This article is published by J-10, Jalubro's proprietary governance enforcement platform. It is part of a series exploring how regulated enterprises can enforce compliance inside operational workflows. To learn how Jalubro's advisory and implementation services support governed enterprise operations, visit our services page.

Most enterprises have an AI policy. Almost none enforce it.

The conversation about AI governance in regulated enterprises has been dominated by principles. Responsible AI frameworks. Ethics boards. Acceptable use policies published on intranets. Internal guidance documents circulated by email.

None of this governs anything.

A governance framework that exists in a document but is not enforced inside the workflows where AI is actually used is not governance. It is aspiration. And the gap between aspiration and enforcement is where regulated enterprises are accumulating risk at a pace most leadership teams have not yet fully understood.

This article sets out a practical framework for AI governance that works in operational reality, not on paper. It covers both sides of the problem: what goes into AI tools and what comes out of them. And it is designed for the enterprises that cannot afford to get this wrong: financial services, energy, healthcare, insurance, government and any organisation operating under regulatory scrutiny.

The two-sided risk most enterprises are not governing

AI governance is usually framed as an output problem. The concern is that AI tools will generate something inaccurate, non-compliant or harmful, and that this output will flow into a business decision without being checked.

That is a real risk. But it is only half the problem.

The other half is the input side. Every time a user interacts with an AI tool, they provide it with data. In a regulated enterprise, that data might include privileged legal advice, client-confidential information, draft settlement terms, commercially sensitive pricing, employee personal data, board strategy documents or regulatory filing content.

Most enterprises have no controls over what data their people feed into AI tools. No classification checks. No sensitivity gates. No policy enforcement at the point of input. The acceptable use policy says "do not share privileged or confidential information with AI tools." The operational reality is that there is nothing preventing it from happening, and no way to know when it does.

A complete AI governance framework must govern both directions: what enters the AI tool and what comes out of it before either side creates risk for the enterprise.

Why existing tools do not solve this

Enterprises typically look to three categories of existing technology to address AI governance. None of them are designed for it.

GRC platforms (Archer, ServiceNow GRC, MetricStream) manage risk registers, control libraries and audit workflows. They can document an AI risk. They cannot enforce an AI governance policy inside a legal workflow, a procurement process or a contract authoring tool. They operate alongside operational systems, not inside them.

Data loss prevention (DLP) tools can detect and block sensitive data leaving the enterprise through defined channels. But AI tools used within the enterprise, CoCounsel inside Legal Tracker, Harvey accessed through a browser, Copilot inside Microsoft 365, operate within the trusted perimeter. DLP was not designed to govern what a user pastes into an AI prompt inside an approved enterprise application.

The AI tools themselves offer some built-in controls. CoCounsel operates within the Thomson Reuters environment with its own security model. Harvey has enterprise deployment options with data handling controls. But neither tool governs whether the specific data a user provides is appropriate under the enterprise's own policies, or whether the specific output the tool generates complies with the enterprise's governance framework before it enters a downstream workflow. The AI vendor secures the tool. The enterprise must govern how it is used.

The gap is structural. No existing category of technology was built to enforce governance policies at the point where humans interact with AI inside enterprise workflows, on both the input and output side.

A practical framework for enterprise AI governance

What follows is a framework designed for operational deployment, not for a slide deck. It has four layers, each of which must be present for AI governance to be enforceable rather than theoretical.

Layer 1: Input governance

Input governance controls what data enters AI tools. It operates at the point of interaction, before the data reaches the AI.

Data classification enforcement. Every document, every data field and every piece of content that a user attempts to provide to an AI tool is validated against the enterprise's data classification policy. Privileged legal advice classified as "Restricted" is blocked. Client-confidential financial data is blocked. Board strategy documents are blocked. The classification is not advisory. It is enforced.

Context-aware sensitivity rules. Not all data of the same classification should be

treated identically. A draft contract marked "Confidential" might be appropriate for AI-assisted clause review within a governed CLM workflow, but inappropriate for a general-purpose AI prompt. Input governance must account for the context of the interaction, not just the classification of the data.

User-level permissions. Different roles have different risk profiles. A senior associate with matter-specific authorisation may be permitted to use AI for contract analysis on that matter. A paralegal without that authorisation should not. Input governance must enforce role-based access, not just data-based classification.

Audit trail on all inputs. Every interaction where data is provided to an AI tool must be logged. What was submitted, by whom, when, under what policy authority, and whether it was permitted or blocked. This evidence must be captured automatically, not reconstructed after the fact.

Layer 2: Output governance

Output governance controls what the AI produces and whether it is permitted to enter downstream workflows and decisions.

Policy validation. Every AI-generated output (a clause, a summary, a recommendation, a risk assessment) is validated against the enterprise's governance policies before it progresses. A contract clause generated by CoCounsel is checked against the approved clause library. A risk summary generated by Harvey is validated against the enterprise's risk classification framework. Non-compliant outputs are flagged, blocked or routed for human review.

