When an advisor asks an AI assistant to summarize a client meeting, the request they type is only part of what the model reads. Behind it sits a layer of instructions they never see. That layer is shadow prompting: hidden AI instructions, written by a vendor or by your firm, that set the tone, permissions, format and limits of every response.
For RIAs, that hidden layer is now an exam topic. In its fiscal year 2027 examination priorities, released October 23, 2026, the SEC's Division of Examinations said it will prioritize examinations of policies and procedures for monitoring firms' use of AI, a technology the document calls super intelligence, including in fraud prevention and detection, back-office operations and trading. Its fiscal year 2026 priorities had already said examiners would assess whether firms have adequate policies and procedures to monitor and supervise their use of AI. The SEC describes these priorities as staff views that create no new obligations.
A firm can't supervise behavior it can't see, and any firm using AI tools is likely already running hidden prompts it has never reviewed.
How shadow prompting works
Every AI tool sends the model a set of instructions before the user's message arrives. These are usually called system prompts in AI products, and because users never see them, they're also described as invisible prompts or shadow prompts. They work in three layers.
System prompts set the rules
A hidden instruction might tell a client service assistant to stay factual, avoid anything resembling investment advice and route account-specific questions to a human. The advisor just types a question; the rules apply automatically.
Context layers supply the facts
Shadow prompts can carry CRM notes, custodian data, a user's role and the firm's approved policies. That lets the model answer from authorized information instead of guessing. It also means the hidden layer decides what client data the model can see, which makes it relevant to your Reg S-P obligations.
Prompt chains break work into steps
Many tools run several hidden prompts in sequence. One classifies the request (onboarding, compliance, service), another retrieves approved data, and a third drafts the answer within policy. This kind of AI response steering is how vendors make complex workflows reliable.
Why shadow prompting matters for wealth management firms
Hidden instructions are what make AI usable at firm scale. They give you LLM output control without asking every advisor to master prompt engineering methods.
- Consistency across advisors. Approved language, disclosure phrasing and escalation rules apply to every response, not just the ones written by your most careful users.
- Less friction. Advisors get answers without learning to structure detailed prompts, which drives adoption of sanctioned tools.
- Built-in AI guardrails. Hidden rules can mask sensitive data, block performance claims, and limit what each user can access based on role.
The same features that make shadow prompting useful also make it a governance question. Whoever writes the hidden instructions is, in effect, setting policy for how AI behaves at your firm.
How to govern shadow prompting: an enterprise AI governance checklist
Governing the hidden layer doesn't require new technology so much as applying controls you already know. Treat prompts like any other policy that touches client data.
- Inventory every AI tool and record who controls its hidden instructions: the vendor, your firm, or both.
- Version-control firm-written prompts with an owner, change date, test results and approval, just as you would for a written procedure.
- Separate governance rules from task instructions. Keep security, data access and disclosure rules in one controlled layer; let task prompts change without touching it.
- Ask vendors the hard questions. Can you see or configure the system prompts? How do they defend against prompt injection? Are inputs, outputs and overrides logged and retained?
- Monitor outputs. Sample responses for errors, unsupported claims and override attempts, and feed what you find back into the prompts.
- Write it into your compliance program. AI use belongs in the written policies required under Rule 206(4)-7 and in your annual review.
Firms that do this are better placed to explain how their AI workflows are governed, which is the question examiners are asking. A platform built for AI compliance in wealth management can apply these controls by workflow instead of leaving them to each vendor.