The Human Insights Engine turns real-world evidence into a living evidence and simulation environment that supports organizations across the life of their programs, policies and products. It sharpens hypotheses faster, focuses real-world engagement where it counts, and informs the analyses behind how resources are committed. Built for decisions that affect people's lives.
Sharpen hypotheses soonerTest options against evidence-grounded virtual personas in a bounded virtual world, broad or narrow, sectoral or thematic, before committing resources at each stage.
Focus real-world engagementAim field research, pilots and community engagement at the people and questions where they matter most.
Keep evidence alive for yearsOne sandbox, refreshed with new data, informing decisions from first design to scale and beyond.
The challenge
Decisions don't stop after launch. The evidence behind them rarely keeps up.
Programs, policies and products run for years, and organizations commit resources and draw on data all the way through. But evidence is gathered at one moment and ages quickly. Using it well depends on teams knowing what to draw on, when, and when to commission something new. The challenge is more than having relevant data. It is keeping evidence usable as decisions, conditions and questions change.
01
Different moments need different evidence
Early strategy needs the broader picture; design needs niche segments; scale-up needs recent data from the places you are expanding into.
02
Research is fragmented and goes stale
Surveys, monitoring data, evaluations and local knowledge sit across programs and partners, and are rarely brought together at the point of decision.
03
Assumptions go untested, and new data comes late
Teams move forward without knowing which assumptions carry the greatest risk, or commission research after the decision it should have informed.
How HIE addresses this
Evidence that stays usable, from first design through scale
HIE brings your evidence together into a living evidence and simulation environment: virtual personas in a bounded virtual world, built around your work and refreshed as new data and implementation results come in. Teams return to it at each decision point, for resource allocation and decision-making across the life of a program, policy, product or investment.
Illustrative use-case journeys
How HIE saves time and effort across a multi-year journey
HIE shortens research and design steps, cuts repeated re-research as work scales, and points field work at the people and questions where it matters most. Each journey shows the full arc, then zooms in on one decision. All examples are illustrative composites.
Headline benefit
Zoom in on one decision · phase
What HIE does
Turn your data and new research into tested options before each decision
HIE brings research data and real-world observations together with contextual understanding, Dalberg's problem-solving expertise and human intelligence into a bounded simulation built around your work. It is set up once, then refreshed as new data comes in, so every later decision is tested against the latest evidence rather than an aging study.
Two inputs keep the environment useful over time: what your organization already knows, and what it continues to learn.
What you already knowYour existing dataProgram, field and implementation data, team knowledge, and unstructured files such as spreadsheets, transcripts, audio, video and reports.
What you continue to learnNew primary researchNew research periodically refreshes the environment, particularly where behavior, markets, policies or implementation conditions are changing quickly.
Supported by
Public contextual dataSurveys, census, DHIS2, national health surveys, CGIAR, Findex, climate series, NGO and multilateral reports.
Dalberg dataTwo decades of behavioral and psychographic research across LMICs, used where consented and available.
The Human Insights EngineEvidence-grounded virtual personas in a virtual world, with a confidence score on every finding
Used again and again for
Sharper hypothesesbefore each design or investment choice
Prioritized segmentsas the population you serve changes
Early warning of drop-offwhere awareness, trust, access or follow-through break down
Tested options and messagesbefore each launch, region or revision
Focused field researchknowing what to commission, and when
Input to resource decisionsallocation, costing, M&E and decision memos
Why behavioral data matters. Demographics tell us who people are. Behavioral and psychographic evidence helps explain why people may respond differently to the same intervention. Every simulated response is decomposed against seven dimensions.
Comprehension
Relevance
Trust
Feasibility
Social permission
Emotional readiness
Practical constraints
What you get back · illustrative
Predicted sign-up, risk-averse farmers
Pay-as-you-harvest
62%
Upfront credit
41%
Virtual World Confidence
81/100One segment flagged as thin evidence
2 priority barriers identified
1 segment needs additional field research
The loop back
Field results
Update evidence and virtual personas
Test the next decision
Deploy
Learn again
Pilot, monitoring and research results flow back in, so each next decision is tested against what actually happened.
What this can reduce or avoid
Narrows many hypotheses to a few worth field testing
Avoids repeating full segmentation or baseline studies at each stage
Targets primary research at the highest-uncertainty questions and segments
Your data, governed
Each dataset is used according to the permissions, consent basis and governance applicable to it.
