The First AI Governance Move Is Not a Policy

Why a Policy Alone Is Not Enough

Most regulated firms have responded to AI by writing a policy. The instinct was a reasonable one. New technology arrived, staff started experimenting, vendors started bolting features onto existing products, boards asked what the firm was doing about all this, and the natural compliance answer was to put a document around it.

In a Jersey financial services business, leaving tool use uncontrolled is not a small matter. A trust company that allows client files to be pasted into public systems without knowing where the data goes has a confidentiality problem before it has an AI problem. A law firm that lets a system draft client advice without proper review is exposed in ways that have nothing to do with novelty.

So the policy arrives, and it usually says sensible things. Do not put confidential information into public AI tools. Do not use AI to give advice without review. Escalate doubts to compliance or IT. The document may pick up data protection, information security, record keeping and professional duties, and it will probably be approved by the board and circulated to staff.

There is nothing wrong with the policy. The difficulty is that it answers the visible part of the question, which is the smaller part, and tends to leave the larger part exactly where it was.

A policy of this kind tends to govern the AI that everyone can name, ChatGPT, Copilot, Claude, Gemini. That is the easiest surface to govern because it is the most obvious. The harder AI is already inside the firm under other names: vendor tooling, legacy automation, and informal staff practice, where what policy prohibits and what daily workload tolerates have quietly diverged.

A firm that writes an AI policy before mapping those three places has produced something. It has not produced an AI governance framework, though. It has produced evidence that the subject has reached the agenda, a useful step, but not one that gets you very far on its own in regulatory and operational terms.

Layer 1: Vendor Tooling - Where Most Unmapped AI Lives

Begin with vendor tooling, because that is where most of the unmapped AI in a regulated firm now lives. Firms buy systems to solve operational problems: client relationship management, entity administration, screening, transaction monitoring, document management, board portals, file review, knowledge management. Those systems are now being updated, quietly in many cases, with AI features.

Sometimes the feature is labelled clearly. More often it appears in a release note as "smart search", "risk scoring", "anomaly detection" or "assistant", with no particular flag that anything material about the firm's exposure has changed. The firm may not think it has adopted AI at all. It may simply think it has renewed a licence and clicked through a change log.

The distinction matters more than it sounds. A screening tool that subtly changes the way alerts are ranked moves a compliance team's working risk population without anyone noticing that operational judgment has shifted. An onboarding platform that flags risk indicators using machine-generated logic is, in practical effect, a participant in customer due diligence, whether or not anyone has classified it that way.

Jersey firms already have a language for this, even where they do not yet call it AI governance. The JFSC's Outsourcing Policy states that a business remains responsible and accountable to the JFSC for outsourced activity, and that the governing body cannot delegate its regulatory responsibilities or accountability to a service provider. The same policy expects suitable due diligence on service providers, attention to material risks, and ongoing monitoring of performance. When an AI feature appears inside a vendor product, it remains the firm's governance problem, regardless of where it sits in the licensing stack.

When a firm has visibility of its vendor AI, negotiations also change. It can ask direct questions: what AI features are in the product, where data is processed, whether client data trains models, whether features can be disabled, how incidents are reported. These are procurement and outsourcing questions, of the kind the JFSC's policy already expects firms using cloud services to put to providers alongside data storage location, information security, audit and access rights, business continuity and exit arrangements. AI sharpens that line of enquiry.

Layer 2: Legacy Automation That Predates the Conversation

Legacy automation is the next layer down. Many regulated firms have systems that have been making consequential judgments for years without anyone calling them artificial intelligence. They may not use machine learning in any modern sense, they may be rules engines, scoring models, screening thresholds, spreadsheet macros.

Some were designed by staff who left the firm long ago. The documentation is thin, and the original business logic has drifted from current practice without anyone revisiting the assumptions. Some still sit inside Excel files that everyone depends on and that nobody really wants to touch.

These systems usually fall outside any AI policy because they predate the current conversation, which is the wrong test. The governance question is whether a system shapes a decision that matters. Consider two examples:

• A transaction monitoring rule, set five years ago to suppress alerts below a particular threshold, may now be shaping AML outcomes against a client base and risk profile that no longer resemble the ones the rule was calibrated for.
• An onboarding workflow that routes "low risk" files away from senior review, because that was efficient when the book was simpler, may have become a governance decision in its own right, one the board has never reconsidered.

