Why banks have paid $3.5bn in fines over WhatsApp

Since December 2021, financial firms have paid roughly $3.5bn in penalties over staff conducting business on WhatsApp and other personal messaging apps. More than a hundred firms, including nearly every major name on Wall Street.

It began with J.P. Morgan Securities admitting to widespread recordkeeping failures and paying $125m to the SEC, plus $75m to the CFTC, in December 2021. It has not stopped since.

The obvious reading is that banks were careless about a messaging app. It is worth resisting, because it is wrong, and because the correct reading tells you something uncomfortable about a problem most companies now have with AI.

Almost every firm that got fined already banned WhatsApp for business use. They had the policy. They had attestations. What they did not have was any way to know the policy was being ignored, or any record of what had been said once it was.

What the fines were actually for

Start with the offence, because it is narrower than people assume.

Regulated firms have to keep records of business communications so that regulators can reconstruct what happened and when. That obligation is old, boring, and absolute. When staff moved conversations onto personal devices, those records simply did not exist.

So the violation being punished is not "your people used WhatsApp". It is "you cannot show us what was said". The SEC's own framing is that these failures likely deprived it of evidence in other investigations. The missing record is the harm.

That distinction matters enormously, because it changes what a fix looks like. If the problem were the app, banning the app would solve it. Every one of these firms had already tried that.

The scale, wave by wave

The listed actions alone exceed $3.1bn, and the total reaches roughly $3.5bn once parallel CFTC orders and other regulators are counted.

Look at the shape rather than the numbers. The first action landed in December 2021. The industry was on notice from that moment, in the most public way possible. Firms then spent eighteen months knowing exactly what the regulator was looking for, and the September 2022 sweep still cost $1.81bn in a single day. Penalties were still being handed out in 2025.

An entire industry, the most heavily supervised one there is, watched the first firm get fined and could not stop it happening to them.

Why the policy never held

Three things, and none of them are about carelessness.

The compliant channel was worse. A trader's client was already on WhatsApp. It was on their phone, it was quicker, and messages arrived. The approved system required a different device, or a desktop, or simply did not reach the person they needed. When the sanctioned path is slower than the unsanctioned one, the policy is a speed bump on a road people have a reason to drive down.

It went all the way up. This is the detail that makes the cases what they are. The SEC's orders repeatedly note that the conduct involved staff at multiple levels of seniority, including supervisors and senior executives, the very people responsible for enforcing the policy. A control that the leadership is personally violating is not a control.

Nobody could see it. Most of these firms genuinely did not know the scale until a regulator asked. There was no dashboard showing off-channel volume, because the whole point of off-channel is that it does not appear on your channels. The absence of the record was also the absence of the warning.

Put those together and you get the actual failure mode: a well-run organisation, with a correct policy, confidently unaware that the policy had been dead for years.

Then everyone fixed it, and a gap stayed open

The response was sensible. Rather than keep banning a channel people obviously needed, most firms brought WhatsApp inside. Archiving went in, capture was wired into the SIEM, DLP rules were written. Today, in most large regulated businesses, the apparatus genuinely exists and people genuinely maintain it.

Here is where it gets interesting, and where Kashif's experience is the useful part of this piece.

Some years ago he was close to a communications monitoring programme at a large energy company. Serious work, well resourced. WhatsApp had been brought inside the perimeter, capture was in place, messages were flowing into the SIEM where they could be searched, retained and alerted on. By every measure anyone was looking at, the channel was covered.

Then someone went looking for a specific exchange, and it was not there. Not deleted. Never captured.

The text in that thread had arrived intact. The voice notes had not. People had been holding entire conversations by pressing and holding a microphone button, and none of it had ever reached the SIEM. No error, no failed job, no gap in the message counts. The pipeline had done exactly what it was built to do, which was to read text.

Nobody had bypassed the control. The control had a shape, the product moved, and the shape stopped matching.

Why voice notes slip through

A voice note is not a message with sound in it. Structurally it is an event plus an encrypted binary attachment, usually Opus-encoded, that has to be fetched, decrypted, decoded and then handed to something that can understand audio. Text needs none of those steps.