Personally identifiable information is removed or appropriately protected before use in simulation environments.
Client deployments run with agreed access controls and limits on how outputs can be used.
Simulated outputs are not fed back in as training data, and offline deployment is possible.
Why HIE
Proven in commercial markets, built for development
Synthetic personas already have real traction in commercial settings such as consumer goods and FMCG, so the approach is not untested. But those tools are usually built to drive purchases in high-income markets. HIE is built to improve outcomes for people in low- and middle-income contexts.
Majority-world led
Not adapted from market research tools tailored to the Global North. Grounded in real-world decision-making models that reflect diverse, lived experience across the majority world.
Actionable and outcome-driven
Designed to illuminate pathways to human impact, not conversion or transactional behavior change. Structured around the frameworks social-sector decision-makers use, like theories of change.
A partner, not a dashboard
Not a consumer synthetic-persona tool or a self-serve dashboard. HIE comes with Dalberg's advisory team, connecting results to your decisions and other analyses.
Proof points
96%average similarity in selected backward-validation studies in four countries. See validation
1.6 Mnpeople's research data represented in the evidence base
3.31 Bnrecorded observations and variables
59.5 Bnsimulation runs to date
Using it well
How HIE works best: alongside field evidence and human judgment
HIE focuses and complements real-world research; it does not replace it. It adds the most value when sharpening hypotheses, comparing design and delivery options, surfacing likely friction, and deciding where real-world research is worth commissioning. Other research, analysis and judgment lead on:
Establishing causal impactSimulations give directional signals; whether an intervention works needs real-world testing.
Understanding lived experience in fullVirtual personas cannot replace direct engagement with people and communities as contexts change.
Feasibility and operational testingField validation is still needed for implementation realities, uptake and sustained use.
Exercising expert and contextual judgmentTeams and experts frame the questions, review assumptions and interpret results.
Decisions requiring human judgmentHIE informs, but does not determine, funding, policy, program or individual decisions.
How we work together
One joint operating rhythm, three roles
Each decision is simulated, connected to supporting analysis and research, acted on, and the new evidence flows back in. Dalberg runs the environment and advises alongside you; you own the decisions and the work on the ground.
Dalberg HIE team: environment upkeep
Set up and calibrate your sandbox
Bring in new data as it arrives
Maintain virtual personas, scenarios and confidence scoring
Dalberg: advisory partner
Turn insights into decisions
Connect HIE with your other analyses and processes
Advise on resource allocation, new investments, and when to commission new research
You and your partners
Bring your data, context and questions
Own the decisions
Run field research, pilots and implementation, and share results back
What you commit
Data
Access to relevant existing data, and to people for new primary research where gaps matter.
Team time
A named lead and regular working sessions, built around your decision calendar.
Time to first result
Sandbox set-up comes first, typically a few weeks, before the first decisions are tested.
Scale
Start with one program, policy or product, and extend the same sandbox as the work grows.
Supporting detail
How HIE compares with other persona and simulation approaches
Most approaches are strong on one thing: readable personas, scale, social realism or survey alignment. HIE combines research grounding, explicit behavioral representation, computational simulation and applied decision support.
Structured models, not prompting
Simulation is driven by structured behavioral and decision models, rather than prompting alone.
Virtual personas in a virtual world
Not just personas, but the world they live in: households, institutions, services, money, information and social norms, with person-to-person, person-to-institution and person-to-object interaction.
Stable and accurate
Results are checked for stability across repeated runs and back-tested against real-world studies, so findings hold when the same question is asked again.
Capability comparison of persona and human-simulation approaches
Approach
Research grounding
Machine-readable schema
Behavioral and psychographic variables
Dynamic state and memory
Localization and overlays
Scenario and intervention testing
Multi-agent and panel simulation
Trust and adoption modeling
Go-to-market fit
Published human-fidelity validation
Core strength / key limitation
Full or nativePartialNot presentDalberg self-assessment, based on published descriptions of each approach.
Schedule a live demo to integrate HIE in your work
Bring a live question and we will walk through how HIE could support it, now and over the years ahead. Contact prerak.mehta@dalberg.com
Note: The underlying training and grounding data for the Human Insights Engine is anonymized and ethically sourced. The outputs are directional hypotheses, and not real responses. They inform, rather than replace, contextual expertise, direct engagement, co-design, real-world testing, and accountable human judgment. Journey examples, durations and outputs are illustrative composites. Dalberg · Confidential.
