Siya Zanwar
Question Journey Finding Implications Next test Evidence
SIYA ZANWAR
UCL School of Management · MSc business research
Digital banking · Identity verification · UK

Customer Trust in
Digital Onboarding

I analysed app reviews and Financial Ombudsman decisions to understand negative experiences with automated identity verification. Among comparable low-rated complaints, verification failures more often described missing explanations or human recourse. The findings informed requirements for clearer decisions, alternative verification routes and escalation.

Individual MSc research project Public secondary data Observational, not causal
01

The research question

Digital banks compete on speed, and automated identity verification is what makes fast account opening possible. It is also the control that decides who gets in. I wanted to know what actually goes wrong for customers when that check fails.

I screened 450 Monzo and Starling app reviews and read Financial Ombudsman decisions covering a wider set of UK providers. The central comparison uses 74 low-rated complaints: 25 about verification and 49 about everything else. Each was coded for journey stage, failure type, and whether it described opacity or an absent route to a human.

02

Mapping the verification journey

Verification failure is not one event, so I mapped where it happens and what changes for the customer at each stage.

01

Entry

Documents will not upload, images cannot be read, codes do not arrive. The customer blames the app.

02

Verification

Document, selfie, biometric or liveness checks reject the applicant. The failure now feels like a judgement.

03

Waiting

No status, timeframe or control over what happens next.

04

Outcome and recourse

A decision arrives. Is there a reason, a correction path, or a person to reach?

03

Key finding

Verification complaints were not more intense than other complaints. They described a different kind of failure.

52%13 of 25

of verification complaints described opacity, an absent route to a human, or both

18%9 of 49

of other low-rated complaints in the same range described the same thing

Coded for opacity and/or absent recourse. No difference in complaint intensity was detected by star rating or sentiment. Observational study of selected complaints: these are not bank-wide rejection rates or population estimates, and the comparison shows association, not cause. Statistics in the evidence section below.

04

Product implications

Proposed requirements, derived from the complaints analysed. None has been tested.

01

Explain

Give a general, compliant reason wherever one can lawfully be given, so a rejection is not silence.

02

Correct and escalate

Offer a visible, reachable review path when automated verification repeatedly fails.

03

Accommodate

Provide alternative routes for customers whose documents, names or circumstances do not fit the expected path. Several Ombudsman cases involved exactly this.

05

Next test

The obvious next step is a comparison I have not run: show customers two versions of the same rejection, one with a general reason and a review route and one without, and measure whether they attempt the correction path. That tests whether explanation changes behaviour, which this study could not.

06

Evidence and method

Sources, sample and coding

Sources. Public Google Play reviews for Monzo and Starling, and published Financial Ombudsman Service decisions covering a wider set of UK providers. The two strands are not the same provider set and are reported separately.

Scope. 450 reviews were screened. The central comparison uses 74 like-for-like low-rated complaints, 25 verification-related and 49 other. Ombudsman decisions were read qualitatively alongside them. The counts are not additive and do not represent 450 individual people studied.

Inclusion. Cases include freezes and closures that occurred after an account was opened, not only account-opening rejections. Where that distinction matters it is kept rather than collapsed into one claim.

Coding. Each complaint was coded for journey stage, failure type, and whether it described opacity and/or absent recourse. Five recurring failure types emerged: document rejection, biometric or liveness failure, waiting and uncertainty, opaque rejection or freeze, and exclusion with denied recourse.

Statistics

Opacity and/or absent recourse: 13/25 (52%) versus 9/49 (18%). Continuity-corrected χ² = 7.43, p = .006. Odds ratio 4.8, 95% CI 1.7 to 14.0. Cramér’s V = .32. Fisher’s exact agrees to the same rounded p value.

An odds ratio of 4.8 describes the odds of a complaint being coded this way. It does not mean 4.8 times as many customers lost trust.

Intensity: Mann-Whitney U on star rating p = .33, and on VADER sentiment p = .69. No difference was detected. With this sample only a large difference would have been detectable.

Limitations
  • App reviews over-represent very satisfied and very dissatisfied users.
  • The 25 versus 49 comparison could reliably detect only a large intensity difference.
  • Observational design. The research identifies association, not causal effect.
  • Random re-coding by the same coder produced Cohen’s κ = .58. This is not independent inter-rater reliability.
  • English-language Google Play data, collected in a defined window.
  • Ombudsman cases are escalated disputes, not typical onboarding experiences.
  • Several cases involved vulnerable customers or reasonable-adjustment failures. The sample is far too small to support any claim of disproportionate impact, and none is made.
Source pages from the dissertation
Dissertation page showing the star-rating baseline and the like-for-like intensity test
Intensity test

No detected difference by rating or sentiment.

Dissertation page showing the 52 versus 18 per cent comparison with chi-square, Cramer's V and odds ratio
Character of failure

52% versus 18% opacity or absent recourse.

Dissertation page showing practical implications and the stated boundaries of the study
Implications and limits

Explainability, human review and reasonable adjustments.