Interest Is Not Suitability: Auditing Predictive Targeting for Digital Wealth Services

Name the target before accepting the score

A dashboard may label a person a good candidate for robo-advice when it actually predicts only that the person will click, express interest or accept a sales conversation. These are different propositions. A family office evaluating digital services should require the vendor to state exactly what the score predicts, how the label was obtained and what action the score triggers. A high likelihood of adoption says nothing by itself about the suitability of an investment, the quality of advice or the person’s understanding. The office should establish this distinction before sharing beneficiary information or permitting automated outreach.

The study’s useful and bounded contribution

Orzeszko and Piotrowski compared tree-based models for predicting acceptance of robo-advisory services using survey data from Poland [1]. Their predictors included demographic and socioeconomic characteristics, attitudes toward technology and experience with banking services. The study demonstrates an approach to an acceptance-prediction problem; it does not validate a family-office suitability assessment. The authors also discuss the possibility that models become obsolete as technology and customer attitudes change. For an institutional buyer, the central question is not whether prediction is possible, but whether the prediction is being used for a purpose that its evidence supports.

Build a label and action register

For every predictive score, record the unit of analysis, outcome definition, observation window and source of the label. Distinguish stated intention from actual use, sustained use, informed use and satisfactory outcomes. Then record the action: invitation, service access, recommendation, pricing or escalation. A marketing label should not silently migrate into an eligibility rule. Require the vendor to identify which inputs are necessary, which are inferred and which can be corrected by the person concerned. The register should include missing-data treatment and a description of who can see the score. Limit access to the people who need it for the approved purpose.

Test the model against the decision it will support

Review performance on a population and time period resembling the proposed use, and compare a simple baseline with the more complex model. Examine false positives and false negatives in terms of their consequences, not only aggregate accuracy. An interested person sent an irrelevant offer may face a different harm from someone excluded from a service they need. Ask whether calibration holds when the prevalence of interest changes and whether small groups are represented adequately. Do not infer fairness from one averaged metric. If the vendor cannot explain the validation population or cannot distinguish survey intention from observed behavior, keep the score outside consequential decisions.

Separate invitation from authorization

An invitation to learn about a service should open a conversation, not manufacture consent or bypass an adviser’s obligations. The person should be able to decline, correct relevant information and request a human explanation. For any consequential use, establish a separate review of the person’s objectives, understanding and authorized mandate using qualified professionals where appropriate. Do not reuse sensitive family information collected for another purpose merely because it improves prediction. Require an auditable record of the decision and the person responsible. A family office can use analytics to improve service design while refusing a conversion-driven definition of client welfare.

Implementation ownership and review

Owner: Data governance lead. Action: Create a register linking each score to its target, labels, inputs and permitted action. Evidence: Label definitions, data sources and approved-use record. Measure: Scores with ambiguous targets or unapproved secondary uses.

Owner: Independent model reviewer. Action: Compare validation performance and error consequences with the intended population. Evidence: Baseline comparison, calibration checks and subgroup coverage. Measure: Consequential false classifications and unexplained performance drift.

Owner: Client service owner. Action: Keep outreach separate from suitability and authorization. Evidence: Consent pathway, human review and complaint log. Measure: Propensity scores used as suitability evidence and unresolved corrections.

Monitor drift and the cost of being wrong

Maintain a reference report identifying model version, approved data, target definition and expected error pattern. Reassess it when the service changes, the population changes or actual response patterns diverge from expectations. Monitor complaints, unwanted targeting, corrections and any case where staff treated a propensity score as suitability evidence. Pause the affected workflow when the target or purpose has drifted. These are proposed governance controls, not findings tested by the original researchers. The practical standard is straightforward: a predictive tool should answer the question it was evaluated to answer, and a human should own every additional inference.

Source, attribution and limits

[1] Witold Orzeszko; Dariusz Piotrowski. Prediction of robo-advisory acceptance in banking services using tree-based algorithms. PLOS ONE (2024-05-06). https://doi.org/10.1371/journal.pone.0302359. CC BY 4.0.

Predictive performance is specific to the survey population, labels, period and validation design. Contextual generalizability and rapidly changing customer attitudes require further evaluation; prediction does not establish causality or suitability.

Prepared by the Family Office Private Wealth Journal editorial research desk with AI assistance at the direction of Dr. Freedom Cheteni. Original research authors retain full credit and are not authors or endorsers of this companion. This proposed process has not been evaluated in a new empirical study; no independent peer-review acceptance is claimed. This is educational analysis, not individualized investment, legal or tax advice.

Implementation record

Open the cited sources

1. [1] Witold Orzeszko; Dariusz Piotrowski. Prediction of robo-advisory acceptance in banking services using tree-based algorithms. PLOS ONE (2024-05-06). ↗