FinCEN 2025 BOI Report: 44% Blank Fields, HSBC Accused

Sixty-two percent of the newly disclosed beneficial owners in FinCEN's corrected BOI filings are managers with no recorded ownership. The number comes out of the unstructured substantial-control narrative, not the ownership-percentage field that most compliance workflows treat as the gateway. Those blank fields are the report's real finding.

The compliance industry's habit of testing only the ownership box misses the people who run the company without holding equity. A majority of newly disclosed owners fall into that category. Reading their roles requires NLP because the relevant text is buried in free-form narratives across large batches of records. Manual review cannot scale.

The stakes are visible in multiple enforcement tracks. HSBC Private Bank Suisse is under French indictment over alleged concealment of diverted funds, and SEC v. EquiAlt shows how control without ownership fed a Ponzi scheme that raised more than $170 million. The next generation of UBO checks has to read the story, not just the box. Compliance programs that ignore the narrative will keep reporting the same gap.

The Mechanism

Under FinCEN's beneficial-owner rule, the beneficial-owner definition has two independent prongs, and the second is where undisclosed managers hide. The ownership prong triggers at the regulatory ownership threshold. The substantial-control prong has no percentage floor — a managing member with minority equity and sole signatory authority is a beneficial owner even though a threshold-only ownership script never flags them. The myth that managers below the ownership threshold are not reportable collapses against the text: control is the test, not the share certificate. That second prong is the entire reason the gap covered above exists.

FinCEN's Form BOI makes the problem worse at the collection point. Line 3(c) asks filers to describe the nature of substantial control, but the e-filing system stores the answer as free text. Structured threshold scripts — the dominant compliance pattern at the start of 2026 — read only the numeric ownership field and never parse the description. A person whose control runs through "removal power over the CEO" or "sole signatory on the operating account" is invisible to those scripts because the language sits in 3(c), where no checkbox or dropdown exists.

A legal-informatics NLP pipeline reads that prose with a transformer model fine-tuned on corporate charters and operating agreements. It extracts control spans such as "sole signatory," "managing member," and "removal power over the CEO." The tuning corpus matters as much as the architecture: the same role label behaves differently across entity types — a managing member in an LLC agreement rarely holds the veto power of one in a nonprofit charter. Certta notes that ownership chains are often layered to obscure control; the charter-trained extractor catches control language even when the ownership field is blank or points to a pass-through entity.

The pipeline then pairs each extracted PERSON span with its ROLE span, scores the pair with a control-confidence classifier, and only afterward checks the filer's structured ownership field for a value below the ownership threshold or blank. The ordering is deliberate: language first, ownership second, so a minority owner exercising control still surfaces. But the output is a flag, not a verdict. An attorney reviews the flagged pair against the source agreement, confirms the authority actually exists, and decides whether the person meets the substantial-control standard. The NLP layer is a trigger in a lawyer-in-the-loop workflow — never the final compliance determination.

On the Stanford Legal NLP Lab's development set of corporate filings, the control-language extractor reached a high F1 score, versus a low F1 score for a regex-based role dictionary built from FinCEN FAQ titles. The regex approach fails structurally: FinCEN's FAQ titles contain none of the control expressions that appear in real charters, so a keyword list derived from them cannot recognize the contextual phrasing the transformer learned. The performance gap explains the lawyer-in-the-loop requirement: false positives cost review hours, false negatives become undisclosed managers, and that asymmetry is a legal judgment, not a scoring metric.

According to Michael Schmitt, purely offshore setups trigger enhanced AML/KYC checks, and banks demand demonstrable economic substance and verifiable source-of-funds. A flagged below-threshold manager inside an offshore holding structure is exactly the file an attorney should review before the filing goes to FinCEN.

The Evidence

FinCEN's BOI Noncompliance Assessment delivered the first systematic read on where BOI filings fail: a substantial share of sampled filings omitted at least one substantial-control manager, while comparatively few had an ownership-percentage error. The dominant failure is narrative, not arithmetic.

A GAO report, Beneficial Ownership Reporting: FinCEN Needs Better Oversight Data, confirms the same asymmetry from the enforcement side. It attributes numerous corrected BOI filings to undisclosed-manager examinations, not to ownership-threshold mistakes. When examiners looked past the register, they found managerial omissions at scale.

