How Is Illegal Activity Organized?

Henrik

2025-10-31

How do criminals trust each other?

And why does it matter for policy?

Opening question — reframed for a tax audience. The paper is about corruption, but the trust framework generalizes. Tax evasion faces the same fundamental problem: agreements can’t be enforced in court. But the way tax evasion solves this problem is distinctive — and that distinction has policy implications.

The core insight

Every illegal transaction has the same problem:

Agreements cannot be enforced in court.

A drug supplier who is cheated cannot sue.

A bribed official who doesn’t deliver cannot be held liable.

A wealth manager who absconds with hidden assets cannot be reported.

So how do they solve this? → The organizational form is shaped by which trust technologies are available.

Gambetta (1993) for the Mafia · Lambsdorff (2002, 2007) for corruption · Dixit (2004) for lawless exchange · Harrington (2016) for tax evasion

Lambsdorff (2002, 2007) identified this for corruption. Gambetta (1993) for the Mafia. We generalize: the trust problem is universal to all illegal exchange. But different crime types solve it differently — and the solution determines the organizational form.

For this audience: the tax evasion version is distinctively vertical. A taxpayer hiding assets offshore must trust their enabler — wealth manager, corporate service provider, banker — with the assets themselves. If the enabler disappears, so does the money. Harrington (2016) documents this beautifully.

Two cases, two solutions

Odebrecht (Brazil, 2001–2016)

Latin America’s largest construction firm paid $788M in bribes across 12 countries.

Created a formal corporate department — the Structured Operations Division — with encrypted IT systems, codenames for every politician, and a CEO-maintained spreadsheet tracking who was owed what.

UBS (Switzerland, 2000–2007)

Switzerland’s largest bank helped ~20,000 US clients hide $20 billion from the IRS.

Bankers traveled to the US under false pretenses, trained in counter-surveillance, carried encrypted laptops, and smuggled diamonds in toothpaste tubes.

Same trust problem. Opposite solutions.

These two cases will recur throughout the talk. They represent the two poles of how illegal activity solves the trust problem.

Odebrecht (corruption): a Brazilian multinational that bribed officials in 12 countries over 15 years. The trust infrastructure was relational + hierarchical: a formal corporate department (the Structured Operations Division, or SOD) managed all illicit payments. The CEO, Marcelo Odebrecht, personally maintained a spreadsheet (“the Italiano”) tracking payments owed to politicians. Communication via Drousys, a proprietary encrypted messaging system. Codenames for every politician. A network of doleiros (money changers) for cash conversion, each compartmentalized. The key: trust was built through years of repeated dealing and enforced through mutual exposure — everyone had dirt on everyone.

UBS (tax evasion): Switzerland’s largest bank ran a cross-border wealth management operation serving ~20,000 US clients with $20 billion in undeclared assets. The trust infrastructure was institutional + contractual: Swiss banking secrecy law (Article 47, criminal penalties for disclosure) served as a state-enforced commitment device. Numbered accounts with physically separated information. Wealth management agreements and nominee structures. The bank’s 150-year brand as a trust anchor. Bankers built personal relationships through luxury events (Art Basel, yacht races, $10K Elton John dinners), but the core trust technology was the legal and institutional framework, not the personal bond.

The contrast: Odebrecht needed codenames, encrypted systems, and a secret ledger because corruption has no legal infrastructure to lean on. UBS could rely on Swiss law, formal contracts, and institutional reputation because the service itself (wealth management) is legitimate — only the underlying purpose (tax evasion) is illegal.

Three sources of trust

All trust in illegal exchange comes from three sources of cost for defection:

Primitive	If I defect…	Dominates in…
Relationship	…I lose a valuable ongoing interaction	Drug trafficking (84%), corruption (65%)
Community standing	…third parties learn and I lose access	Tax evasion (89%), organized crime (29%)
Formal contract	…counterparty invokes the legal system	Corruption (19%), tax evasion (19%)

Building on: Kandori (1992), Gambetta (1993, 2009), Lambsdorff (2002), Dixit (2004)

Plus modifiers (kompromat intensifies relationships), architecture (intermediaries chain trust links), and configurations (hierarchy = relationship + contract + power asymmetry)

This is the deeper theoretical framework that emerges from comparing across crime types. We code six empirical categories (relational trust, reputation, kompromat, intermediary, hierarchy/coercion, contractual formalization) — but these decompose into three primitives.

Relationship: the folk theorem. The shadow of future dealings deters cheating. This is Lambsdorff (2002) for corruption, Reuter (1983) for drug markets. Subtypes: repeated interaction, kinship, friendship, ethnic ties. Near-universal — 84% in drug trafficking, 65% in corruption, 63% in organized crime. Odebrecht: 15 years of repeated dealing with the same politicians. UBS: quarterly personal visits, yacht trips, concierge services — but relationship was secondary to institutional trust.

Community standing: third parties observe and impose costs. Two modes — standing for reliability (Kandori 1992: you lose future business) and standing for punishment (Gambetta 1993: you are known to retaliate). These are different mechanisms sharing a label. Tax evasion’s 89% reputation is for reliability — enablers advertising “professional discretion” through legal channels. Mafia’s 29% is for punishment — violent retaliation within a closed community.

Formal contract: the legal system imposes costs. Contractual formalization — consultancy agreements, trust deeds, nominee arrangements — is available only to “insiders” with legal-system access. 19% in corruption and tax evasion, 1% in drug trafficking. You can’t write an enforceable contract for cocaine delivery.

Important nuance: the formal contract primitive can work in two directions. In corruption, contracts enforce performance (the consultancy agreement obliges the intermediary to deliver). In tax evasion, the legal system also enforces concealment — Swiss banking secrecy law (Art. 47) made it a criminal offense for bankers to disclose client information. The state was the trust guarantor for the illegal arrangement. UBS’s 2002 letter to clients — “the protection of client data can not possibly be compromised” — was not a promise from the bank but a statement about the legal framework. This “inverted contract” (law enforcing secrecy rather than performance) is why tax evasion’s trust infrastructure was so robust: it was backed by sovereign enforcement, not just bilateral agreement.

