stakes launches, regulatory submissions, and major pricing changes should pass through recruited human validation first.
A mature research stack does not rely on a single vendor or method. It positions synthetic research as an upstream accelerator that makes downstream human recruitment more focused.
In discovery, researchers use AI personas to explore customer language, map objections, and brainstorm value propositions. In concept screening, teams run structured methods such as MaxDiff or conjoint on synthetic Audiences to cut dozens of concepts to a shortlist.
The shortlisted concepts then go to human testing with recruited panels. Because weak options and confusing messages were removed earlier, human fieldwork becomes cheaper and more focused on confirming results.
By keeping realistic methodological boundaries and pairing self-serve tools like Minds with disciplined human validation, insights teams speed up research while keeping decision integrity. Teams ready to try structured synthetic research can register for access.