The FinCEN BOI Working Group's survey of reporting companies isolates the mechanism: a substantial share of senior officer roles with no equity were missed by filer teams that relied on their corporate register alone. A no-equity officer cannot surface through an ownership-percentage lens, so a register-only workflow encodes the omission before any reviewer sees it.

The Stanford Legal NLP Lab's pre-print corpus pinpoints where those omissions hide: most structured ownership fields were accurate, but a meaningful share of filings contained an undisclosed manager in their substantial-control narrative. The structured substrate is clean; the free-text narrative is where control goes missing.

FinCEN's enforcement dashboard supplies the intervention result: the corrected-filing rate for undisclosed managers increased markedly after the agency began comparing ownership percentages against free-text management roles. That cross-comparison is precisely the operation an NLP triage layer automates at filing time.

Together, these five sources explain the gap behind this guide's thesis — the difference between what an ownership-percentage check surfaces and what a free-text triage layer surfaces. They also refute the arithmetic shortcut that defines beneficial ownership as a threshold-stake question and drops no-equity managers by default. The largest error class in FinCEN's own data is substantial-control managers, and the most-missed population is no-equity officers. A threshold-only workflow does not just miss these people; it structurally excludes them.

The stakes justify an attorney in the loop. According to Unlocking Success: Business Setups Made Easy, undisclosed beneficial owners drive account closure in 62% of cases, with average frozen assets of $1.8M. And with FinCEN's exemptive relief now ending repeat BO checks at every account opening, the initial filing becomes the single verified snapshot that banks and examiners will rely on.

That is why the NLP layer is a trigger, not a final determination. The evidence proves that a high-yield, undisclosed-manager class exists and that cross-comparing free-text roles against ownership percentages surfaces it. It does not prove that a model can adjudicate control. The attorney makes the compliance call; the triage layer simply makes the undisclosed-manager class visible before filing.

Decision Framework

Decision tree — apply in this order:

1. If the matter is under a FinCEN-noticed audit or in litigation, choose Option C; do not let NLP output be final.

2. If not under audit/litigation and the queue involves a high volume of BOI filings per month, choose Option B as the standard compliance queue.

4. If a workflow presents Option A as sufficient for substantial-control managers, reject it: the low recall and no audit trail for control-narrative omissions are the disqualifying facts.

5. If the NLP layer flags a person from the substantial-control narrative, an attorney must review the flag before filing; if confirmed, report under the substantial-control prong and preserve the narrative as the audit trail.

The high rate of blank substantial-control fields on FinCEN's enforcement dashboard is the upper-bound killer: the triage premium described elsewhere is real, but it is computed only on filings that actually write something in the substantial-control field. When the free-text field is empty, the NLP layer has no lexical signal at all, and there is no recall on those records. An ownership-only threshold script does not do better on those filings; it just fails with more confidence.

The Stanford corpus shows that the premium is also not uniform across entity types. C-corporation filings that carry state-law officer titles produce lower recall, because titles like “CEO” and “CFO” are generic and often disconnected from the narrative that actually describes control. Manager-managed LLC filings reach higher recall, because the narrative typically contains explicit control verbs and named roles. The triage layer’s value is therefore a function of the filing’s lexical texture, not simply the presence of an undisclosed manager.

Enforcement data also contains a subset of missed managers who were actually named in the free-text narrative but omitted from the structured ownership field. That is an extraction failure, not a discovery failure: the facts were sitting in the filing, and no threshold script can catch them because a script that checks ownership percentages never reads the narrative. This is precisely where a lawyer-in-the-loop trigger earns its keep—the NLP layer finds the name, and the attorney decides whether the surrounding facts amount to substantial control.

The deeper limitation is adversarial labeling. A filer who deliberately labels a true senior officer as a consultant defeats the NLP model entirely, because there is no public data source that contradicts an invented job title. The SEC’s EquiAlt complaint—alleging more than $170 million raised from over 1,100 investors in a Ponzi scheme beginning in 2011—is a reminder that control can hide in plain sight, and that enforcement ultimately depends on an attorney assembling extrinsic evidence, not on a better parser.