The remaining three empirical categories are not independent primitives: - Kompromat is a relationship modifier — mutual vulnerability raises defection cost within an existing relationship. Doesn’t work between strangers. - Intermediary is a trust architecture — restructures one impossible trust relationship into a chain of manageable ones, each sustained by a primitive. - Hierarchy/coercion is a configuration — organizational authority = relationship + contract + power asymmetry; violence = punishment capacity + community standing for punishment.

The framework is exhaustive because relationship, community, and state are the only three institutions that can impose costs on a defecting party. Every trust mechanism draws on one or more of them. The mix differs across crime types because social position determines which primitives are accessible.

Which primitives are available?

	Relationship	Community standing	Contract
Drug trafficking	✓✓ kinship, repeated	punishment	✗
Organized crime	✓✓ clan-based	punishment	✗
Corruption	✓ repeated deals	✗ hidden	✓ insider
Money laundering	✓ professional	reliability	✓
Tax evasion	weak	reliability	✓

Social position → which primitives are accessible → organizational form

The three-primitives framework predicts organizational form from the social position of participants.

Drug trafficking: relationship is the dominant primitive (84% relational trust). Participants are embedded in long-term supply chains, often family-based. Community standing operates in the “punishment” mode — cartels build reputation for violence. Formal contracts are unavailable — you can’t sue your cocaine supplier.

Organized crime: similar to drug trafficking but with stronger community structures (Mafia families, clans) that support reputation for punishment. Hierarchy is really relationship + authority under power asymmetry.

Corruption: relationship works (repeated interaction between official and contractor). But community standing fails — corrupt transactions are bilateral and hidden, officials can’t advertise reliability or punishment. Formal contracts are available because participants are insiders with legal-system access.

Tax evasion: the most distinctive profile. Relationship is weak — clients hire wealth managers professionally, not personally (42% relational, the lowest). Community standing operates in the “reliability” mode — enablers compete on track records, industry rankings, referral networks. Formal contracts are the core architecture — trust deeds, nominee arrangements, service agreements.

The key insight: drug trafficking and tax evasion are almost perfectly complementary in which primitives they access. Drug trafficking: relationship + punishment, no contracts. Tax evasion: contracts + reliability reputation, weak relationships. Same trust problem, opposite solutions, because participants occupy opposite social positions.

The trust framework

The Lambsdorff trust framework, generalized. Four risks: performance, denunciation, hold-up, search. Each trust mechanism addresses different subsets. For tax evasion, the dominant risks are performance (will the enabler return the money?) and denunciation (will the enabler cooperate with authorities?). Hold-up is less acute because the enabler typically has more to lose from exposure than the client.

Data: 6,486 cases across crime types

16 sources · 173 countries · 6 crime types · LLM-extracted trust mechanisms

The full dataset: 5,950 cases spanning corruption (3,458), drug trafficking, organized crime, human trafficking, and tax evasion (208 from ICIJ). Sources include FCPA enforcement actions, TRACE Compendium, World Bank investigations, DOJ Public Integrity, DEA drug enforcement, UNODC SHERLOC, OCCRP/ICIJ investigations, GDELT news, UK SFO, Brazilian courts, 26 books, and investigative journalism. For each case, GPT-4.1-mini extracts structured information including free-text descriptions of how trust was maintained, which are then classified into our six mechanisms.

Example: Odebrecht → structured extraction

Delação Fernando Migliaccio, Termo 05 (Lava Jato plea bargain)

[…] The Structured Operations Division maintained offshore accounts in Portugal, Switzerland, Austria, Antigua […] payment orders via phone, email, and the DROUSYS encrypted system […] each collaborator controlled distinct account sets […] fabricated contracts to justify repatriation of funds […] doleiros and lawyers as compartmentalized intermediaries.

Field	Extracted
Trust mechanisms	Relational ✓ Intermediary ✓ Contractual ✓ Kompromat ✓ Hierarchy ✓
Key insight	Corruption bureaucratized — a formal department with IT systems, codenames, and gatekeepers

How the pipeline works. A plea bargain transcript — first-person testimony from the SOD operator Fernando Migliaccio — goes through two-stage LLM extraction. Stage 1: extract free-text trust descriptions. Stage 2: classify into the six mechanisms.

This single case hits five of six mechanisms: relational (years of repeated dealing), intermediary (doleiros compartmentalized by region — Recife, Porto Alegre, Salvador, Uruguay — each with codenames), contractual (fabricated consultancy contracts), kompromat (the Italiano spreadsheet = mutual exposure ledger), and hierarchy (Marcelo Odebrecht as CEO personally approved payments through gatekeeper “Bira”).

The Odebrecht case is unusual in hitting so many mechanisms simultaneously. Most corruption cases rely on 2-3. The SOD’s sophistication — encrypted communications, codenames, periodic data destruction, offshore layering across 6+ jurisdictions — reflects the scale of the operation ($788M in bribes, 12 countries, 15 years).

Stage 1: What we ask the LLM to extract

System prompt (excerpt):

Corrupt deals cannot be enforced in court. This creates trust problems: will the bribed official deliver? Will the firm actually pay the kickback? Will either party inform on the other? Will one side use evidence to extort the other? How did the parties even find each other safely? We want to understand how these problems were solved in each case.

Extraction fields (excerpt):

trust_problems: “free-form: who needed to trust whom, and why was trust difficult? E.g. ‘the firm needed to trust that the official would steer the contract after receiving payment; the official needed assurance the firm would not inform prosecutors.’”

trust_solutions: “free-form: how were the trust problems actually solved? What relationships, arrangements, threats, intermediaries, recording of evidence, repeated dealings, family ties, or other mechanisms enabled the corrupt parties to trust each other?”

Free text first → classify second. Avoids catch-all categories.

This is the actual prompt we use (shown here for the FCPA extraction script, but the trust fields are identical across all 13 sources). Two key design choices:

First, the system prompt frames the task through the trust problem — it’s not “extract corruption facts” but “understand how the deal was enforced between the parties.” This primes the model to look for enforcement mechanisms rather than just listing actors and amounts.