Accuracy and consistency checks. AI outputs are compared against source materials, approved templates and existing enterprise data to identify hallucinations, contradictions or deviations. This is particularly critical in legal environments where an AI-generated citation, clause or summary that is factually incorrect can create significant liability.

Downstream flow control. An AI output that passes validation can enter downstream workflows automatically. An output that fails validation is held. It does not flow into a contract, a procurement decision, a financial report or a client communication until the governance issue is resolved. The AI tool does not get the final say. The governance layer does.

Audit trail on all outputs. Every output, its validation result, and its onward journey through enterprise workflows must be evidenced. If an AI-generated clause ends up in an executed contract, the enterprise must be able to trace the full chain: the AI tool that generated it, the policy it was validated against, the result of that validation, who approved it, and when.

Layer 3: Cross-system enforcement

AI tools do not operate in isolation. CoCounsel sits inside the Thomson Reuters legal technology stack. Harvey connects to document management systems and contract repositories. Copilot operates across Microsoft 365. Enterprise AI agents interact with ERPs, procurement platforms and financial systems.

AI governance must travel with the data across systems. A governance policy that applies when a user interacts with CoCounsel must also apply when the output of that interaction flows into Legal Tracker, then into a procurement workflow, then into a financial commitment. If governance only exists within the AI tool's boundary, it disappears the moment the output moves downstream.

Cross-system enforcement means that the governance layer is not embedded inside any single tool. It sits across the enterprise technology stack and enforces policy wherever data flows, regardless of which system originated it or which system receives it.

Layer 4: Continuous evidence and reporting

Regulated enterprises must be able to demonstrate their AI governance to regulators, auditors and boards. This requires more than a policy document and an annual review.

Continuous evidence capture. Every governed AI interaction, input and output, produces an evidence record automatically. No manual logging. No retrospective assembly.

Real-time governance dashboards. Leadership can see, at any point, how AI is being used across the enterprise: how many interactions, how many blocked inputs, how many flagged outputs, which policies are being triggered most frequently, where exceptions are occurring.

Regulatory reporting readiness. When a regulator asks "how do you govern AI use in your organisation?", the enterprise can produce timestamped, tamper-proof evidence of every governed interaction, not a policy document and a promise.

Applying the framework: what this looks like operationally

Consider a global financial services firm with legal teams using CoCounsel for contract review, compliance teams using Harvey for regulatory research, and procurement teams using an AI-powered supplier risk tool.

On the input side, a lawyer attempts to paste a privileged board memorandum into CoCounsel for summarisation. The document is classified as "Board Privileged" under the firm's data classification policy. The governance layer blocks the input, logs the attempt and notifies the lawyer that the document's classification does not permit AI processing. The lawyer can request an exception, which is routed to the appropriate approver with a full audit trail.

On the output side, Harvey generates a regulatory summary for the compliance team. The governance layer validates the summary against the firm's regulatory framework and identifies that one of the cited regulations has been superseded. The output is flagged before it reaches the compliance report. A compliance analyst reviews the flag, confirms the issue, and requests a corrected output. The entire chain is evidenced.

Across systems, CoCounsel generates a limitation of liability clause for a vendor contract. The clause passes output governance checks against the approved clause library. It enters the CLM workflow, where it is incorporated into a contract that, once executed, triggers a procurement commitment. The governance controls that applied to the AI-generated clause continue to apply as the data flows from the CLM into the procurement system and into financial reporting. The governance layer does not stop at the boundary of the AI tool.

This is not theoretical. This is what governed AI looks like in a regulated enterprise.

Why this cannot wait

The window for getting AI governance right is closing. Three pressures are converging.

Adoption is accelerating. Legal, compliance, finance and procurement teams across regulated industries are adopting AI tools now. Every month without enforceable governance is a month of ungoverned AI interactions accumulating risk.

Regulation is catching up. The EU AI Act, FCA expectations on AI use in financial services, sector-specific guidance from regulators across healthcare, energy and insurance. The regulatory direction is clear: enterprises will be expected to demonstrate enforceable AI governance, not just policy documentation.

The evidence gap is growing. Every ungoverned AI interaction is an interaction that cannot be evidenced. When the regulator or the auditor asks for evidence of how AI was governed across the enterprise for the past twelve months, the enterprise either has it or it does not. Retrospective reconstruction is not credible.

How J-10 delivers this framework

J-10 is a business-side governance enforcement platform that implements all four layers of this framework across your existing enterprise technology stack.

It governs AI inputs by enforcing data classification and sensitivity policies at the point of interaction, before data reaches the AI tool. It governs AI outputs by validating what AI produces against your governance policies before outputs enter downstream workflows. It enforces governance across systems, so that a policy applied to an AI-generated clause in your CLM continues to apply when that clause creates a procurement commitment or a financial exposure. And it captures continuous, audit-grade evidence of every governed interaction automatically.

J-10 works with CoCounsel, Harvey, Copilot and enterprise AI agents. It does not replace your AI tools. It governs how they are used inside your business.

To learn more about how J-10 governs AI inputs and outputs across enterprise workflows, visit j-10.ai or contact the Jalubro team to book a briefing.