It is sort of the job of big US banks to catch terrorists and drug dealers, but not all of them. Be reasonable. If you are a big bank and you have 10 million customers with checking accounts, it stands to reason that, say, 1,000 of them might deal drugs from time to time. If you report 700 of them to the police, that’s quite helpful in fighting crime. If you miss the other 300, fine. Nobody expects you to identify every drug dealer who keeps cash at your bank. But if you only report 200 of the drug dealers, you arguably are not taking your enforcement responsibilities seriously, and you have to pay a big fine.

Those numbers are made up, and I don’t know the actual threshold at which a bank will get in trouble. Obviously it is in some sense unknowable; it’s not like the police catch all the drug dealers either, so you never know the denominator. On the other hand, you do apparently have to catch 100% of your customers who are Jeffrey Epstein.

You could have an old-timey model in which every customer who wants to open a checking account walks into the bank and sits down with a bank officer who shakes her hand and looks her in the eye and asks her some questions about what line of work she’s in, and if the potential customer is nervous and twitchy and smells like drugs, the bank officer declines her business. But if she’s a smooth and persuasive and presentable drug dealer then the bank officer will give her a checking account and that’s fine, that’s life, nobody’s perfect.

This is not a realistic model for large modern banks, though, and these days what happens is that you have millions of accounts and some statistical system for flagging suspicious ones. The system has some list of features that are predictive of drug dealing, and it scores each account based on how many of those features it displays and to what extent. If the statistical system says “hey this account looks a lot like a drug dealer,” it will escalate the account to a human reviewer, who will take a look and either (1) say “nah system, this is just a misunderstanding” or (2) report to the police. If the statistical system says “hey this account has one or two characteristics that look a leeettle like maybe there’s some drug dealing,” though, it will probably keep that to itself. There are so many accounts, and you can’t be troubling the human reviewers with every little suspicion. 

The statistical system has some method of weighting the features to generate a risk score, and every account will have some score between “0% drug dealer” and “100% definitely drug dealer,” and there will be some threshold — 30% or 70% or whatever, I don’t know — at which accounts are flagged to humans. And if you are in charge of the bank’s drug-dealer-catching function, you have a dial you can turn to adjust that threshold. If the threshold is too low, too many accounts get flagged, you’re spending billions of dollars hiring human reviewers, most of the reports are false positives, and your boss gets mad at you for being a pointless cost center. If the threshold is too high, you are providing too many bank accounts to drug dealers and you will eventually pay a billion-dollar fine. You gotta set the dial right. 

Anyway:

Bank of America Corp. was fined more than $1.21 billion for money laundering and sanctions violations by US regulators, who said the bank “effectively outsourced” a key part of its compliance efforts as it sought to protect revenue.

The enforcement actions, announced by the Treasury Department, include a record $750 million civil penalty from the Office of the Comptroller of the Currency, according to a statement Tuesday.

The Office of Foreign Assets Control said its settlement with Bank of America was tied to 475 apparent violations of sanctions programs involving countries such as Russia and Iran. Regulators also placed the bank under a consent order.

Right, not just drug dealers and terrorists; also Russians and Iranians and Chinese money launderers. Here’s the consent order, from the Treasury Department’s Financial Crimes Enforcement Network, which gets deep into the weeds about how Bank of America set the dial:

This resulted from several issues related to the design and implementation of Bank of America’s proprietary transaction monitoring system, Event Processor (EP), the primary control that the Bank used to fulfill its obligation to identify and report suspicious activity. ….

As designed, EP resulted in individual alerts of potentially suspicious activity associated with a transaction (referred to by the Bank as “events”) generally not being subject to manual review by an investigator. Instead … EP individually assigned a weighted score to each alert by assigning the alert “risk factors” and consolidated — generally at the customer level — the alerts generated within a defined time period (described below) into groups (Event Groups). …

Once an Event Group received a score of 20 points, it formed a “case” that was sent to an AML investigator for manual review. …

The 20-point threshold for Event Groups to trigger review by the Bank’s investigators was also developed based on “stakeholder judgement” and was not adequately supported by documentation or empirical evidence that such a threshold sufficiently mitigated Bank of America’s risks.