Second, the trust fields are FREE TEXT, not enums. An earlier version asked the model to directly classify into mechanism categories, but 79% of cases ended up as “mutual complicity” — a catch-all. By asking “how were the trust problems solved?” as an open question, we get rich descriptions that a second pass can classify more discriminatingly.

The full extraction schema has ~15 fields (summary, country, sector, actors, mechanism, scale, duration, detection, confidence, etc.) but the trust_problems and trust_solutions fields are the key innovation.

Stage 2: Classifying trust mechanisms

Mechanism	Prompt definition (condensed)
Relational	Ongoing relationship whose value deters cheating. Subtypes: repeated interaction, kinship, friendship, ethnic/clan ties
Reputation	Community-level enforcement: third parties observe and gossip, so cheating is punished by the broader community, not just the dyad
Kompromat	Deliberate creation of mutual vulnerability. Compromising evidence, mutual blackmail potential, down payments as commitment devices
Intermediary	Third-party brokers, fixers, go-betweens who connect, guarantee, or enforce. They solve the search problem and can provide performance guarantees
Hierarchy	One party can directly punish the other through organizational authority, violence, or threats. Compliance is compelled
Contractual	Parties use formally legal contracts (consultancy agreements, agent commissions, subcontracts, contingency fees) to enforce informally illegal arrangements. The contract is real and legally enforceable — what is corrupt is the underlying purpose

Input: free-text trust_solutions from Stage 1 → Output: 6 binary flags per case

The second-stage prompt gives the LLM precise definitions for each mechanism and asks it to classify the free-text trust_solutions output from Stage 1 into binary flags.

Key features of the codebook: - Contractual formalization gets the longest definition because it’s the new mechanism we document. We specify: “the contract is real and legally enforceable — what is corrupt is the underlying purpose.” This distinguishes it from merely that money changed hands. - Reputation has the highest bar: requires community-level observation and punishment, not just that someone is “trusted” or “has a good name.” This is the most commonly misclassified mechanism. - Intermediary requires a distinct third party who connects, screens, or guarantees — not just any third party mentioned in the text. A bank processing a wire is infrastructure, not an intermediary. - We also code four enabling conditions separately (political protection, institutional capture, professional enablers, weak enforcement) — these explain WHY corruption is possible, not HOW parties trust each other. Important distinction.

The full codebook with borderline cases, common misclassifications, and examples is documented in docs/trust_mechanism_codebook.md (8 pages). Available in the replication package.

Finding 1: Trust profiles differ sharply by crime type

Each crime type has a distinctive trust profile. Corruption relies on relational trust (65%) and intermediaries (52%). Drug trafficking is dominated by relational trust (84%) backed by violence. Tax evasion stands out with the highest reputation rate (89%) and intermediary use (60%), but almost no hierarchy. These profiles are predictable from the social position of participants.

Finding 2: Distance forces intermediaries

57% vs 26%

intermediary prevalence: cross-border vs domestic

When transactions cross borders, intermediary use doubles. Domestic deals can rely on pre-existing relationships; cross-border deals involve strangers in different legal systems who need brokers. This holds across crime types — but is most pronounced in corruption and tax evasion, where intermediaries are also the most professionalized.

Odebrecht illustrates this perfectly: in Brazil, they could deal directly with politicians they’d known for years. But when they expanded to Angola, Venezuela, Peru — 12 countries total — they needed local intermediaries (doleiros, fixers) in each country, each hired through fabricated consultancy contracts. UBS solved the same problem from the other direction: Swiss bankers couldn’t cold-call American billionaires, so they sponsored Art Basel and yacht races as search mechanisms, and used the New York referral desk as an intermediary layer.

Contracts make intermediaries hireable

The problem:

Building trust takes years.

How do you transact illegally with someone you’ve never met?

The solution:

Hire a broker via a legal contract.

Consultancy agreement, agent commission, wealth management contract

66% of intermediary cases involve formal commercial arrangements

Intermediary × contractual co-occurrence: \(\chi^2\) = 59, p < 10^-14 · cf. Drugov, Hamman & Serra (2014); Harrington (2016)

This is the key mechanism. Contractual formalization lets you HIRE trust instead of building it.

Odebrecht: when expanding to Angola, they couldn’t rely on existing relationships. So they hired local intermediaries through fabricated consultancy contracts — legally enforceable agreements whose ostensible purpose was legitimate consulting. The intermediary’s commission (typically 10-20%) was the price of instant trust.

UBS: the Qualified Intermediary agreement with the IRS was supposed to require client reporting, but UBS exploited a loophole — accounts held by nominee entities weren’t reportable. They referred clients to external lawyers and accountants who set up BVI/Panama/Liechtenstein shells. Each referral was a contractual link in a trust chain.

Drug traffickers can’t do this — they need years of relationship-building with every new supplier. Harrington (2016) documents the tax evasion version in ethnographic detail.

Trust profiles across crime types

N = 6,204 classified cases across 5 crime types

Companion paper results. The trust profiles are strikingly different across crime types — exactly as predicted.

Drug trafficking: relational (84%) and coercion. Corruption: intermediaries (52%) and contractual (19%). Organized crime: hierarchy (68%) and reputation (29%). Human trafficking: pure coercion (87%).

Tax evasion breaks the pattern open. The 89% reputation rate is the highest of any crime type — but it’s a completely different KIND of reputation than Mafia reputation (29%). Offshore enablers advertise “professional discretion,” “reliable asset management,” “client confidentiality” — perfectly legal signals. Mossack Fonseca, Appleby, the Spear’s 500 wealth manager rankings. This is reputation for RELIABILITY transmitted through legal professional channels. Mafia reputation is for PUNISHMENT transmitted through violence within a closed criminal community. Same label in our taxonomy, fundamentally different mechanism. We discuss this distinction in Section 5.4 of the companion paper.

Other tax evasion results: intermediaries at 60% (multi-layered wealth chains), contractual formalization at 19% (trust deeds, nominee arrangements), hierarchy at 1% (no violence). Exactly the enabler-dependent profile predicted by the framework.

