9
Every firm we walk into has an AI policy. Most of them are thoughtful documents, carefully drafted, approved by leadership, and distributed with a memo. And in almost every firm, the policy is failing. Not because it is badly written, but because a policy was never designed to do the job it has been asked to do.
The gap between what a policy says and what staff actually do is the single most common finding in our assessments. A partner pastes a client agreement into a public tool to summarize it. An associate drafts a memo with a consumer assistant and never checks the settings. A paralegal uses a free transcription service on a deposition. None of these people are careless. They are solving real problems with the tools that are easiest to reach, and the policy did not give them a better option.
This is shadow AI, and it is not a compliance failure so much as a design failure. When the approved path is slower, harder, or less useful than the unapproved one, staff will take the unapproved one every time. No amount of training or warning changes that arithmetic. People do not violate policy out of malice; they violate it because the policy asks them to choose inconvenience, and convenience wins.
The deeper problem is that most policies are written as if the risk lives in the document. It does not. The risk lives in the architecture. A policy that says “do not send client data to external tools” is meaningless if the firm’s own systems make it effortless to do exactly that. The policy is a statement of intent; the architecture is the enforcement. Where they disagree, the architecture wins.
Enforcement is the word most policies are missing. A real control is not a rule on paper. It is a technical boundary that cannot be crossed, or a workflow that makes the compliant path the natural path. It is logging that shows what actually happened, not a signature page that shows what was supposed to happen. It is a review process that catches the exception before it becomes an incident.
This is why policy without architecture fails. You can write the perfect rule, but if the firm’s data is scattered across personal drives, shared folders, and unmanaged cloud accounts, no rule can govern it. You can mandate training, but if the approved tools are inferior, training just teaches people to hide their workarounds. You can threaten consequences, but consequences discovered after the fact are not prevention.
The firms that get this right do not start with the policy. They start with the data. They map where confidential information lives, they bring the AI to the data instead of sending the data to the AI, and they build enforcement into the systems people actually use. Only then does the policy become something other than a hope.
If your policy is failing, the answer is not a better policy. It is the architecture that makes the policy true. That is the work we do, and it begins with a question most firms have never answered: where does your data actually live?
Start with the Confidential AI Exposure Checklist at 9tailsadvisory.com/checklist/. Nine lives. Zero leaks.