Reconciliation Complaints Get Refunds Nearly Three Times as Often as Other Fintech Disputes

When a consumer files a complaint with the Consumer Financial Protection Bureau about a fintech payment problem, the company response falls into one of three buckets: closed with an explanation and no payment, closed with some non-monetary fix, or closed with monetary relief, meaning the company paid the consumer to resolve the dispute. That last category is the strongest outcome the CFPB records. It reflects a company’s own determination, made under regulatory review, that the consumer was owed money.

A study of 97,028 complaints pulled from the CFPB Consumer Complaint Database filed against eleven major US consumer fintechs between January 2021 and December 2025 found that complaints describing a reconciliation failure, cases where two systems disagree about the same transaction, close with monetary relief at 9.96%. Complaints describing everything else close with monetary relief at 3.51%. The gap holds in direction across every product category in the dataset and ranges from 1.4x to 4.0x depending on the type of product involved.

What Counts as a Reconciliation Failure

The study defines reconciliation as the process of confirming that two or more records of the same money movement agree. A reconciliation failure is what happens when those records diverge in a way the consumer can see: a transfer that leaves one account and never arrives in another, a duplicate charge that never gets reversed, a refund the merchant logged but that never posts to the consumer’s balance.

This is a narrower category than “payment problem.” A complaint about an unauthorized transaction is a fraud complaint, not a reconciliation complaint, because both systems may have recorded the transaction correctly; the failure is in who authorized it, not in whether the records agree. A complaint about a frozen account is an access complaint. A complaint about a disputed fee is a pricing complaint. Reconciliation failures are specifically about disagreement between records of the same event.

To classify complaints at this scale, the study used a reasoning-capable large language model to read each complaint narrative alongside its CFPB metadata (product type, sub-product, issue category, company, and company response) and assign a binary label: does the described sequence of events match a recording, matching, settlement, or ledger disagreement, regardless of whether the consumer used the word “reconciliation” or correctly identified the cause themselves. A consumer who writes “they stole my money” but describes a settlement failure is classified as reconciliation-related. A consumer who writes “my transfer was lost” but describes what reads as a fraud dispute is not.

The classifier was validated against 80 hand-labeled complaints, coded independently by two people with disagreements resolved by discussion. The final model reached 88.75% accuracy (95% CI: 80.0%-94.0%), with precision and recall balanced at 89.19% and 86.84%. Confidence intervals were calculated using the Wilson score method. Balanced precision and recall matter here specifically because it means the model isn’t systematically over-flagging borderline complaints to inflate the reconciliation share, or under-flagging them to suppress it.

The Monetary Relief Gap

Of the 97,028 complaints, 7,880 were classified as reconciliation-related. Table 1 shows how those complaints resolved compared to the other 89,148.

Reconciliation-related complaints are also about 6 percentage points less likely to close with an explanation alone, meaning companies more often concede a financial obligation when these specific disputes get escalated to a federal regulator.

There are two non-exclusive explanations for the gap. The structural one: reconciliation failures tend to be traceable. A duplicate charge can be matched against the original transaction. A transfer that left one account but never arrived in another can usually be traced through provider logs. That traceability makes it more feasible for a company, or the CFPB, to identify exactly where records diverged and reach a financial determination. Fraud and access disputes are harder to resolve from documentary evidence alone, because the central question, whether an action was authorized, isn’t something transaction records settle on their own.

The behavioral explanation: companies may simply treat reconciliation failures as unambiguous operational errors they’re obligated to fix, while treating fraud and access complaints as disputes with more room for judgment calls.

Neither explanation implies that this gap describes all reconciliation failures happening inside these companies. The 7,880 classified complaints are the fraction that got escalated to a regulator. The unmeasured population includes discrepancies resolved before a consumer ever noticed, duplicates that closed without a refund, and settlement breaks written off internally. The CFPB data only measures what reaches that federal complaint pipeline.

How Common Are These Complaints

Across the full five-year dataset, reconciliation-related complaints are 8.12% of all complaints. That pooled figure isn’t the headline prevalence number in the study, and the reason is specific: the 2025 denominator is distorted.

In December 2024, the CFPB sued Early Warning Services, the operator of Zelle, along with Bank of America, JPMorgan Chase, and Wells Fargo, alleging failure to protect consumers from fraud on Zelle’s network. That action generated a wave of near-identical template complaints about Zelle service issues rather than transaction-level errors. Total fintech payment complaints in the dataset rose roughly sevenfold from 2024 to 2025, with 67,585 of the study’s 97,028 complaints, nearly 70% of the five-year total, filed in a single year. Reconciliation’s share of that inflated 2025 pool drops to 5.39%, not because reconciliation failures declined, but because the denominator ballooned through a category of complaint that has nothing to do with transaction-level errors. The raw count of reconciliation-related complaints kept rising anyway, from 1,537 in 2024 to 3,642 in 2025.

