Skip to Content

Write the AI Policy

This is the fifth step in The Adoption Path. The first four steps produced four documents: a use-case list, a duty map, a data matrix, and a set of approved tools. This step turns those into one written policy — not a new, standalone AI policy sitting off on its own, but a section inside the compliance manual the firm already has. A policy that just restates the four documents in prose adds nothing; a policy that invents rules those four documents don’t support is ungrounded, and hard to defend in an exam.

Why it matters

Without this step, one of two things happens. Either use continues informally, back to the exact problem Decide What AI Is For exists to prevent, or the firm writes generic AI-safety language that sounds responsible but doesn’t reflect anything that was actually decided in the first four steps. The policy is what turns four documents of internal reasoning into something enforceable and inspectable — a firm that has done the first four steps but never written the fifth has done real work that nobody outside the firm can see or check.

What actually happens

Decide where the policy lives

The policy is a new section inside the compliance manual the firm already maintains (for an SEC-registered adviser, the compliance-program rule described in Map the Rules That Apply requires one) — not a separate AI policy living on its own, disconnected from the supervision the rest of the manual already gets. Attach the four documents from the first four steps — the use-case list, the duty map, the data matrix, and the tool approvals — as dated appendices, rather than retyping their contents into new prose. The policy’s job is to say what those four documents mean for day-to-day conduct; it is not a fifth copy of them.

Write what the policy actually has to say

Beyond the appendices, the policy needs to say a few things plainly, each pointing back to a specific document rather than standing alone. Who may use which tools for which approved uses — referencing the use-case list by name, not “AI” generally. What must never be entered into a general-purpose tool — referencing the data matrix, so the rule is the one already worked out, not a new one written from memory. Who reviews AI-assisted output before it is used internally or sent to anyone outside the firm — referencing the review point without repeating Handle Client-Facing Use in full; link to it instead of restating it. And who can approve a new use or a new tool going forward — referencing the approval decision from Vet and Choose Tools, so approval stays where step four already put it. The duty map is what every one of those provisions is checked against: a provision that answers no duty the map names is one the four documents do not support.

Set the review cadence

Tie the policy’s own review cadence to two things already decided rather than inventing a third schedule: the review step later in the Path, and the re-check trigger set in step four for individual tools. A policy reviewed on its own separate timeline, out of step with the rest of the firm’s compliance calendar, tends to get missed; folding it into the same rhythm the firm already runs is what keeps it current.

Decide how the policy is communicated and acknowledged

A policy nobody has read is not a policy, it is a document. Decide how it gets communicated — at a minimum, everyone it covers reads it and acknowledges that in writing — and make sure that includes the assistant, since the use-case list from step one covers them too. An unacknowledged policy is functionally the same as no policy at all if a use goes wrong and nobody can show anyone was told the rule.

Address what happens when the policy is silent

A use nobody thought of yet will come up — say plainly, in the policy itself, that “not addressed” is not the same as “allowed.” Name who decides in that gap: the same person who approves new tools in step four is a reasonable default, since a use the policy hasn’t addressed is functionally the same question as a new tool or a new use case. Without this line, silence in the policy quietly becomes permission by default, which defeats the entire point of writing one.

Questions to ask yourself

  • Does the policy live inside the compliance manual, or somewhere separate that supervision might not reach?
  • For each rule in the policy, can you point to which of the four earlier documents it comes from?
  • Has everyone the policy covers, including the assistant, actually acknowledged it?
  • If a genuinely new use came up tomorrow, does the policy say who decides, or would it just get decided informally?
  • Is the policy’s review date tied to the firm’s existing compliance calendar, or off on its own?

Common mistakes

  • Writing a policy that describes what the firm hopes is true rather than what steps one through four actually found.
  • Creating a second, freestanding AI policy instead of folding it into the existing compliance manual.
  • Writing a policy so general it can’t actually be violated — “use AI responsibly” is not a rule anyone can check against.
  • Never updating the policy when a new tool is approved in step four.
  • Forgetting to have the assistant acknowledge it, because the policy quietly gets read as being for advisers only.

Which Firm Functions apply

  • Compliance and Supervision: The policy is this Firm Function’s core document — it is what supervision is actually checking use against.
  • Records and Documentation: The policy, its appendices, and every acknowledgment of it are the kind of records a firm may be asked to produce.

At a larger firm, the AI policy is not a new document the adviser writes. It is a set of provisions already sitting inside the firm’s existing supervisory and compliance manual, and the adviser’s own use has to fit inside them. The job is finding those provisions and following them, not drafting new ones from scratch.

Next step

Build the Workflows: Turn the policy into the specific, repeatable procedures for each approved use.