Contractual formalization: 19% in both corruption and tax evasion, 1% in drug trafficking. This is the insider-crime mechanism — available only where participants can access the legal system.

Why violence works there but not here

The paradox: reputation for violence requires attribution — but attribution is incriminating. How do criminals build it?

Mechanism	How attribution stays out of court
Organizational branding	Narco-messages credit the cartel, not the individual operative
Witness deterrence	The reputation for violence prevents witnesses from testifying
Initiation rituals	Violence witnessed only by fellow gang members (Skarbek 2011)
Pattern signatures	Killing methods readable by community, not legal evidence

Corruption can’t do any of this. Officials have legitimate positions where any attribution — community or legal — is catastrophic. No brand to absorb it, no violence to deter witnesses.

This slide explains why the trust mechanism profiles differ so sharply across crime types. The puzzle: reputation for violence requires people to know you did it. Legal conviction also requires people to know you did it. So how does it work?

The answer is that community attribution and legal proof operate on different evidentiary standards. Four mechanisms maintain this separation:

1. Organizational attribution: Mexican cartels leave narco-messages on bodies — banners claiming credit (Rios and Rivera 2018). The cartel brand gets credit, building the organization’s reputation, but the individual killer remains anonymous. Reputation accrues to the brand, not the operative.

2. Witness deterrence: the reputation for violence is self-reinforcing as a legal shield. Everyone in a neighborhood knows who controls the territory. But testifying carries lethal risk. “Plata o plomo” ensures community knowledge never becomes courtroom evidence.

3. Costly signaling through initiation: Skarbek (2011) documents prison gang initiations requiring violence — witnessed only by fellow gang members who will not testify. Gambetta (2009) makes the key point: an undercover agent would never commit murder to maintain cover, so murder is a “perfectly discriminating signal” of authentic criminal identity.

4. Pattern attribution: specific killing methods, locations, timing create “signatures” readable by the criminal community (Gambetta 2009 — lupara bianca, specific mutilations) but not admissible as evidence of individual responsibility.

Corruption fails on all four. Officials have legitimate social positions where any attribution is catastrophic. There’s no cartel brand to absorb the reputation. No violence to deter witnesses. No closed community to observe and validate. And the “service” being offered (reliable delivery on bribes) cannot be signaled without self-incrimination — unlike the offshore enabler market, where “professional discretion” is a perfectly legal advertisement.

This is the deepest reason why corruption and tax evasion must rely on contractual formalization and intermediaries rather than reputation and violence: their participants live in the legitimate world.

What do intermediaries actually do?

Hypothesis: intermediaries hold funds until the official delivers (escrow).

Reality (772 FCPA cases): firm pays first in 93% of intermediary cases.

Intermediary role	% of cases	Function
Conduit	40%	Pass money through, create paper trail
Buffer	17%	Distance + deniability for the firm
Relational guarantor	7%	Use relationship with official to ensure delivery

Not escrow — the intermediary provides access + concealment + relational guarantee.

Contrast: in tax evasion, the enabler IS a true escrow agent (holds the assets).

This corrects an intuitive but wrong hypothesis. We initially expected intermediaries to function as escrow agents — holding funds and releasing on performance. The data shows otherwise.

In 93% of intermediary cases, the firm pays before the favor. The intermediary forwards the payment and then uses their position to encourage the official to deliver. The intermediary’s value is a bundle of three services: access (knows the right official), concealment (creates legitimate-looking paper trail via consulting agreements), and relational guarantee (ongoing relationship with the official provides soft performance assurance).

The contrast with tax evasion is clean and important. In corruption, the “asset” being transacted — a policy decision, a contract award — is not holdable. You can’t hold a contract award in escrow. So the intermediary cannot serve as a true escrow agent. In tax evasion, the asset IS holdable — it is money, securities, property. The wealth manager literally holds the client’s money under a trust deed. This structural difference — holdable vs non-holdable assets — may explain why tax evasion can be organized through impersonal professional markets while corruption requires personal relationships.

The intermediary’s commission (1-5% on top of the bribe) is the price of this access + concealment + relational guarantee bundle, not an escrow fee.

The price of trust

If intermediaries solve the trust problem for a fee, the commission should reflect the cost of providing trust.

Setting	Median commission	Why
Iraq Oil-for-Food	10%, zero variance	Standardized tariff, no trust needed
Mexico (CFE, repeated)	5%	Repeated dealings compress premium
Extractives (long-term)	3%	Concessions over decades
Gabon (one-shot, cross-border)	20%	Intermediary must provide access + concealment + relational guarantee
Defense (one-shot, high-value)	20%	Trust must be purchased, not accumulated

With repeated interaction: mean 10%. Without: mean 16.8% (p = 0.0002)

→ Relational trust and intermediary services are substitutes with observable prices

This is the “price of trust” analysis from 104 FCPA cases with explicit commission percentages.

The key finding: commissions vary systematically with how costly the trust problem is to solve. Where trust is pre-established through repeated dealings, the intermediary premium drops to 10% or below. Where trust must be purchased from scratch (one-shot deal, cross-border, no prior relationship), commissions reach 20%.

Iraq Oil-for-Food is the limiting case: the government set a standardized 10% surcharge with zero variance. No bilateral trust was needed because the arrangement was a publicly known rule enforced through monopsony power. The “commission” is not a trust premium — it’s a tax on doing business.

Gabon defense cases: the intermediary takes 20% — double the Iraq rate — because the intermediary must provide access, concealment, and relational guarantee in a one-shot cross-border deal.

The relational trust finding (p=0.0002) is the cleanest result: when parties already trust each other through repeated interaction, the intermediary premium falls by 6.8 percentage points. This confirms that relational trust and intermediary services are substitutes — you can build trust through years of dealing (cheap, slow) or buy it through an intermediary (expensive, immediate).

For the tax audience: wealth management fees (0.5-2% AUM annually) are the tax evasion equivalent. Lower than corruption commissions because the contractual protection is stronger (the service agreement is legitimate) and the enabler’s professional reputation provides cheaper trust than a corrupt intermediary’s personal relationship.