*2025 share excluded from prevalence estimates due to the denominator distortion described above; the raw 2025 count is included for transparency.

Using the four years before the distortion, reconciliation-related complaints represented between 12.95% and 15.64% of fintech payment complaints annually, and 14.39% pooled across 2021-2024, or roughly one in seven complaints. They weren’t just tracking overall complaint growth either: total complaints in the dataset roughly doubled from 2021 to 2024, while reconciliation-related complaints more than doubled over the same stretch, meaning their share of the total was climbing on its own.

As a sensitivity check, applying the classifier’s precision (89.19%) to the 2021-2024 count produces a lower-bound estimate of about 12.8%; applying recall (86.84%) to account for missed positives produces an upper estimate of about 16.6%. The reported 14.39% sits inside that range.

Where Complaints Concentrate

Reconciliation-related complaints cluster at the highest-volume platforms in the study.

These are raw counts, not rates. The CFPB dataset has no transaction volume or active-user denominator for any company, so these numbers scale with platform size and complaint propensity as much as anything else. They shouldn’t be read as a ranking of operational quality, and the study makes no claim that any named company has confirmed unresolved reconciliation failures. What the concentration is consistent with is a structural pattern: the platforms at the top of this list route payments across the most external banks, processors, wallets, and internal ledgers, and more system boundaries mean more chances for two records of the same event to disagree.

By Product Category

The near-3x monetary relief gap holds across every product category in the dataset, though the size of the gap varies.

Money transfer shows the widest gap, at roughly 4x, which tracks with the structural explanation above: money transfer involves multi-party routing across external processors, wallets, and bank accounts, which creates more boundaries where records can diverge and, correspondingly, more traceable discrepancies once they do. Checking accounts and card products show the narrowest gap, around 1.4x, which may reflect that simpler transaction flows in those categories more often get resolved through ordinary account processes before a dispute ever reaches the monetary-relief threshold.

What the Gap Implies

The core implication is upstream of the CFPB. Monetary relief reflects a company’s own decision to pay, not a regulatory order. If reconciliation failures are consistently identifiable and compensable after a consumer has already escalated to a federal regulator, that same traceability should, in principle, make them detectable and resolvable before the complaint is ever filed, provided the company has transaction-level visibility across the systems the payment touched. The near-3x gap is really a statement about the evidentiary clarity of these disputes once you go looking, not about how often they occur in the first place.

Limitations

The study is explicit about what it can’t establish. It classifies consumer-described events, not internal ledger data. A complaint’s narrative may point toward a reconciliation failure whose actual internal cause was something else, and the study makes no claim about what happened inside any specific company’s systems. Complaint counts by company scale with platform size, not failure rate, since there’s no transaction denominator in the CFPB data to normalize against. The 2025 numbers are distorted by the Zelle enforcement action and are excluded from prevalence estimates for that reason, though included in full for transparency. The 80-complaint validation sample produces reasonably wide confidence intervals, particularly on recall and specificity; the study notes that a roughly 200-complaint sample would tighten the accuracy interval to about plus-or-minus 5 points. And the monetary relief rate only covers complaints that got escalated: discrepancies resolved before a consumer noticed, or absorbed internally as write-offs, don’t appear in this dataset at all.

Conclusion

Reconciliation-related complaints close with monetary relief at 9.96%, against 3.51% for everything else, a gap that holds across every product category studied. In the four years before a regulatory event distorted the 2025 numbers, these complaints made up roughly one in seven fintech payment complaints annually. The most likely explanation is that reconciliation failures come with traceable transaction-level evidence that fraud and access disputes generally don’t, which makes them easier for a company to resolve with a definitive financial payout once a regulator is involved.

The full classified dataset, the validation set, and the CFPB source query used to reproduce the underlying 97,028-complaint dataset are published for independent audit.

How Rexi approaches this

Rexi is an agentic reconciliation platform built for fintechs, neobanks, payment companies, and marketplaces that move money across fragmented systems. It ingests transaction data from processors, banks, wallets, ledgers, and ERPs, reconciles across sources, and surfaces discrepancies before they become consumer complaints.

Two capabilities are directly relevant to the failure pattern this data describes. Exception management identifies transactions that fail to match, routes them for investigation, and tracks resolution at the transaction level. Revenue leakage detection catches duplicate charges, settlement breaks, and unrecovered fees before they accumulate. Both depend on transaction-level visibility across every record boundary the payment crosses.

The CFPB data in this analysis measures the cost of those gaps after the fact. Reconciliation software addresses them before the consumer notices.

Data availability: - Classified dataset (97,028 records, narratives removed for privacy) - CFPB source query (reproduces exact 97,028-complaint dataset) - Classification prompt available on request