Compliance · KYB / KYC
The AI Perception Layer for KYB and KYC Compliance
Registries prove a company exists on paper. Large language models show how it actually lives in the market.
Know Your Business (KYB) and Know Your Customer (KYC) processes are built on a clear foundation: legal existence, beneficial ownership, documentary consistency, and sanctions or adverse-media screening. These checks answer essential questions — does this entity exist, who controls it, and does the paperwork align?
They leave a second question unanswered. That question is already being asked inside the tools many analysts and reviewers now use every day: how do large language models currently describe, recommend, rank, or omit this company?
The answer is generated in private sessions, usually with a single model, and leaves almost no auditable trail. Different models frequently return different answers about the same counterparty. One model may treat the firm as a credible category participant; another may omit it from the same category query entirely; a third may surface risk language that never appears in the data room or registry extract. We call this structured disagreement split-perception. It is currently invisible in most onboarding workflows and therefore unpriced.
VeritasLinks makes that perception layer measurable, decomposable, and reviewable so it can sit alongside traditional KYB and KYC evidence rather than remaining an unrecorded influence on first impressions.
Live report surface
Perception affect field · split detected
Northline Analytics · illustrative fixture on the real UI. Not a screenshot.
Perception affect field
Models split into two emotional camps
Polarization
64%
Division is material, not noise.
Dominant zone
Activated
Why traditional KYB leaves a blind spot
What registry and document checks cannot see
A company can clear every registry query, ownership verification, and document review and still present a fractured or inflated picture inside generative models. The reasons are structural.
Large language models compress public digital footprint — websites, reviews, press, forums, structured data, and third-party discussion — into short verbal judgments. Those judgments are delivered at the precise moment a reviewer forms a first impression. Because the models differ in training data, retrieval behavior, and internal ranking of authority signals, they do not converge on a single “fact of the matter” about a company’s standing. The same firm can be described as a defensible platform by one model and as a commodity tool (or omitted) by another.
In the measured data underlying the Split-Perception working paper, 6.4% of companies covered by at least two models on category queries were simultaneously omitted entirely by one model and ranked in the top three by another. For those firms, whether they appear in an AI-assisted short list or research note is decided by which assistant the reviewer happens to open. That is not random noise; it is a property of the model–evidence pair.
Traditional KYB systems have no mechanism to surface this divergence. They were never designed to. The result is that a material information risk — the model-dependence of the first AI judgment — remains outside the compliance record.
What the perception layer actually measures
Signals relevant to onboarding and ongoing monitoring
Each VeritasLinks assessment executes a structured battery of 1,000–1,500 prompts across five production models (ChatGPT, Claude, Grok, Perplexity, DeepSeek) under a fixed orchestration layer. Responses are parsed into measurable signals and aggregated into a composite VeritasScore on a 300–870 scale. The scale is deliberately modeled on consumer credit scoring: bounded, monotonic, and decomposable into named components whose contributions are disclosed in every report.
For KYB and KYC teams the following outputs are particularly relevant:
Cross-model score distribution
The range between the highest- and lowest-scoring models on the same measurement date. In the fully instrumented subset the median range is 100 points and the 90th percentile reaches 134 points on a scale whose reporting bands are 200 points wide. A wide split is itself a finding: the AI-assisted view of the counterparty is highly model-dependent.
Mention rate, position, and omission
Whether the company appears in buyer-shaped, category, and diligence queries, how prominently it is named, and whether any model omits it while others rank it highly. Simultaneous omission and top-three ranking is a high-severity signal.
Polarization and affect field
Responses are mapped onto a valence–arousal field. The polarization index quantifies whether models cluster into emotionally different camps. High polarization means that averaging model outputs hides reputational and recommendation risk.
Coded objections from synthetic panels
Models conditioned on distinct evaluator roles are asked to discuss positioning, make forced preference choices, and articulate risks. Recurring objections (platform dependency, competitive defensibility, unit economics, and others) are coded and retained with full transcripts for auditability.
Citation and authority profile
The evidence base models actually draw upon. Thin or conflicting authority signals concentrate exactly where cross-model divergence is largest.
Risk deduction component
An explicit subtractive term for objections, negative attributions, and instability surfaced anywhere in the assessment.
Together these outputs convert an invisible single-model draw into a reviewable distribution with clear severity markers.
Live report surface
VeritasScore · multi-model distribution
Northline Analytics · illustrative fixture on the real UI. Not a screenshot.
Score by LLM model
620Overall · 300–870
How the layer fits existing KYB/KYC workflow
Complementary evidence, not a replacement
VeritasLinks does not replace registry checks, beneficial-ownership verification, document review, or sanctions screening. Those remain primary. The perception layer answers a different question and is designed to be recorded alongside the traditional file.