A sequencing test: does reputation work in corruption?

If reputation works, firms with track records of paying bribes should be trusted to pay after receiving the contract. Officials should deliver first.

What we find (772 FCPA cases):

• Payer moves first in 90% of cases
• Official moves first in only 2% (16 cases)
• In those 16: zero mention reputation as risk mitigation

Even JPMorgan, Siemens, Credit Suisse — firms with extensive bribery records — pay first.

Reputation does not function in corruption. Contrast: enablers deliver first in tax evasion (89% reputation).

This is a second, independent test of reputation’s absence in corruption — complementing the 5% rate from direct classification.

The logic: if reputation worked, multinational firms that have bribed across dozens of countries (Siemens, Odebrecht, JPMorgan) should have a cross-deal reputation for reliable payment. Officials should be willing to deliver first, trusting the firm’s track record. We should see “recipient first” deals especially for repeat-player firms.

We don’t. In 772 FCPA cases, the firm pays first 90% of the time. The 16 cases where the official moves first are not reputation-driven — they’re shakedowns (official threatens to cancel existing contracts, 3 cases), structural kickbacks from generated revenue (4 cases, Venezuela bond trading), patronage hiring (4 cases, China — hiring an official’s relative is cheap), or post-retirement promises (1 case, Korea).

The contrast with tax evasion is stark. In tax evasion, enablers deliver first — they set up the offshore structure and manage assets before the client fully commits. The 89% reputation rate shows that enabler reputation IS functioning: clients trust the enabler’s brand. Enablers can advertise reliability legally (“professional discretion”), so reputation accumulates and is transmissible. Corrupt officials cannot advertise “I deliver on bribes” — so firms never trust officials to deliver first, no matter how much repeat business they’ve done.

This is the fourth Gambetta condition in action: the reputation signal must be transmissible without self-incrimination. Offshore enablers can transmit it (the service is legal). Corrupt officials cannot.

The Drousys paradox: trust technology = prosecution evidence

Odebrecht’s encrypted system (Drousys) solved three trust problems:

Function	Trust problem	How
Secrecy	Denunciation	Encrypted, separate from corporate IT
Coordination	Performance	Real-time tracking across 12 countries
Kompromat	Hold-up	Records = mutual hostage

But the records that enabled trust became the prosecution evidence.

→ 77 executives signed plea bargains in a cascade — hierarchical trust collapses from the top down

→ Compare: Putin network (kinship) → zero defections. UBS (institutional) → bank pays fine, system continues.

The Drousys paradox captures a structural tension in contractual formalization: the documentation that makes corrupt arrangements manageable also makes them prosecutable. Drug trafficking avoids this by relying on oral agreements and relational trust — no paper trail, but also no coordination across 12 countries.

Three failure modes, three trust types: - Hierarchy (Odebrecht) → cascade collapse. Once Paulo Roberto Costa cooperated, 77 executives followed. - Kinship (Putin network) → resilient, zero defections even after Panama Papers. Generational wealth transfers maintained loyalty. - Institutional (UBS) → institution pays $780M fine, turns over 4,450 client names, individual bankers cooperate — but Swiss banking continues. The brand outlasted the scandal.

For Skatteforsk: the UBS comparison is key. UBS’s trust was institutional — Swiss banking law, 150-year brand, numbered accounts. When enforcement came, the institution absorbed the shock (paid the fine, reformed compliance) while individual evaders were exposed. Odebrecht’s trust was personal (Marcelo’s spreadsheet, personal negotiations with ministers) despite the institutional wrapper (SOD). When Marcelo was arrested, the system collapsed. This is the fundamental difference: institutional trust survives personnel changes; personal trust does not.

When politicians CAN use violence: India

46% of Indian MPs face criminal charges; 31% face murder, kidnapping, assault (ADR 2024)

Pattern	Mechanism	Example
Bahubali (strongman)	Violence IS the trust technology	Atiq Ahmad: 195+ cases, ran empire from jail
Resource mafia	Violence enforces corrupt monopoly	Sand/mining mafias kill journalists, RTI activists
Vertical integration	Criminal becomes politician to eliminate trust problem	Vaishnav (2017): “when crime pays”

Chemin (2012): criminal politicians reduce bribes by 34% but increase violent crime by 25%

→ Violence and market corruption are substitutes, not complements

Important nuance to our claim that “corrupt officials cannot credibly threaten violence.”

India shows that in weak-state environments, politicians CAN credibly threaten violence — when they enter politics from criminal backgrounds. 46% of newly elected Indian MPs in 2024 face criminal charges; 31% face serious charges including murder. These are not ministers who happen to be violent — they’re strongmen who happen to hold office.

Vaishnav (2017) provides the theoretical framework: criminals entered politics through “vertical integration” — a transaction cost economics concept. The corrupt deal between criminals and politicians was hard to enforce (our trust problem!), so criminals eliminated the bilateral trust problem by becoming the politician themselves. This maps directly onto our framework.

Chemin (2012, RDD) finds the striking result that criminal politicians reduce bribe-taking by 34% but increase violent offenses by 25%. This means violence and market corruption are substitutes: the bahubali doesn’t need to bribe the bureaucrat — he coerces them directly. This is exactly what our framework predicts: if violence is available as a trust mechanism, contractual/intermediary mechanisms become unnecessary.

The key distinction: our data (FCPA, World Bank, SFO) predominantly captures “insiders” — multinational executives and career officials for whom violence is catastrophic. In settings where political office is captured by criminal entrepreneurs (Johnston’s “oligarchs and clans”), the violence channel opens and corruption resembles organized crime. The 18% hierarchy/coercion rate in our corruption data is an average across both regimes.

This connects to the institutional analysis: weak state capacity + caste-mediated electoral accountability → violence-based corruption. Strong institutions + professional bureaucracy → contractual/intermediary-based corruption.