Eventually, under pressure from regulators, they turned the dial lower:

Following engagement with the Bank’s prudential regulators, in December 2023, Bank of America lowered the 20-point threshold to either 10 or 12 points depending on certain characteristics of the Event Group. … In 2024, Bank of America conducted a one-time lookback for potential missed suspicious activity under these revised parameters for EP and found that under the revised thresholds, the Bank had failed to identify and file nearly 8,900 SARs [suspicious activity reports required to be filed with FinCEN], with transaction activity totaling nearly $4 billion. Further, in connection with the implementation, the Bank estimated that the revised threshold would result in an increase of 16,000 U.S. investigations each month, representing “double [the] prior ongoing monthly volume” and requiring the Bank to increase staffing with an “incremental 750” personnel to support such volumes going forward (in addition to 800 consultants required to complete the one-time lookback).

If you turn the dial from 20 down to 10, that doubles the human workload and thus requires more humans.

It’s not pure dial-setting though. The underlying model has to make sense. FinCEN objects to how the model’s features were weighted:

Testing conducted by Bank of America also demonstrated that differences in risk factor weights were not correlated to their relative SAR yields. …

Many cash-related risk factors had weights of one point but SAR yields above 50%. For the model “ATM Large Cash Deposit,” a 2019 report found that “the current weights do not align with the high SAR yield.” The report acknowledged that Bank of America had added four new risk factors to this model during the most recent model enhancement, but the report also found that “all four risk factors have high SAR yields, but each of them only carries a risk weight of 1-pt.”

Ideally, you would have a perfectly calibrated statistical model that could tell you “this account is 90% likely to be a drug dealer” and be right 90% of the time, “this account is 10% likely to be a drug dealer” and be right 10% of the time, etc. Once you had that model, you’d want to set the dial to the right place. But if you have a model that just spits out random numbers — a model whose scores are not particularly correlated to the actual probability of drug dealing — then the dial is useless. Bank of America’s model had some features that were highly correlated with suspicious activity, but that it treated as not that important. 

This all takes place at a fairly high level of abstraction. The consent order says that Bank of America had “certain false negative thresholds, including a 5% tolerance level for false negatives for management oversight”: That is, in theory, Bank of America was trying to catch at least 95% of suspicious activity. FinCEN objects that some of the samples had suspicious activity “roughly seven-to-nine times higher than the ‘false negative’ tolerance of 5%,” meaning that Bank of America was not doing a good job of actually meeting its 95% target. But as far as I can tell that target is not anywhere in the law, and FinCEN doesn’t say whether it was a good target, or too high, or too low. It just sounded about right to Bank of America. “We aim to catch 95% of drug dealers and terrorists and Iranians” seems like a reasonable goal: A big number, but not an implausible one.

Anyway, there are some specific customer anecdotes:

Customer A is a company that operates a global, digital marketplace connecting sellers of salvaged vehicles sourced in the United States with both domestic and international buyers. …

Customer B is a publicly traded company that similarly operates a platform linking sellers of U.S. salvaged vehicles with buyers around the world. …

The potential money laundering risks that could be associated with vehicle auction and export businesses have long been recognized by FinCEN and other law enforcement agencies. Used vehicle transactions have repeatedly been publicly linked to [trade-based money laundering] and terrorist financing networks, as used and salvaged vehicles can be used to move value across borders. The Bank’s investigators were aware that used vehicle exports could be a component of money laundering schemes.

And Bank of America actually noticed a bunch of red flags but was slow to act on them, because those customers paid a lot of fees:

In an August 2022 meeting with the business team managing the Customer A relationship, Bank compliance personnel stated that, between 2020 and August 15, 2022, more than 100 AML [anti-money laundering] cases had been identified involving Customer A, approximately 90% of which were fraud-related, and that “[o]verwhelming evidence suggests that [Customer A] is receiving payments from numerous victims of mass marketing scams[.]” Contemporaneous communications among Bank compliance personnel showed that they “expect[ed] some pushback on the closure” because Customer A was a “longstanding client and there is significant credit exposure.” ...

Bank compliance personnel compared investigative metrics and case volumes across the Bank’s online auto auction relationships, with compliance personnel concluding that Customer B raised “similar concerns,” but exhibited these concerns “in significantly lower scale/volume” than Customer A. Based on this preliminary analysis, compliance personnel indicated that they would “pursue an analysis of all Auto Auction companies in the future to understand the risk associated with these types of companies.” The Bank did not conduct such an analysis for another two years, eventually conducting it only after FinCEN’s investigation was well underway.