A practical sequence used by teams that already incorporate AI-assisted research:
- Complete the standard KYB or KYC process and retain the usual evidence.
- Run a VeritasLinks assessment on the same legal entity or trading name.
- Record the overall VeritasScore, the cross-model range, any material omissions, the polarization index, and high-severity coded objections.
- Treat a wide split, consistent omission by one or more models, or elevated risk deduction as a diligence note that requires either an internal explanation or further human review of the public footprint.
- Retain the timestamped report so the perception state at the time of onboarding can be reconstructed if questions arise later.
Because the methodology is versioned and the orchestration layer is fixed, assessments under the same vintage remain comparable. The score is designed so that absence of new public signals produces relative stability, while real signals (new third-party discussion, authoritative coverage, or material news) appear in subsequent runs.
For higher-risk or higher-value relationships, periodic re-assessment creates a light time series of perception state that can be reviewed alongside other ongoing monitoring.
What “inflated, split, or phantom” looks like in practice
Three patterns that matter for onboarding risk
Inflated perception
Models describe the company in language that significantly exceeds the strength of its observable public footprint or third-party corroboration. High mention rate or positive sentiment paired with weak citation quality and low authority signals is a common signature. The risk is that internal or counterparty AI research creates expectations the company cannot operationally support.
Split perception
Material disagreement between models on positioning, category membership, or risk language. The same firm is treated as a leader by one model and marginal or absent by another. Pipeline or short-list membership becomes tooling-dependent. The spread itself is the finding.
Phantom perception
The company is largely invisible across models despite having a legal existence and some public presence. Omission from category and buyer-shaped queries means that any AI-assisted research conducted by partners, customers, or internal teams will simply not surface the name. That invisibility is rarely recorded in the KYB file.
Each pattern is detectable from the combination of VeritasScore components, per-model distribution, mention/omission data, and authority profile.
Live report surface
Perception affect field · split detected
Northline Analytics · illustrative fixture on the real UI. Not a screenshot.
Perception affect field
Models split into two emotional camps
Polarization
64%
Division is material, not noise.
Dominant zone
Activated
Research foundation
Built on measured divergence
The methodology and quantitative findings are documented in the working paper “Split-Perception: Measuring Divergence in How Large Language Models Assess Company Credibility” (July 2026). The paper draws on assessments covering more than 5,000 companies and several million individual model responses. Quantitative claims rest on the fully scored current-vintage population of 139 companies, with per-model persistence and longitudinal series providing the divergence and stability statistics cited above.
The paper explicitly frames split-perception as an unpriced information risk for KYB review, credit assessment, and M&A due diligence workflows that rely, formally or informally, on AI-assisted research. It distinguishes stable model–evidence disagreement from volatility imported by live retrieval, and shows that divergence concentrates where authority signals are weak or conflicting.
Make AI perception visible before onboarding decisions are finalized
If reviewers or internal tools already consult large language models for first-pass research on counterparties, the perception state is already influencing attention. The only operational question is whether that state is measured, decomposed, and retained in the compliance record or left as an invisible single-model draw.
Free analysis
Run it on any company
Public URL. Same funnel as the homepage. No card required.
Enter a public website
Multi-model · VeritasScore · ~5 minutes
See a sample dossier →FAQ
Short clarifications — positioning and proof live in the sections above.
Does VeritasLinks replace official KYB or KYC providers?+
No. It is an independent perception layer that sits alongside registry, ownership, and documentary checks. It answers a different question.
How should a wide cross-model split be treated in an onboarding decision?+
As a diligence note. A wide split indicates that the AI-assisted view of the counterparty is highly model-dependent and usually points to a thin or contested public evidence base. It warrants either explanation or additional human review of the footprint.
Can the assessment be run on private, pre-revenue, or thinly documented companies?+
Yes. The battery works on any public digital footprint. Thin footprints typically produce wider splits and higher retrieval volatility. That outcome is itself informative for onboarding risk.
How often should a counterparty be re-assessed?+
After material public events (funding, product launches, regulatory actions, major press) or on a periodic cadence for higher-risk relationships. The score is designed to remain relatively stable when no new signals appear and to reflect real signals when they do.
Is the output auditable?+
Yes. Full session transcripts (including adaptive probes), coded objections, component contributions, and per-model scores are retained. The methodology is versioned so that later reviewers can understand the vintage under which the assessment was produced.
What is the relationship between mention rate and actual recommendation?+
High mention rate does not automatically translate into selection or recommendation in key queries. The system explicitly separates visibility signals from positioning and preference signals. A company can be frequently mentioned yet rarely chosen when models are placed in an evaluative role.
Does this create additional regulatory burden?+
The layer is designed to be recorded as supplementary evidence, not as a new primary control. Teams that already use AI research informally gain a way to make that usage visible and consistent rather than leaving it unrecorded.