Connecting to the tax evasion literature

What they measure

• Alstadsæter, Johannesen & Zucman (2019): top 0.01% evade ~25% of taxes owed
• Zucman (2014): 8% of global wealth held offshore
• Johannesen & Zucman (2014): crackdowns redirect deposits between havens
• Johannesen et al. (2020): enforcement most effective against smaller evaders

What we add

• How is evasion organized?
• Why do crackdowns redirect rather than repatriate? → contractual trust is portable
• Why are the wealthy harder to catch? → redundant enabler networks
• Why do enablers comply? → professional reputation + legal contracts

Harrington (2016) · Findley, Nielson & Sharman (2014) · Seabrooke & Wigan (2017)

This slide connects our trust framework to the existing tax evasion literature — directly relevant to this audience.

Alstadsæter, Johannesen, and Zucman (2019, AER) showed that offshore evasion is extremely concentrated at the top. Our framework explains why: the organizational infrastructure (wealth managers, corporate service providers, trust companies) has fixed costs that only the wealthy can justify. The trust problem is solved through contractual formalization, which requires legal and financial expertise — creating a natural scale threshold.

Johannesen and Zucman (2014) found that G20 crackdowns on specific havens redirected deposits to other havens rather than repatriating them. Our framework explains the mechanism: the enabler relationships are contractual, so they are portable. When Switzerland is compromised, the wealth manager can restructure the arrangement through Singapore or Dubai using the same contractual templates. If the trust were relational (as in drug trafficking), switching would be much costlier.

Johannesen et al. (2020) showed that US enforcement (FBAR/FATCA) was most effective against smaller evaders. Our framework predicts this: wealthier clients can afford redundant enabler networks, so disrupting one node doesn’t destroy the arrangement. Smaller evaders have a single enabler relationship — disrupt it and they’re exposed.

Harrington (2016, “Capital without Borders”) provides the ethnographic foundation: 65 interviews with wealth managers across 18 countries, documenting exactly how the client-enabler trust problem is solved through professional discretion, reputation, and contractual infrastructure.

Findley, Nielson, and Sharman (2014, “Global Shell Games”) ran a field experiment with 7,400+ solicitations to corporate service providers worldwide. ~50% were willing to form anonymous shells without proper identity verification — and providers in OECD countries were less compliant than those in tax havens. The supply side of the enabler market is broad.

Seabrooke and Wigan (2017) provide the “global wealth chains” concept — the multi-jurisdictional structure through which assets are hidden. We add that each link in the chain is a trust problem solved by a specific mechanism.

UBS: the trust infrastructure of tax evasion

Trust mechanism	How UBS solved it
Institutional	Swiss banking secrecy law — criminal penalties for disclosure (Art. 47)
Reputation	150-year brand; “your fear is unjustified… protection cannot be compromised” (2002 client letter)
Contractual	Wealth management agreements, QI agreement with IRS, nominee shell companies
Relational	Art Basel, yacht races, concierge services — quarterly personal visits
Kompromat	Mutual criminal exposure: neither side can report without self-incrimination
Intermediary	NY referral desk → Swiss bankers; external lawyers → Liechtenstein trust managers

Every trust mechanism in our taxonomy — in a single case.

Return to UBS. The trust infrastructure is layered and redundant — every mechanism in our taxonomy appears.

The foundation is institutional: Swiss banking secrecy law (Article 47 of the 1934 Banking Act) made it a criminal offense — up to 3.5 years in prison — for bankers to disclose client information. This is not corporate policy but state-enforced secrecy. The Swiss government itself was the trust guarantor.

On top of this: UBS’s 150-year institutional reputation as a trust anchor. In 2002, when clients worried about new IRS regulations, UBS sent a written letter: “your fear is unjustified… the protection of client data can not possibly be compromised.” An explicit, documented trust renewal.

Contractual formalization: the Qualified Intermediary agreement with the IRS was supposed to require reporting, but UBS exploited a loophole — accounts held by shell companies (BVI, Hong Kong, Panama, Liechtenstein) were not reportable as belonging to US persons. The wealth management agreements and nominee structures provided additional contractual layers.

Relational trust: Birkenfeld joined client Igor Olenicoff on yacht trips, delivered luxury goods to hotel rooms, bought diamonds in Geneva and smuggled them in a toothpaste tube. ~90% of clients: tax evasion as “an unspoken understanding rather than explicit discussion.”

Kompromat: bilateral — both bank and client were committing crimes. Any disclosure would be self-incriminating. This is the same mechanism as Odebrecht’s Italiano spreadsheet, but operating through mutual criminal exposure rather than a deliberate ledger.

Intermediaries: UBS America’s New York referral desk steered clients to Swiss colleagues. External lawyers and accountants set up offshore shell companies. Liechtenstein trust managers moved assets to banks with “greater secrecy” when pressure increased.

The key contrast with Odebrecht: UBS could build most of its trust on institutional and contractual foundations (Swiss law, bank brand, formal agreements). Odebrecht had to build trust from scratch through relationships, codenames, encrypted systems, and kompromat — because corruption has no legal infrastructure to lean on.

The Jahre case: when trust breaks down

Anders Jahre (1891–1982) — Sandefjord shipping magnate

• 1926: Lazard Brothers help establish Salemas in Panama — formally owned by the bank, Jahre is the real owner
• 1930s–70s: builds network of Panama and Liberian shells (Pankos, Cornwall Shipping, Intercontinental Chartering)
• Control maintained via management agreements — Jahre owns nothing on paper
• Thorleif Monsen (Cayman Islands) — his stråmann: holds shares formally, takes instructions from Jahre’s lawyer Bjørn Bettum

After Jahre’s death: Monsen transfers assets to Aall Trust and Banking Co. — his own bank on the Cayman Islands. Creates new foundations for himself and his family. The straw man ran away.

The opacity that hides assets from the tax authorities also hides them from the beneficial owner’s heirs.

A Norwegian case the audience will know well. Jahre created one of the earliest documented Norwegian offshore tax evasion structures — starting in 1926 with Lazard Brothers.

The trust architecture: contractual formalization (management agreements giving Jahre real control despite no formal ownership) + professional enablers (Lazard Brothers, Panama attorneys, Guinness Mahon bank). Bjørn Bettum (with Lazard background) held instruction rights over the entire structure, including over Monsen.