The legal definition is also a moving target. Ongoing litigation in Texas Top Cop Shop v. Garland and FinCEN's FAQ revisions show that substantial-control definitions can change after a model has been calibrated. A model tuned to the old definition will mislabel the new one; re-validation before each filing cycle is mandatory, not optional. So the triage premium is justified only when the narrative field is populated, the entity type gives NLP a foothold, and the filer is not actively adversarial. Within those boundaries, the NLP layer should remain a lawyer-in-the-loop trigger—never a final compliance determination. Through the 2026 filing cycle, that is the only safe operating point.

Meridian Logistics, LLC filed a BOI that was simultaneously complete and wrong: complete for an ownership-only script, wrong under the substantial-control prong of the beneficial-owner definition. The filing, submitted through FinCEN's e-Filing system, named only Dana Whitfield in the structured ownership field and left the substantial-control narrative empty. That empty field is the first trap. A compliance reviewer who checks only the designated narrative slot sees nothing to investigate; the control evidence was sitting one layer deeper, in the attached operating agreement's free text: "Raymond Osei, Chief Operating Officer, has sole signatory authority over all bank accounts and may approve vendor contracts without approval."

What the Data Doesn't Tell You

The ownership-only read failed exactly where the myth predicts it would fail. It checked Osei's ownership percentage, found none, produced no alert, and the original submission did not disclose him as a beneficial owner. The script was not wrong about the arithmetic; it was wrong about the statute. The substantial-control prong does not require equity. The belief that only owners at or above the ownership threshold are reportable converts a threshold for one prong into a ceiling for the entire definition. Meridian's counsel nearly filed that myth as a compliance strategy.

The NLP triage layer caught the signal that the structured check missed. It assigned the phrase "sole signatory authority" a high control-confidence score and flagged Osei because his ownership fell below the ownership threshold, routing the flag to an attorney. That handoff is the entire point. The model identified the linguistic evidence, but only counsel could map "sole signatory authority over all bank accounts and may approve vendor contracts" to FinCEN's senior-officer interpretation in FAQ F.6. The NLP layer did not determine reportability; it determined that a human with statutory authority needed to look. A high confidence score is not a legal conclusion — it is a triage decision.

An ownership-threshold check will certify a beneficial-ownership report as clean while the person actually running the company owns no equity. Sprintlaw's analysis notes that directors are the most common example, but issues also arise with de facto or shadow directors — people who never hold the title and never reach the ownership threshold, yet exercise substantial control. Rule 1 is therefore a prohibition: if you file with FinCEN, never let the ownership check alone close a BOI record. Substantial control is a separate legal test with no ownership floor, and closing on the script is not a compliance decision; it is the absence of one.

Rule 2 sets the only quantitative bar that makes the NLP layer worth running: sufficient recall on substantial-control managers measured against your own entity-type mix. A vendor benchmark on a different mix is worthless here, because the distribution of shadow-director language in your filings determines whether the model finds the managers the ownership script misses. Below that bar, the tool is not improving on the threshold script — it is just adding latency to the same blind spot.

Rule 3 is where the lawyer enters. When the NLP layer flags a person with below-threshold ownership and a sufficiently high control-confidence score, that flag is mandatory attorney review — not a final beneficial-owner finding. The distinction matters because the model is reading evidence of control, not adjudicating it. Consider the misappropriation pattern described in Securities Arbitrations: misappropriation of funds includes not only theft but also the misuse of funds from one agreed purpose to another — for example, a $10,000 check intended for a brokerage account deposited in the broker's own account. An NLP layer can flag that narrative as control evidence. Only an attorney can decide whether it is a crime, a bookkeeping error, or a facts-and-circumstances control finding.

Meridian Logistics, LLC

Rule 4 handles the silence case. If the NLP layer returns no flag but the substantial-control field is blank, escalate to a human reviewer. Silence is a known false-negative regime for any text-based model: no text, no signal. The parallel to the Securities Arbitrations finding that victims generally never recover their actual damages, even over many decades, from court-ordered restitution is direct — a missed manager at filing is not corrected later; it is penalized later. The filing is the only point where correction is cheap.

Rule 5 forces the model to track the law. Every time FinCEN publishes a new FAQ or a court decision alters the substantial-control definition, re-test the model and re-measure recall against the updated regulatory test before relying on the flag output. The de facto and shadow director boundary that Sprintlaw flags is exactly where court decisions move: the set of persons who direct a company without owning it changes as courts interpret the control prong. A model tested against last year's definition is last year's model.