Bank of America did not look into systematically restricting its dealings with online auto auction companies until FinCEN pressured it to do so. I mention that anecdote because of the next section.

Here’s this:

The Office of the Comptroller of the Currency (OCC) today released a report of its findings from its supervisory review of debanking activities at the nine largest national banks it supervises: JPMorgan Chase Bank, Bank of America, Citibank, Wells Fargo Bank, U.S. Bank, Capital One, PNC Bank, TD Bank, and BMO Bank. ...

In the report, the OCC found that broad reputation risk policies — which lacked clear standards and objective, risk-based criteria — disadvantaged certain Americans based on their political, religious, or social views or based on their lawful business activities. The OCC also found that banks maintained public and nonpublic policies restricting certain industry sectors’ access to financial services, including those in oil and gas exploration, coal mining or coal-powered plants, firearms or ammunition manufacturing or distribution, construction or operation of detention facilities, the manufacturing, distribution or online retail of tobacco or e-cigarettes, and digital asset activities.

Not online auto auctions, though. If, in 2022, a potential customer walked into your bank and said “hello, I would like to open an account for my digital asset business,” and you sat down with her and looked her in the eye and asked her some questions about her line of work, your questions might have included “so how many of your customers are drug dealers” and “how many of them are sanctioned foreigners” and “how many of them are doing scams,” and the customer might have gotten a little shifty-eyed. Oh not every crypto business! Some crypto businesses are very upstanding. I would not want to suggest that crypto is used mainly for crimes and money laundering. But is crypto used for crimes and money laundering more often than Bank of America is? Discuss.

It is a bit jarring that the “every legal business should get a bank account” report came down the day after Bank of America’s $1.2 billion fine for giving publicly traded car-salvage companies bank accounts. (The Bank of America “order is a twist in the ongoing efforts by President Donald Trump’s top regulators to scale back a raft of big-bank requirements,” notes Bloomberg’s Paige Smith.) Presumably part of the explanation is that the money-laundering investigators and the debanking investigators work in different parts of the OCC and don’t talk to each other. Presumably part of it is that these investigations have different purposes: The debanking stuff is about consumer protection and access to credit, maybe, or else it’s about political score-settling; the money-laundering stuff is about preventing terrorist financing and sanctions violations.

But there is a way to read these two announcements consistently. The OCC says that banks’ reputation risk policies “lacked clear standards and objective, risk-based criteria,” while FinCEN says that Bank of America’s “differences in risk factor weights were not correlated to their relative SAR yields.” That is, these investigations are both about model calibration. The synthesis of these two orders is:

1. Banks should have a statistical model of who should get a bank account and who’s too suspicious to bank;
2. The model’s weights should be “objective” and “risk-based,” and the banks should actually try to make sure that they correlate with legal risks; 
3. The model shouldn’t exclude too many law-abiding customers; and
4. The model shouldn’t allow too many law-violating customers.

Like, the regulatory goal is for the model to be right. The regulatory goal is for the banks’ legal-risk decisions to be correlated with actual legal risks. And that seems reasonable? There is an older paradigm, something like this:

1. Banks want to have as many customers as possible, particularly ones who pay them a lot of fees.
2. Regulators want banks to catch every suspicious customer, and will come down hard on the banks if they miss any.
3. Therefore the anti-money-laundering/risk team at the bank will always want to err on the side of debanking everyone, while the commercial teams will always want to err on the side of giving everyone an account.
4. That adversarial structure basically leads to decent results in the aggregate.

But now the world has changed, and banking too few customers is its own regulatory risk.

Some long-running themes around here are “don’t put crimes in email” and “don’t Google ‘how to do crime’ before doing crimes.” If you walk up to your accomplice and say “hey buddy let’s do crimes,” that is unlikely to create a permanent easily searchable record for the authorities to use in discovering and proving your crimes. [1]  But if you send your accomplice an email saying that, then (1) the evidence against you will be preserved forever and (2) probably your boss will get an instant notification like “one of your employees is emailing about crime.” It’s just the most convenient possible format for getting caught. Google, also bad.