From real-world evidence to dynamic virtual worlds
How the Human Insights Engine turns evidence into bounded, context-aware, persistent in memory, and auditable virtual personas, and how it checks that what they tell us holds up.
Real world evidence, simulated feedback, informed decisions.
HIE is the simulation layer within a living evidence environment. Once established around a program, policy, product or investment, the same evidence base, virtual populations and behavioral models can be updated and reused as new questions, data and implementation results emerge.
This means HIE is not rebuilt for every question. It becomes an evolving environment that teams can return to across design, implementation, adaptation and scale.
HIE does not write virtual personas as character descriptions or prompts. It converts real-world evidence into machine-readable behavioral agents, places them in structured social and research environments, and runs repeated simulations under explicit rules, constraints and confidence checks. The result is simulation that is bounded by evidence, persistent in memory over time and auditable.
01The evidence base
Every virtual persona, behavior, interaction and sandbox in HIE is grounded in real-world evidence: qualitative research, surveys, population datasets, longitudinal records and other public and private sources. HIE has an ever-growing evidence corpus.
3.31 BnRecorded observations and variablesUsed to ground virtual personas, behaviors, interactions and sandboxes
1.6 MnPeople's research data representedRepresented directly in the underlying observational and research studies
1.67 BnPopulation representedThrough weighted cohorts built on that evidence
59.5 BnSimulation runs to dateEvaluations, state transitions and simulations run on the platform
94–99.5%Similarity in selected backward-validation studiesKenya, India, Bangladesh and Nigeria; see section 06
That evidence is mapped into a common schema. It is never pasted directly into a language model prompt. From there, one logic chain runs through the whole engine:
Evidence
Virtual persona schema
Instantiation
Rules
Interaction
Cohort
Instrument
Simulation
Confidence
Validation
02How the engine was trained
HIE began as a shell: machine learning code that defines how a virtual persona is constructed, with metrics across four categories: demographics, psychographics, behaviors and context. On its own, the shell knows the shape of a virtual person but nothing about specific characteristics that define it.
We trained it on a wide range of real-world data, including human-centered design primary research, in-field observations, household and national surveys, census data, quantitative research studies, clinical and program data, longitudinal series data, and more. Large language models (LLMs) made the training fast and feasible, reading and structuring very different sources and mapping them into the shell's common schema. That structured evidence trained the machine learning code and let us find the right weights, turning a static shell into a dynamic machine learning schema that can generate new virtual personas whenever new data is loaded.
Data-sets by Dalberg
Engagements over 25 years
Dalberg brings over two decades of primary research experience on macroeconomic, demographic, behavioral, and psychographic factors across multiple LMICs. We use the insights to co-design and problem-solve with communities, governments, social/private enterprises, CBOs, and NGOs.
National household and health surveys
Population baselines
For example, GAVI / UNICEF / WHO datasets, NFHS-5 (India), DHS (Kenya, Nigeria, Ghana, Tanzania, Bangladesh), NAIIS (Nigeria), IHDS-II and NSSO 75th Round, plus census and open government data, among others.
Health research and clinical evidence
Behavior in care
For example, IBBSS key population surveillance and PrEP cohort studies (Nigeria), Nivi.io data from Uganda/India etc., the National Mental Health Survey of India, the HAP and Atmiyata community mental health trials, and conversational maternal health data, among others.
Livelihoods and enterprise
Economic lives
For example, DAY-NRLM self-help group registries, the PRADAN women's collectives panel, IWWAGE enterprise diaries, Dalberg's "Becoming Someone" aspiration study, and rural digital enterprise field audits, among others.
Agriculture and climate
Farms and weather
For example, smallholder agronomic and input adoption studies in Kenya, agricultural extension records, and India Meteorological Department climate and heat data, Digital Green data from their primary research, longitudinal ten-year climate-heat series data across countries, among others.
Education
Learning
For example, ASER learning outcomes, alongside national school health and well-being frameworks; Acumen Academy published results, Teach For All and country-level published results, national education surveys, among others.
The above are a few examples of the types of input data we have used to train the model. The Human Insights Engine doesn't retain the prior training data. The training taught us the quality and quantity of data needed to generate new virtual personas and virtual worlds for a specific problem. That formula is what let us build demo centers, mini virtual worlds dedicated to specific problem-solving spaces.