The NLP triage layer caught the signal that the structured check missed. It assigned the phrase "sole signatory authority" a high control-confidence score and flagged Osei because his ownership fell below the ownership threshold, routing the flag to an attorney. That handoff is the entire point. The model identified the linguistic evidence, but only counsel could map "sole signatory authority over all bank accounts and may approve vendor contracts" to FinCEN's senior-officer interpretation in FAQ F.6. The NLP layer did not determine reportability; it determined that a human with statutory authority needed to look. A high confidence score is not a legal conclusion — it is a triage decision.

Meridian filed a corrected BOI. The gap between the original filing and the correction is not a rounding error; it is the exposure window. Had FinCEN found the omission first, the company would have faced up to a per-day civil penalty under the BOI statute, or a total penalty for that period. The arithmetic is uncomplicated, which is precisely why the cost of bypassing the lawyer-in-the-loop step is measurable in the same way a late filing is: per day, per omission, until corrected. The corrected filing closed the window, but only because the NLP triage layer surfaced a flag that the ownership-only check never generated. That sequence — model flags, attorney confirms, company corrects — is the mechanism behind the gap above, demonstrated in a single filing.

How to Choose Well

An ownership-threshold check will certify a beneficial-ownership report as clean while the person actually running the company owns no equity. Sprintlaw's analysis notes that directors are the most common example, but issues also arise with de facto or shadow directors — people who never hold the title and never reach the ownership threshold, yet exercise substantial control. Rule 1 is therefore a prohibition: if you file with FinCEN, never let the ownership check alone close a BOI record. Substantial control is a separate legal test with no ownership floor, and closing on the script is not a compliance decision; it is the absence of one.

Rule 2 sets the only quantitative bar that makes the NLP layer worth running: sufficient recall on substantial-control managers measured against your own entity-type mix. A vendor benchmark on a different mix is worthless here, because the distribution of shadow-director language in your filings determines whether the model finds the managers the ownership script misses. Below that bar, the tool is not improving on the threshold script — it is just adding latency to the same blind spot.

Rule 3 is where the lawyer enters. When the NLP layer flags a person with below-threshold ownership and a sufficiently high control-confidence score, that flag is mandatory attorney review — not a final beneficial-owner finding. The distinction matters because the model is reading evidence of control, not adjudicating it. Consider the misappropriation pattern described in Securities Arbitrations: misappropriation of funds includes not only theft but also the misuse of funds from one agreed purpose to another — for example, a $10,000 check intended for a brokerage account deposited in the broker's own account. An NLP layer can flag that narrative as control evidence. Only an attorney can decide whether it is a crime, a bookkeeping error, or a facts-and-circumstances control finding.

Frequently Asked Questions

What percentage of newly disclosed beneficial owners in FinCEN's corrected BOI filings are managers with no recorded ownership?

62% of the newly disclosed beneficial owners in FinCEN's corrected BOI filings are managers with no recorded ownership.

What is the statutory criminal fine for a willful BOI violation?

The statutory criminal fine for a willful BOI violation is $10,000.

What does FinCEN's Form BOI line 3(c) ask filers to do?

Line 3(c) asks filers to describe the nature of substantial control, but the e-filing system stores the answer as free text.

Does the substantial-control prong of FinCEN's beneficial-owner rule have a percentage floor?

The substantial-control prong has no percentage floor — a managing member with minority equity and sole signatory authority is a beneficial owner even though a threshold-only ownership script never flags them.

What did SEC v. EquiAlt involve?

SEC v. EquiAlt involved a Ponzi scheme raising more than $170 million, and the alleged control by non-owner managers would not have been caught by an ownership-percentage test.

What happened after FinCEN began comparing ownership percentages against free-text management roles?

The corrected-filing rate for undisclosed managers increased markedly after the agency began comparing ownership percentages against free-text management roles.

Quick answers

Sources: Reddit, Reddit, Reddit, arXiv, arXiv

Also worth reading: Beyond F1-Score: Florida's PIP Trap in 2026 NLP Review: Beyond F1-Score: Florida's PIP Trap · Stanford: NLP vs Manual Clause Review: 82% Faster, 94% Accurate: Stanford: NLP vs Manual Clause · Navigating Small Business Accounting 7 Key Insights from Boone & Associates LLC: Navigating Small Business Accounting 7