Before building Compiled, we spent several months in structured conversations with more than fifty CISOs and chief compliance officers — across global banks, regional broker-dealers, hospital systems, and high-growth technology companies. We asked them where their programs were most exposed, what they had tried, and why the gap persisted. What follows is a synthesis of those conversations: the patterns that appeared consistently enough to constitute findings, and what they told us about the state of behavioral risk management in enterprise organizations.
The research was qualitative, not a survey with statistically representative sampling. We recruited participants through professional networks and inbound interest, which means the population skews toward leaders who were already thinking carefully about the problem. The findings should be read as a portrait of the concerns that thoughtful security and compliance leaders carry — not as a census of the industry.
The most striking finding from these conversations was not a revelation about a hidden problem. It was that the problem is known, accepted, and in many cases precisely quantified by the people who own it. Every CISO we spoke to could tell us, with reasonable accuracy, what percentage of their organization's communications traffic was actually inspected. The numbers ranged widely — from roughly ten percent at the lower end to around forty percent at the upper — but none claimed total coverage. The remainder was described, consistently, as a calculated risk.
“We know what we're not seeing. We accept it because the cost of seeing all of it doesn't fit the budget. That is exactly the kind of answer that doesn't survive a regulator asking the question.”
This acceptance of the gap was not a sign of negligence. It was the rational response to the cost structure of the tools available. The inspection methods capable of detecting subtle, novel, or context-dependent risk were, in practice, too expensive to run on the full communication volume of a large organization. The inspection methods cheap enough to run on everything were not capable of detecting what mattered. Leaders had internalized this tradeoff and calibrated their programs around it. The coverage gap was not a failure of the program; it was the program.
When we asked participants to characterize the risk they were most concerned about that their current program was least capable of detecting, the answers clustered consistently around a category we came to call behavioral risk: conduct that is wrong not because it matches a prohibited pattern but because of what it implies about intent, context, or relationship.
Insider trading is the canonical example. The communication that precedes an insider trade does not typically say “I have material non-public information and I am going to trade on it.” It says something that, in context — considering who is talking to whom, about what, at what time relative to a corporate event — implies that. A keyword filter does not have context. A statistical anomaly detector sees the pattern only after the corpus is large enough to establish a baseline. A human reviewer can recognize the implication, but human review cannot scale to total coverage at human cost.
The same pattern applied to a broad range of concerns: coordinated market manipulation, front-running, implicit investment recommendations, data exfiltration that does not use recognized sensitive-data patterns, and conduct that violates the firm's policies without violating any external rule. In every case, the concern was the same: the behavior is real, the risk is quantifiable, and the tools available to detect it were either insufficiently capable or impractical to deploy at scale.
We asked every participant the same question about detection latency: if a violation occurred right now, in what timeframe would your program detect it? The answers were illuminating. For keyword-based violations — communications that contained explicitly prohibited content — detection was typically described as occurring within a day, during the next batch-processing cycle. For behavioral violations — the kind that require context to interpret — detection was described as occurring in days, weeks, or not at all unless flagged through another channel.
None of the participants described anything approaching real-time behavioral detection as a current capability. Most had not thought to want it, because the tools available had not offered it and the use cases that would benefit most from it — agentic actions, real-time trading communications, live customer interactions — had not yet been fully incorporated into the surveillance perimeter.
When we described real-time inline inspection — a finding rendered before the communication was delivered or the action completed, not in the next batch cycle — the response was consistently similar. Participants recognized immediately that this changed the value proposition of surveillance from a retrospective reporting function to an active governance mechanism. The CISO at a mid-size financial services firm put it directly: “If you can stop it before it happens, you have a control plane. If you can only find it after, you have an audit trail. Those are very different things.”
A significant proportion of participants — particularly those at institutions that had moved quickly on enterprise AI adoption — described a version of the same scenario: AI agents deployed in production roles whose outputs were not being reviewed by the communications surveillance program. The agents were generating customer communications, summarizing research, drafting responses to regulatory inquiries, and in some cases executing operational steps with downstream consequences. In most cases, these outputs were outside the surveillance perimeter because the perimeter had been designed for human communications and had not been extended to cover agent outputs.
The compliance implication was not lost on participants. An AI agent generating an implicit investment recommendation in a customer interaction has the same regulatory exposure as a human representative doing the same. The fact that the output came from an agent does not reduce the obligation; in some readings, it increases the firm's responsibility because the firm chose to deploy the agent. Several participants described this as a gap they had identified but had not yet resolved — waiting for clearer regulatory guidance, for technical solutions, or simply for the issue to be escalated to a level where resource allocation could be considered.
The picture that emerges from these conversations is not one of ignorance about the problem. The security and compliance leaders we spoke to were, nearly without exception, intelligent, experienced, and clear-eyed about where their programs were exposed. The persistence of the behavioral blind spot was not a failure of awareness. It was a failure of available tooling.
The tools capable of detecting behavioral risk required either unacceptable compromises on data residency, or cost structures that made total coverage impossible, or latency profiles that produced retrospective reporting rather than real-time governance. Most programs had made the best of these constraints. They sampled. They focused keyword detection on the highest-risk populations. They accepted the gap in the middle and hoped that the violations that mattered most would surface through other channels — audits, whistleblowers, regulatory inquiries — before the damage was too large.
The behavioral blind spot is not inevitable. It is the residue of a generation of tools that could not simultaneously provide semantic precision, real-time operation, total coverage, and data residency. When those constraints are removed — when the inspection can be fast enough to be inline, capable enough to recognize behavioral patterns, economical enough to run on every message, and architecturally suited to run inside the organization's own environment — the gap closes. The CISOs we spoke to did not describe this as a nice-to-have. They described it as what they had been waiting for.
Compiled was built on exactly these conversations. If you recognize your program in what you read above, we would like to talk.