Zero-dose immunization
TB vaccine readiness
Maternal health (ANC and PNC)
Heavy menstrual bleeding
Community mental health
Women's digital enterprise
SHG leadership
Smallholder farming
Climate and heat stress
Financial security and fraud
Classroom pedagogy
This has helped us codify the approach needed to build any new set of virtual personas and virtual worlds for specific problem areas, and do it quickly and efficiently. It tells us what kind of data to collect first. Once that foundation is built, it can be reused to solve many more problems, not just the first one.
Two principles that keep the engine honest
Principle 1
Trained only on inputs
When we use a study, we train only on its original responses, never on the study's outputs or findings.
Principle 2
Simulated data never goes back in
Nothing the engine generates is ever fed back into the platform as training data.
Both guard against the same failure. Feeding in simulated data risks hallucination compounding over time. Feeding in a study's outputs risks the model, over time, referencing those answers instead of reasoning from what people actually said and did.
It also means there is only one route to improving HIE: more detailed, more comprehensive, higher-quality input data. When a simulation needs to be more accurate, we improve what goes in, never by feeding back what comes out.
Inputs are the past, a study's outputs are the present, and simulations are the future.Trained only on inputs, the engine has no sense of past, present or future. Reproducing a study that has already happened and anticipating one that has not are the same task, which is why HIE's validation results are a strong anchor for trusting its simulations.
The animation below walks the build end to end: from an empty shell, through training and calibration, to hundreds of distinct virtual personas interacting inside dynamic virtual worlds. Use the controls to pause or step through it.
03How a virtual persona is built
A virtual persona in HIE is a structured state model, not a character sketch. Each one is represented through four layers, which together decide what it can be, how it tends to behave, what can change, and where every part of it came from.
A · Fixed core
Who the virtual persona is
The segment boundary it cannot casually drift outside.
Building a virtual persona follows five steps: retrieve evidence, map it to the virtual persona schema, calibrate attributes and behavioral ranges, instantiate stochastic latent variation, and validate constraints and provenance.
Calibration is where the shell becomes dynamic, and it never involves feeding in outputs. Simulated behavior is compared with what was observed in the real world. Where accuracy falls short, the weights for new virtual personas are adjusted by feeding in more detailed, more comprehensive, higher-quality input data until simulations reach the accuracy needed. Once calibrated, the same machinery can take in new data and generate new virtual personas without being rebuilt.
Stochastic Latent Seeding (SLS) is what turns one segment into many distinct virtual personas. The segment defines the behavioral envelope, and SLS decides where a particular virtual persona starts inside it. Two virtual personas from the same segment might begin with slightly different levels of trust, household pressure or risk tolerance, while both staying inside the calibrated evidence boundaries.
Virtual persona schema = what this person can plausibly be. SLS = which plausible version is instantiated for this run.This is why HIE produces a distribution of plausible responses rather than one deterministic answer.
What a virtual persona is, and is not, allowed to do
Hard constraints do not drift.Core demographic and structural attributes stay fixed unless a simulation explicitly models their change.
Variation stays inside calibrated boundaries.Stochasticity creates heterogeneity, not arbitrary behavior.
Dynamic states change through events and interaction.Trust, intent, relevance, emotion and feasibility move as the scenario unfolds.
Evidence gaps remain visible.Where evidence is thin, the system surfaces uncertainty rather than silently generalizing.
State transitions are traceable.Changes in state, assumptions and evidence used in a run are logged.
The model does not redefine the virtual persona.The underlying language model is frozen and works inside boundaries set by the engine.
04From virtual personas to virtual worlds
Individual virtual personas become useful when they are placed in context. HIE layers four structures on top of them.
The interaction graph governs how virtual personas affect each other. HIE does not simply concatenate responses. Each exchange is interpreted through relationships, trust, authority, social norms, influence and previous interactions, and can shift dynamic states such as trust or social permission.
Units and cohorts assemble virtual personas into households, panels, networks and communities against a sampling frame, with weights so the cohort represents a larger real population. Aggregation does not erase individual state.
The instrument runner applies qualitative, quantitative and specialised research methods across the cohort. A virtual persona can move along a behavioral pathway, from exposure and attention through comprehension, credibility, personal relevance, social consultation and intent to action or drop-off. HIE records where and why drop-off occurs, not just the final answer.