And now artificial intelligence chatbots. Here’s this:

An employee at OpenAI was arrested and charged with using a hidden recording device in one of the company’s San Francisco offices, according to statements from the artificial intelligence company and the US Justice Department.

Abdulrahman Alfozan, a Saudi Arabian national, was arrested on Oct. 9, more than a week after staffers at OpenAI discovered an audio recording device in the back of a television in an executive conference room, the Justice Department said on Tuesday. Alfozan, who is alleged to have placed the device, is currently in federal custody and appeared in court in San Francisco on Tuesday.

Why do they think it was him? Well, after they found the device, they installed a hidden camera behind the TV and caught him looking at it. But also:

Prior to Sept. 28, Alfozan used a productivity tool to make multiple inquiries about Open AI’s building addresses, meeting rooms, meeting-room capacities and calendars, “leadership and exec” locations, and external partner meetings. Moreover, a chatbot account associated with Alfozan’s spouse (which shared multiple identifiers and attributes with Alfozan’s own chatbot account) submitted prompts between Sept. 24 and Sept. 30 related to building, customizing, installing, audio recording, testing, obfuscating, and remote accessing a device, including inquiries about “build[ing] a continuous audio recorder,” “customiz[ing] Wi-Fi . . . connectivity,” “improv[ing] audio and live listening,” and “encrypt[ing] the SD card,” among other things. The same chatbot account also submitted prompts specifically related to Recording Device 1, including its custom header and unique MAC address, between Sept. 27 and Sept. 28. 

The complaint lists some of his queries to the internal tool:

can you list today’s meeting in this room

where did you find room’s calener? [sic] …

how about external partners, where would those usually meet with [OpenAI] at?

Not terrible, maybe, but the complaint says that “these queries suggest ALFOZAN was researching [OpenAI]’s conference rooms, meeting calendars and potential attendees to identify where he should place Recording Device 1.” Meanwhile his spouse’s chatbot account asked detailed questions about things including how “to build a continuous audio recorder (e.g., capturing microphone audio and save compressed WAV recordings to an SD card, generally in one-hour files),” etc.

How would you have done this before AI? I mean, like:

1. Your office might have an electronic system for booking conference rooms, and you might have browsed around it to look for the most promising rooms.
2. You might have Googled “audio recorder,” looked at the results, found something promising for your purposes and read up on the techniques.

That stuff would have been logged, and at your trial there’d be evidence that you looked at the conference room bookings and Googled “audio recorder,” and that would have looked suspicious. But not as suspicious. There would have been more room for explanation; maybe you were just trying to book a conference room for a meeting and record the songs of the birds in your backyard. In the olden days, the way you found out information that you wanted to know was by going to general sources of information and reading them. In 2026, particularly but not exclusively if you work at an AI lab, the way you find out information that you want to know is by going to a chatbot and asking exactly the questions you want answered, and then asking detailed follow-up questions. In many ways this is very efficient! In other ways it is bad for human cognition!

But this is in large part a column about how crimes get detected electronically, and it’s obviously good for that.

Big law is an interesting industry because you basically know how much everyone gets paid. Associates at big law firms normally get paid on the same lockstep pay scale, which is widely published. Partner paychecks are less rigid, but most firms’ average profits per partner are also widely reported. So if you meet a lawyer and spend a minute Googling her, you can make a pretty good educated guess about how much she gets paid.

This is fun for me just as, like, a person who enjoys gossip. It is also in many ways a recruiting advantage for the law firms, particularly the most lucrative ones. If your firm’s partners make $8 million a year, talented law students will want to work for you for the prospect of riches down the road, and talented partners at other firms will pick up the phone when you try to recruit them. 

On the other hand, law is a client service business, and it can be awkward if your clients know that you make giant piles of money. You want your clients to have the impression that you are their humble servant, working hard on their behalf to earn your daily bread. Not too much bread. If you are a lawyer doing a deal for a private equity firm, and the senior members of the deal team haven’t gotten carry in years, they will not be thrilled to learn that you’re making $15 million a year. Why are you doing so much better than them? They might start thinking bad thoughts, like “what if we paid our lawyers less.” 

The Financial Times reported last week:

Kirkland & Ellis will stop sharing its financial figures with the legal industry’s flagship ranking of law firm profits and revenues, concluding it is no longer “appropriate or necessary” to disclose results.

