Frameworks Group
CLARITY
THE CRISIS
Customer-Facing AI Has Entered A Production Governance Crisis.
Enterprises are deploying AI agents into live customer interactions. The agents are performing. The outcomes are not always what was declared. In most enterprises, no one can show exactly what the agent did, why it did it, or whether it met the standard they said it would.Seventy-four percent of enterprises have already rolled back or shut down a deployed AI customer communications agent because of a governance failure. Among those with the most mature governance frameworks, that figure rises to 81%. More than forty percent of agentic AI projects are forecast to be cancelled by the end of 2027 because of inadequate risk controls.The most governed enterprises are failing at the highest rate. Technology is not the problem. The governance of what the technology does at the customer boundary is.
Sinch, The AI Production Paradox, 2026. Gartner, June 2025.
WHY IT EXIST
The current production-control stack is built from the inside out.
It governs permissions, memory, orchestration, runtime execution, and technical performance. It does this well. What it does not do is answer the question the enterprise is actually accountable for: is the agent conducting itself, in real conversations with real customers, the way the enterprise declared it would?The question cannot be answered from inside the stack. It requires a different starting point: the conversation itself, read after it happened, assessed against a defined standard, by a party with no stake in the answer. No layer in the current stack provides that. Not because the builders failed.Because independence cannot be built in. It has to come from outside.
This is the Conduct Interface: the boundary where AI behavior meets human reality and becomes accountable enterprise conduct. It is where technical outputs become customer experiences, where performance becomes responsibility, where the enterprise's declared standard either holds or it doesn't.Every enterprise deploying a customer-facing AI agent is accountable for what happens at the Conduct Interface. Almost none can currently demonstrate what happened there.
The Governance Loop
Governing the Conduct Interface requires a defined sequence: define the conduct standard, observe real-use behavior, diagnose meaningful deviation, remediate what must change, verify that conformance has been restored and remains reliable.The loop cannot be run from inside the stack. An operator assessing its own agent's conduct is not an independent assessment. It is self-reporting.CLARITY runs this loop as a genuinely independent third party. It reads real customer conversations after they happen, assesses conduct against the standard the enterprise itself declared, and identifies where the agent's behavior departed from that standard, traceable to a specific clause, not to a vendor's taxonomy or a platform's own judgment.It flags. It never acts. The enterprise governs. CLARITY makes that governance possible.
Frameworks Group
Frameworks Group was built on one finding: the signals that determine whether a customer conversation resolves, escalates, or quietly fails are consistent, measurable, and present inside the conversation itself, before any outcome is recorded.That finding came from frontline human operations. Contact centres, service environments, anywhere people were under pressure to resolve another person's problem quickly and well. The behavioral signals were identified, validated, and applied across industries over two decades.When we examined AI agent transcripts, the same signals were present, without modification, without recalibration. The gap between what enterprises believe their agents are doing and what the conversations actually contain is significant. CLARITY was built to close it.Frameworks Group is headquartered in Singapore. CLARITY is the firm's primary instrument.
The Origin
I never went down the psychology road. I went down the road of observation.I believe that specific patterns trigger specific responses. Not because of what either party is thinking or feeling - thoughts are invisible, and motivations are irrelevant to the person on the other side of the conversation. What matters is the behavior pattern produced in response to what the customer says and does. That is all the customer ever experiences.A film director knows exactly how to get an actor to produce the right behavior: the pace, the tone, the pause, the stillness that creates a specific response in the audience. The director works entirely on the surface. On behavior. Not on internal states. That principle became the foundation of everything I built.The biggest challenge was measurement. You cannot improve what you cannot measure. For years, that was the obstacle -pattern recognition that was instinctive but not systematic. So I began mapping the precise behavioral moments that triggered a response in the customer, in the agent, and in the direction the conversation then took. Over time that became a behavioral coding system: a framework capable of identifying the signals that predict whether a conversation will escalate, resolve, or quietly fail - before any outcome is visible. Validated across multiple industries. Consistent. Repeatable. Measurable.The structural parallel arrived without warning. The transcripts an AI agent produces in a live customer conversation are identical in form to the ones I had been coding for years. I applied the behavioral coding system directly. The outcome was decisive.The exact same signals that predicted escalation in frontline human conversations were present in AI agent transcripts. Without modification. Without recalibration. On the first run.The reason is not mysterious. AI agents are trained on conversational data. They absorb behavioral patterns - including the ones that cause conversations to fail. They replicate behavioral failure modes not because they are malfunctioning, but because they were built from language produced by people who were not behavioral experts.What made the AI finding significant beyond anything in the human context was this: in frontline environments, remediating a behavioral failure pattern is slow, expensive, and constrained by what a person can learn and retain. An AI agent has no such limitation. The pattern causing the failure can be identified precisely, amended directly, and verified immediately. No learning curve. No resistance. No forgetting.
The constraint that made behavioral governance difficult to scale disappears entirely when the agent is an AI.Rob O'Connell, Co-Founder, Frameworks Group
The Conversation
The production governance crisis is not evidence that AI agents cannot work. It is evidence that governing them requires a capability the current stack was not designed to provide.If you deploy customer-facing AI agents and want to understand what is happening at the Conduct Interface, from the evidence of the conversations themselves, we would welcome the conversation.