The JFSC has made a similar observation in adjacent language. Its feedback on virtual asset service providers' suspicious activity reporting called for documented understanding of monitoring tools, including coverage, limitations, external data sources, how the system identifies unusual activity, and how outputs and alerts are analysed. The underlying observation carries straight across: you cannot supervise a tool you have not described.

Layer 3: Informal Staff Use - The Most Uncomfortable Layer

The third layer, informal staff use, is the most uncomfortable of the three because everyone knows it exists. Staff are under pressure. Client expectations keep rising. Margins are tight. The administrative load on regulated work is heavy. Younger staff in particular have grown up alongside tools that produce text, summaries, translations, explanations and first drafts on demand, and they will use those tools where the firm gives them a safe route. They will also use them where the firm forbids it, if the workload remains and the prohibition feels detached from how the work actually gets done.

A ban that is clean on paper can be porous in practice, and the result is often worse than properly supervised use would have been. Staff strip out names but leave enough context to identify a client. They use a public tool to rewrite a suspicious activity narrative, summarise a tax memo, or sketch a first draft of board minutes. Sometimes the result is plausibly worded and quietly wrong, or stripped of the qualifications that mattered. The firm does not see most of this because the policy has pushed it outside the visible system.

This is also where the opportunity sits. A firm that has a realistic picture of what its people are actually doing can permit broader AI use rather than narrower, because it knows which uses are harmless and which carry real exposure. It can:

• Approve internal tools for summarisation, formatting, translation, drafting and research support
• Prohibit public-tool use on confidential client material without pretending that staff have no need for assistance
• Train people to treat AI as a supervised drafting aid, rather than leaving them to develop the habit as a private shortcut they keep off the record

Supervised use, in nearly every realistic scenario, carries less risk than prohibited use that quietly continues.

The Jersey Office of the Information Commissioner has identified artificial intelligence as one of its 2026 to 2028 strategic priorities, with a particular focus on AI systems in human resources and on privacy risks including misuse of personal information, excessive surveillance, data security and insufficient transparency. The pragmatic response to that environment is to make AI use visible and to govern it deliberately.

The Visibility-First Approach to AI Governance

The useful re-ordering is to begin with visibility, build a framework on that visibility, and write the policy as an expression of the framework, rather than the other way around. Many firms begin with the policy and try to push practice into the document; that sequence rarely holds for long.

Visibility, in this sense, means knowing where AI and AI-like decision support actually sit in the firm. Not only the approved GenAI tools, and not only the frontier models. The full map:

• Vendor products with AI features and their outsourcing implications
• Legacy automation that shapes decisions and the business logic behind it
• Staff use in real work, the decision points each one touches, and who owns each one
• Security controls, failure modes, and contractual protections

The exercise can be made to feel like bureaucracy if it is done badly, but properly handled it is management information. A firm that has it is in a different operational position from one that does not.

With a framework in place, most questions are routed by type, sensitivity and consequence, and the firm spends less time arguing about whether AI is allowed at all. Without that framework, every AI question becomes an individual judgment call, and the compliance burden grows precisely because nothing is settled.

What This Means in Practice

The practical implications differ by business line, although the pattern is the same.

• Trust companies: AI can help summarise correspondence chains, extract deadlines, and improve board reporting. The question is when the tool is assisting administration and when it is starting to influence a fiduciary decision.
• Law firms: Drafting, chronology and research support are all useful, provided the tool does not become an unrecorded source of judgment.

A firm that reduces the time spent finding, formatting, summarising and checking routine information can move faster without lowering its standard of review. A firm that gives junior staff safe tools can train them properly, instead of leaving them to learn by trial, error and private prompts. A firm that understands its own AI exposure can speak more clearly in client pitches, board meetings, regulatory discussions and transaction due diligence.

Jersey has its own reason to take this seriously. The jurisdiction's reputation rests on doing serious things properly, and a firm with credible AI governance is in a position to show clients, regulators and acquirers that it understands how its operations actually work. That will matter more, not less, as AI features disappear into ordinary software. The competitive question is unlikely to be whether a firm uses AI, that has effectively been settled in practice. The question will be whether it can explain its use.

The aim is not perfect visibility, which is not available in any real firm, but manageable visibility: enough to know where AI sits, what it touches, who owns it, what can go wrong, what controls exist, and which decisions still rest with humans. That gives clients, regulators and acquirers something more durable than reassurance. It gives them evidence.

The first AI governance move, in the end, is the work of finding out what the firm is already doing.

If you would like to map your firm's AI exposure before writing a policy, we would be happy to discuss. Contact Sindri Consulting to get started.