The sandbox engine is a persistent environment for repeated and counterfactual simulation. Conditions such as price, service availability, policy, weather, rumours or delivery channels can change virtual personas' dynamic states without rewriting the evidence base. The same world can then run scenario A against scenario B, before and after, Monte Carlo runs, sensitivity tests, or a historical replay.
How virtual personas learn during a run
Current state›receives a stimulus, question or interaction›interprets it through evidence, priors, trust and context›decides or responds›sees an outcome, social response or new information›updates its dynamic state›next decision
Virtual personas are persistent in memory within a simulation, but the underlying foundation model is never retrained.
The engine decides; the language model speaks
HIE engine determines
Evidence retrieval, virtual persona state, constraints, behavioral parameters, decision pathway, state transitions, sampling and weighting, and simulation conditions.
Frozen language model provides
Situated reasoning, interpretation, natural-language expression and dialogue, using the evidence placed in its context.
A structured decision policy produces the choice or propensity. The language model then explains it in the virtual persona's voice. It expresses the virtual persona; it does not define it. This separation makes simulations more reproducible, auditable and less dependent on the wording of an individual prompt.
05Confidence at four levels
HIE scores confidence at four levels, because a well-grounded virtual persona can still give a weak answer on an unfamiliar topic, and a strong answer can still come from a poorly designed study.
Level 1 · Virtual persona build
Is this virtual persona sufficiently grounded?
Evidence coverage, source quality, schema completeness, recency, calibration quality, provenance and how much inference was needed.
Level 2 · Response
How trustworthy is this specific answer?
Retrieval relevance, evidence availability, state consistency, stability across reruns and unsupported inference.
Level 3 · Panel or study
How reliable is the aggregated finding?
Population coverage, sampling-frame fit, weighting, segment representation, method fit, variance across runs and sensitivity to assumptions.
Level 4 · Virtual World Confidence
Is this simulation fit for this question?
Confidence in the quality of the simulation set-up, not the probability that a simulated outcome will occur.
The Level 4 score, the Virtual World Confidence, is a weighted composite of six components:
Virtual persona fidelity: how well they represent the population and its heterogeneity
Method fit: whether the method suits the question
Delivery and execution fidelity: channels, stimuli and journeys realistically represented
Assumptions: importance and uncertainty of what is not grounded in evidence
Audit and validation: traceability, stability tests, backward validation, expert review
06Validated against real-world studies
The clearest test of HIE is backward validation: give the engine only the inputs of a completed real-world study, and compare what its virtual personas say or do with what real people said or did. Across four exercises spanning agriculture, financial inclusion and health, HIE reached 94% to 99.5% benchmark similarity, with a substantially lower sample size and over multiple runs of the same study.
Kenya · Farmers95%
Digital Green, FarmerChat survey by 60 Decibels. Roughly 2–8% variance from real-world findings across tested survey measures.
India · Mothers and frontline workers97%
Government of Maharashtra, NFHS-5 antenatal care. 31,643 households, 33,755 women and 5,497 men on comparable ANC indicators.
Bangladesh · National population94%
BRAC, nationwide study on financial behaviors. 3,592 respondents across all eight divisions and six behavioral segments.
Nigeria · Female sex workers99%
Lagos University Teaching Hospital, HIV PrEP study. 310 respondents; mean positive response 59.35% human versus 59.83% simulated.
Backward validation sits alongside five other checks: holdout comparison against real observations not used to build the virtual personas; repeated-run stability across many SLS instantiations; sensitivity testing of assumptions, weights and conditions; expert review of behavioral mechanisms and delivery assumptions; and field calibration, where small real-world samples of original responses are added as new input to recalibrate parameters without replacing the evidence base.
Remains controlled and versioned
Source evidence, virtual persona schema, fixed demographic core, structural constraints, model weights, simulation rules, provenance, sampling frame.
Can change during a simulation
Trust, relevance, emotion, intent, perceived risk, feasibility, social permission, memory and interaction history.
07What this means for decisions
HIE lets a team test a policy, program, product, service, or message against thousands of evidence-grounded virtual personas before committing resources, see where and why people would drop off, and compare scenarios side by side. Its outputs are directional hypotheses with an explicit confidence score and a traceable evidence trail. They sharpen where to invest, what to test in the field, and what to ask real people next, rather than replacing them.
Dalberg · Human Insights Engine · Confidential. Animation labels marked "Illustrative" show example variables, outcomes and sectors, not specific model outputs.