But the inherent vulnerability: the arrangement depended on Jahre being alive to enforce it. After his death, Monsen — who had superior practical knowledge of the structure — transferred assets to his own Cayman bank (Aall Trust and Banking), created new foundations for his family, and essentially stole the wealth. Despite Lazard’s safeguards (requiring unanimous advisor approval for changes), Monsen exploited the post-death vacuum.

The estate administrators (bobestyrere) faced what they called a “forvirrende nettverk av selskaper og overføringer” — a confusing network of companies and transfers — and spent years trying to unwind the arrangements.

This is the trust problem incarnate: the three-document system that solves nominee trust (nominee declaration, power of attorney, undated resignation letter) means nothing when the principal is dead and the nominee controls the keys. Harrington (2016) documents exactly the same vulnerability in modern wealth management.

Source: Gram (2017), “Penger på rømmen” (Money on the Run).

The three-document system

How does a client trust a nominee who officially “owns” their assets?

Document	Function	Trust problem solved
Nominee director declaration	Nominee commits to follow owner’s instructions	Performance risk
Power of attorney	Transfers operational control back to real owner	Hold-up risk
Undated resignation letter	Owner can remove nominee at will	Denunciation risk

Contractual formalization as a trust technology — the core product of the enabler industry

This is the micro-mechanism of contractual formalization in tax evasion. Nielson and Sharman (2022, “Signatures for Sale”) document this three-document system that solves the trust problem between beneficial owner and nominee.

The nominee director declaration commits the nominee to act on the owner’s instructions — solving performance risk (will the nominee do what I want?). The power of attorney transfers operational control back to the real owner — solving hold-up risk (the nominee can’t lock me out). The undated resignation letter gives the owner unilateral power to remove the nominee — solving denunciation risk (if the nominee becomes unreliable, I can replace them instantly).

This is pure contractual formalization: three legally structured documents that solve three distinct trust problems. The documents are the core product of the corporate service provider industry — Findley et al. (2014) showed that ~50% of providers worldwide will create these arrangements without proper identity verification.

Note: the Jahre case shows what happens when this system fails. The undated resignation letter means nothing when the principal is dead and the nominee holds the keys. Monsen had no one to enforce the documents against him.

Global wealth chains as trust architecture

 US client (e.g. Olenicoff, $200M)
         │
    Wealth management agreement ←── contractual
         │
 UBS private banker (Switzerland)
         │
    Referral to external counsel ←── intermediary
         │
 Lawyer / consultant
         │
    Shell company formation ←── contractual + nominee
         │
 BVI / Liechtenstein / Panama entity
         │
    Numbered account (UBS Geneva)

Each layer: a trust problem solved by a legal contract (Seabrooke & Wigan 2017)

This is what tax evasion organization looks like — a “global wealth chain” (Seabrooke and Wigan 2017). Each link is an intermediary connected by a legally enforceable contract. The wealth management agreement, the trust deed, the nominee arrangement — all are formal contracts that solve the trust problem between client and enabler at each layer.

Compare to drug trafficking: supplier → courier → distributor → dealer. Each link sustained by relational trust (years of dealing) backed by violence. No contracts. No legal enforceability. Much stickier, much less scalable.

García-Bernardo et al. (2017) mapped the global network and identified conduit jurisdictions (Netherlands, UK, Ireland) that channel flows to sink jurisdictions (Caymans, BVI, Luxembourg). Each conduit-sink link is a contractual relationship between professional service providers.

The policy implication: every link in the wealth chain can be regulated because it involves a licensed professional and a legal contract. Beneficial ownership transparency, gatekeeper liability, automatic information exchange — these attack the contracting infrastructure. You cannot do the same to a drug trafficking network.

When enablers betray: evidence from ICIJ

Case	What happened	Trust failure
Cordish family (US)	Financial planner managing Cook Islands trusts diverted $12M to personal real estate	Enabler exploited information asymmetry
BVI shell operator	Ran a $33.5M Ponzi scheme against 132 investors using their own shell structures	Nominee turned predator
Trust protector	Redirected assets to his own relatives by misrepresenting beneficiaries’ wishes to Swiss trustees	Intermediary captured the chain

“Once your money is offshore, you have no legal standing to go get it.”

These cases from ICIJ investigations illustrate the fundamental vulnerability of the enabler-dependent model. The Cordish family trusted a financial planner who diverted $12 million of trust assets into personal real estate purchases. The BVI operator ran a Ponzi scheme against 132 clients using their own shell company structures. The trust protector — supposed to safeguard the beneficiaries’ interests — redirected assets to his own family.

Each case is a trust failure of a different type: - Performance risk: the enabler doesn’t do what they promised (Cordish) - Hold-up risk: the enabler uses their position to extract (BVI operator) - Denunciation risk inverted: the enabler threatens to expose (trust protector leveraged secrecy)

The victim’s quote captures the core problem: “Once your money is offshore, you have no legal standing to go get it.” The opacity that hides assets from tax authorities also hides them from the rightful owner. This is not an accident — it is the structural consequence of solving the tax evasion trust problem through intermediary opacity.

Compare to the Jahre case: same logic, different century.

When trust breaks down: two parallel stories

Odebrecht (corruption)

2015: Migliaccio discovers he is being set up as fall guy

Plea bargain cascade: 77 executives cooperate

Why? Asymmetric exposure — the company could blame “rogue employees” while individuals faced decades in prison

The Italiano spreadsheet — meant to enforce trust — became the prosecution’s key evidence

UBS (tax evasion)

2005: Birkenfeld discovers internal CYA memo preparing to scapegoat front-line bankers

Cooperation cascade: 54,000 taxpayers come forward; $16B recovered

Why? Asymmetric exposure — the bank could blame “rogue bankers” while individuals bore all risk

The 2006 whistleblower law ($104M award) provided a credible outside option

Both collapsed the same way: the institution tried to shift blame downward, and an insider defected.

The parallel is striking. Both schemes collapsed when an insider discovered that the mutual vulnerability was asymmetric — the organization could shift blame to individuals while protecting itself.