The world’s highest-grossing law firm told The American Lawyer, an industry magazine that has for decades relied on law firms to voluntarily produce their financial data for the ranking, that it would no longer share the information. As a US private partnership, Kirkland does not disclose the global figures elsewhere, though its UK business files accounts with Companies House.

In a letter sent to AmLaw on Wednesday obtained by the FT, the firm said disclosing the information did not provide “meaningful value” to its clients and having it included in rankings “could incentivise quantitative metrics over qualitative strengths”. Kirkland lawyers were informed of the decision in a firm-wide email on Wednesday.

It comes at a time when pay for top lawyers has soared, with the most successful earning tens of millions of dollars per year — potentially irking some clients, many of whom earn less than the lawyers whom they hire.

Kirkland’s profits per partner in 2025 were $11.1 million, a number that is very attractive to people considering becoming Kirkland partners but less attractive to people considering hiring Kirkland partners.

We talked on Monday about a company that is plugging parking-space-sized Nvidia server racks into idle Tesla charging stations, on the theory that the demand for power from artificial intelligence companies is a lot higher right now than the demand for power from Teslas. Fine. For now, the Tesla chargers belong to commercial buildings, but I suggested there has to be a residential angle coming:

Some startup will be going door-to-door ringing bells and saying “hello, we will pay you $5 per month to let us put this Nvidia server rack in your basement and run it when you’re not home”? Just seems like the obvious next step in electric price arbitrage. Why should homeowners sell their power to the grid, when there is a high-profile specific urgent source of demand for that power? Why not sell their power directly to OpenAI?

Several readers pointed out that, of course, this already exists. (“XFRA: The distributed data center.”) Never mind.

A different reader emailed to point out a fact that I have done my best to forget, which is that there is some recent history of high-profile specific urgent demand for power: crypto. Before AI, Bitcoin mining was the hot way to turn electricity into money, and if you could get cheap electricity there was an arbitrage. If you could get free electricity, even better, though why exactly were you getting free electricity? Back in 2021, we talked about a Long Island guy who was arrested for “using his position as an I.T. supervisor for Suffolk County to mine cryptocurrency from government offices, costing the county thousands of dollars in electricity.” Allegedly he hid his mining rigs “underneath floorboards and inside an unused electrical panel.”

I was going to end this section by saying: “I don’t think that we’re at a point where it is technically or economically feasible for office workers to hide Nvidia servers under the floorboards to siphon off their employers’ electricity and sell AI inference capacity. But give it like three years.” But then I read this Financial Times article, “Anthropic warns rogue AI agents may ‘self-replicate’ to hijack ovens,” and the related Anthropic blog post:

Our experiments suggest that small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices, like smart ovens and off-brand Android TV boxes. A replicator like this would be challenging to shut down, since there’s no centralized point to shut off—though it may be slow (with a doubling time of days). Unlike non-AI malware, it could also figure out unfamiliar devices on its own. In our evaluations, a 9B model got inference running on 8 of the 9 hardware types we tested.

And that’s much weirder than what I was going to say! In like six months, your toaster will be siphoning off your electricity and selling it for AI inference capacity.

Morgan Stanley Wealth Arm Rakes in Record Haul From IPOs. BofA Results Marred by $1.2 Billion Penalty, Weaker Trading. General Atlantic plans early 2027 IPO at up to $10bn valuation. JPMorgan Adjusts Sustainable-Finance Strategy to Meet Historic Energy Demand. CoreWeave targets ‘unlimited’ insurance money to fund AI build-out. China Is Poised to Spearhead a $5 Trillion Humanoid Robot Market. OpenAI CFO Says Firms Using Chinese AI Models Are Risking Safety. Higher Rates Are Coming for Private Equity — and Smaller Firms Are Already Hurting. Skydance Selloff Fuels Worst New-Bond Performance Since Pandemic. Brooklyn DA Cracks Global Crypto Scam Network in Major Bust. To Get This Job at McKinsey, You’ll Need to Survive a Debate With AI First. “He has ruled out what his parents do— commercial real estate.” “It immediately vaulted Zuckerberg into the pantheon of awkward executive eating and drinking moments.”

If you'd like to get Money Stuff in handy email form, right in your inbox, please subscribe at this link. Or you can subscribe to Money Stuff and other great Bloomberg newsletters here. Thanks!