Data loss prevention grew up around the file. Classify the document, watch the endpoints, block the upload, log the attempt. That model works because a file holds still long enough to be inspected. The material most likely to cause damage now rarely holds still, and often does not look like a file at all.
The file is no longer where the sensitive part lives
A contract in a document repository is protected reasonably well. The negotiating position that preceded it, typed into a chat window at eleven at night, is not. Neither is the client instruction relayed as a voice note, the credential pasted into a thread and deleted an hour later, or the screenshot carrying the same figures as the spreadsheet without any of its labels.
Content inspection built around card number patterns and document fingerprints finds none of it.
Financial services was forced to think in these terms earlier than most sectors. Firms working through MiFID II compliance requirements discovered that a conversation could itself be the regulated artefact rather than the wrapper around one, and that the obligation attaches to what was said as much as to whatever was eventually filed.
What a conversation-aware control has to see
The word conversation is doing real work here, and it is worth being precise about it. A conversation is not a channel and not a transcript. It is a bounded exchange with participants who join and leave, a subject that drifts, and an outcome that may never be stated explicitly anywhere inside it. Controls that treat the channel as the unit of analysis end up monitoring a firehose.
Controls that treat the individual message as the unit end up reading fragments.
Understanding a conversation means holding context that file-centric tooling deliberately discards. Who is in the thread, including anyone added late. What was said before the risky line, since the risk usually lives in the exchange rather than any single message. Whether an attachment was later edited or withdrawn. Whether the channel is one the organisation can actually produce records from. A control that evaluates each message independently raises alerts on the wrong things and misses the pattern entirely.
The adjacent market has already solved parts of this for unrelated reasons. Platforms built around turning talk into insight analyse sentiment, intent and outcome across whole interactions rather than isolated utterances, which is precisely the analytical shape a conversational control needs, pointed at a different question.
Regulators drew the boundary years ago
The UK rules on recording the calls that relate to client order services make the point with unusual precision. They cover reception, transmission and execution of client orders, and explicitly include communications intended to result in a transaction even where none ultimately occurs. Records must be kept for at least six months from creation. Internal conversations sit outside the requirement, and analysts, retail advisers and back office staff are generally not expected to be recorded when acting in those capacities.
That carve out is instructive rather than reassuring. The boundary between covered and uncovered is drawn by role and purpose, not by channel or file type. A control that can only reason about where a message was sent, and never about what it was for, cannot enforce a boundary of that shape without either over-collecting or missing the material that counts.
The enforcement record is a channel problem
Penalties have followed the channel rather than the content. Orders published in March 2024 imposed civil monetary penalties of six million and one million dollars on two firms over business communications sent through unapproved methods including personal text, records that went unpreserved and that neither firm could have produced promptly if asked.
The same releases note that since December 2021 more than a billion dollars in penalties had been imposed on twenty two financial institutions for the same category of failure, and that some of the supervisory staff responsible for enforcing the policy had used unapproved channels themselves.
No classifier was defeated in any of those cases. The material simply travelled somewhere the controls did not reach. That is the awkward truth about conversational risk: the dominant failure mode is not a clever exfiltration technique but a sanctioned employee choosing a faster application. The motive is almost always convenience rather than concealment, which is why policy alone rarely fixes it. People route around friction, and an approved channel that is slower than the alternative will lose that contest quietly and continuously until somebody measures where the traffic actually went.
Detection has to move upstream of the message
Which suggests the useful investment is coverage before cleverness. A mediocre control on every channel people actually use beats an excellent control on the two channels the policy prefers. In practice that means capturing from the messaging platforms already in daily use rather than mandating a platform nobody wants, and treating an unmonitored channel as a finding rather than an accepted exception.
It also changes what a good alert looks like. File-based prevention produces a verdict on an object. Conversational analysis produces a judgement about an exchange, which is inherently probabilistic and needs a reviewer attached to it. Teams that deploy it expecting a clean block or allow decision drown in false positives within a fortnight. Teams that deploy it as triage, feeding a human queue with ranked context, get value from the same signal within weeks rather than quarters.
Watch the exchange, not the attachment
The reframing is uncomfortable because it makes the control less definitive. A file either matched a policy or it did not. A conversation is a moving argument between several people, and any honest assessment of it arrives with a confidence level rather than a verdict. The alternative, though, is a programme that guards the filing cabinet with real rigour while the decisions capable of embarrassing the organisation are made two applications away, in a thread nobody is reading and nobody has kept.