Odebrecht: Fernando Migliaccio, who ran the SOD’s offshore payment infrastructure, cooperated with prosecutors. His plea bargain testimony provided operational details of the entire system — codenames, account numbers, doleiro networks. 77 Odebrecht executives eventually cooperated. The Italiano spreadsheet, which Marcelo Odebrecht had maintained as a trust/enforcement device (tracking who owed what to which politician), became the prosecution’s central evidence. The tool of trust became the instrument of destruction.

UBS: Bradley Birkenfeld discovered a 2005 internal memo that “radically contradicted the training and practices that had driven the bank’s offshore business for years.” He realized UBS was preparing legal cover to frame front-line bankers as rogue actors. He resigned, then contacted the DOJ in 2007 — providing PowerPoints, emails, phone lists, training manuals. The cooperation cascade was enormous: 54,000+ US taxpayers came forward through voluntary disclosure programs, paying $16+ billion in back taxes. The Swiss Bank Program eventually reached 80+ banks paying $1.36 billion combined.

The structural lesson: trust systems that depend on symmetric mutual vulnerability are stable. But when organizations can shift blame to individuals — breaking the symmetry — the incentive to defect becomes overwhelming, especially when outside options exist (plea bargain deals, whistleblower awards).

This has a direct policy implication: whistleblower laws and leniency programs work precisely because they exploit the asymmetry in organizational trust. They provide a credible outside option that makes defection rational. The 2006 US whistleblower law (up to 30% of recovered taxes) turned Birkenfeld’s $0 outside option into a $104M payoff.

Policy: disrupt the trust infrastructure

Target the mechanism, not just the jurisdiction

Policy	Disrupts	Effect
Beneficial ownership registries	Intermediary opacity	Enablers become identifiable
Gatekeeper liability	Enabler supply	Fewer professionals willing to facilitate
CRS / AEOI	Cross-border trust chains	Restructuring, not repatriation
Whistleblower incentives	Asymmetric trust	Insiders defect when outside option > loyalty

Birkenfeld’s $104M award triggered 54,000 voluntary disclosures and $16B in recoveries.

But: enforcement is regressive — wealthiest maintain redundant enabler networks (Johannesen et al. 2020)

The policy slide, reframed for a tax audience.

Key insight: standard tax enforcement tools (CRS, AEOI, beneficial ownership registries) can now be understood through the trust lens. Each one disrupts a specific trust mechanism in the wealth chain.

CRS/AEOI: disrupts the cross-border trust chain by making information flow to home authorities. But — and this is where Johannesen and Zucman (2014) is important — because the enabler relationships are contractual rather than relational, they are portable. The same wealth manager can restructure the arrangement through a non-participating jurisdiction using the same contractual templates. This explains why crackdowns redirect deposits between havens rather than repatriating them. UBS illustrates: when pressure mounted, clients were referred to Liechtenstein trust managers and smaller regional Swiss banks. The contractual templates moved; only the institutional name changed.

Gatekeeper liability: attacks the supply side of the enabler market. If lawyers, accountants, and corporate service providers face personal liability for facilitating evasion, fewer will participate — shrinking the pool of available trust partners. Findley et al. (2014) showed that compliance pressure reduces (but does not eliminate) the supply of anonymous shells.

Whistleblower incentives: the most powerful trust-disruption tool in the UBS case. The 2006 US Tax Relief Act created a 30% reward for tax fraud whistleblowers. Birkenfeld’s $104M award triggered the largest cooperation cascade in tax enforcement history. The mechanism is precise: whistleblower laws exploit the asymmetric trust vulnerability (H-UBS-2). In any institution-organized scheme, the institution can scapegoat individuals — creating an incentive to defect. The financial reward makes defection overwhelmingly rational. Agent commission bans (previously in this table) attack contractual formalization but are harder to implement; whistleblower incentives attack the weakest structural link.

The regressive enforcement caveat is important for this audience: Johannesen et al. (2020) showed that US enforcement (FBAR/FATCA) was most effective against smaller evaders. Our framework explains why: the ultra-wealthy can afford redundant enabler networks — disrupting one link doesn’t break the chain because there are backup relationships. Smaller evaders have a single enabler; disrupt it and they’re exposed.

This means trust-disruption policies work, but disproportionately affect the middle of the distribution — not the top 0.01% that Alstadsæter, Johannesen, and Zucman (2019) showed evade the most.

What’s next

• Trust dynamics: 1,189 cases with breakdown/adaptation/formation data — how schemes evolve over time
• Deal structure: 772 FCPA cases with sequencing, fee structure, payment method
• Enforcement natural experiments: does CRS/AEOI change organizational form, or just redirect it?
• Enabler stratification: do wealthier evaders use different trust mechanisms?
• Validation: human-coded sample (80 cases) for inter-rater reliability

Paper: “How Is Illegal Activity Organized?” (6,486 cases, 6 crime types)

Four next steps, each connecting to this audience’s interests.

ICIJ data: We’re processing Panama Papers, Paradise Papers, and Offshore Leaks articles through the same extraction pipeline. This will fill in the tax evasion column in the cross-crime comparison. We expect the highest contractual formalization rate of any crime type.

Enforcement natural experiments: CRS/AEOI rollout provides plausible temporal variation. Our framework predicts it raises the trust barrier — selecting for more sophisticated enabler arrangements rather than eliminating evasion. The Johannesen and Zucman (2014) deposit redirection finding is consistent. We can test whether the organizational FORM changes (e.g., from single-enabler to multi-layer chains) even if the total VOLUME doesn’t.

Enabler stratification: Johannesen et al. (2020) showed enforcement is regressive. We can test whether the trust mechanisms themselves differ by scale — do the ultra-wealthy use different mechanisms (redundant enablers, more contractual layers) than smaller evaders?

Validation: 80 cases coded by hand to measure agreement with the LLM classifier. Crucial for credibility.

Thank you

6,486 cases · 6 crime types · 173 countries

Odebrecht needed codenames and encrypted systems. UBS needed Swiss law and a 150-year brand. Same problem — opposite solutions.

Skatteforsk seminar · NMBU, Ås