That produces three failure modes, which need different fixes.

1. The connector never receives the audio. Plenty of integration layers pass text and images downstream but do not pass audio messages at all. The event surfaces with an empty attachment field.
2. The connector receives it and cannot use it. The audio lands as an opaque blob. It gets stored, or worse it gets counted, but it is not searchable and not readable by surveillance rules.
3. Nothing downstream can read it even if it arrives. Lexicon surveillance, DLP and SIEM correlation all operate on text. An audio file with no transcript is inert. It satisfies retention while defeating supervision.

All three produce the same symptom, and it is the dangerous one: the system reports success. Partial capture is worse than no capture. No capture is a known gap that gets managed. Partial capture is an unknown gap that gets trusted.

That last row is the whole problem in one line. The metric that tells you the control is working is the same metric that hides the fact that it is not.

Voice notes are just the one that got found

The reason to tell that story is not that audio is uniquely hard. It is that the same thing keeps happening, on a schedule set by someone else's product team.

WhatsApp is a consumer product shipping features continuously to three billion people, and its roadmap has never once considered your surveillance stack. A capture project, by contrast, produces a sign-off: on this date, against the app as it then existed, messages were captured correctly. That certificate starts decaying immediately.

Message editing deserves a moment, because it is subtle where voice notes are blunt. A sender can revise a message within fifteen minutes. Recipients see an "Edited" label but not the previous text. If your archive stores the current state of a message rather than every version, you are holding a record that is complete, searchable, entirely compliant looking, and different from what was actually sent.

Ask who in your organisation is accountable for re-testing capture when WhatsApp ships a feature. In most places the honest answer is nobody. Vendor management owns the contract, compliance owns the policy, security owns the SIEM, and the space between "the app changed" and "our capture still works" belongs to no one on the org chart.

Now do the same story with AI

Read back the three reasons the WhatsApp policy failed and change the noun.

The sanctioned tool is worse than the one people can reach. The approved AI is locked down, waiting on a review, or does not connect to the systems where the work is, while a personal account is one tab away and genuinely excellent.

It goes all the way up. The people using AI most aggressively are frequently the most senior and most capable, the same population the SEC found on WhatsApp.

Nobody can see it. There is no dashboard for work done in a personal AI account, because the whole point is that it does not touch your systems.

Same three conditions, and this time the tool does not just carry the conversation. It reads documents, holds credentials, and acts on systems. The banks had four years between the first fine and the last, and they still could not close it.

There is even a direct handoff between the two problems. Somebody who receives a two minute voice note and does not want to listen to it now transcribes it, using whatever tool is nearest, which is often a personal AI account. Regulated communications your pipeline could not read get uploaded into a system your organisation does not control, has no contract with, and cannot audit. The content did not merely escape the archive. It left the building.

What to actually do

1. Test the pipeline, not the dashboard. Send one of each message type through a monitored account and retrieve the content from the archive. Not the metadata row, the actual audio, image or document. Assume nothing you have not personally retrieved.
2. Record coverage per message type. "WhatsApp is covered" is not a coverage statement. Coverage is per message type per channel, because that is the granularity at which it fails.
3. Give release notes an owner. Somebody should read them and ask one question: does this create a new message type, and can we capture it?
4. Make the sanctioned path the fast one. This is the only lesson that generalises properly. Every one of these fines traces back to a compliant route that was slower than the alternative. If your approved AI is harder to use than a personal account, you are running the 2021 playbook and you already know how it ends.
5. Find out what is actually being used. Before governing anything, discover what is already connected to your systems. You cannot supervise what you have never seen, which is the entire $3.5bn lesson in one sentence.

The short version

Banks were not fined for using WhatsApp. They were fined for being unable to say what was said on it.

The record is the control. Everything else, the policy, the ban, the attestation, is a statement of intent that turns out to be unverifiable at exactly the moment somebody asks.

That is the problem Connor works on, one technology earlier this time: find the AI already connected to your systems, give people sanctioned Capabilities that are better than the workaround rather than worse, and keep the record of what was actually done.

If you want to know where your own gap is, start with step one. Send yourself a voice note, then go and try to find it.