A client AI policy gap audit for agencies using generative AI

Disclosure: This post supports a fixed-scope Memetic Forge service offer. It is not legal advice and does not contain affiliate links.

Most small agencies already use generative AI somewhere: research, drafts, briefs, images, meeting notes, code, reporting, client-service macros, proposal writing, or internal ops.

The risk is not that a team uses AI. The risk is that the agency cannot answer simple client questions consistently:

• Which client data can staff put into AI tools?
• Which outputs need human review before they leave the agency?
• When should AI assistance be disclosed?
• Which tools are approved vs. experimental?
• Who owns quality, confidentiality, and recordkeeping when AI helps create the work?
• What happens when a client asks for the policy?

If the answer is "we handle it case by case," the agency has a client AI policy gap.

The audit lens

NIST describes the AI Risk Management Framework as voluntary guidance for improving how organizations incorporate trustworthiness considerations into AI design, development, use, and evaluation. NIST's generative AI profile adds that organizations should identify unique generative-AI risks and choose risk-management actions aligned with their goals and priorities.

For a small agency, that does not mean building an enterprise governance office. It means creating a short, usable operating system:

1. approved use cases;
2. prohibited data inputs;
3. required human review;
4. disclosure and attribution rules;
5. vendor/tool approval criteria;
6. client exception handling;
7. incident and escalation path;
8. staff rollout checklist.

The policy should be boring enough that staff actually follow it.

The seven gaps I would check first

1. Client-data handling gap

Can the team distinguish public information, internal agency notes, confidential client material, personal data, regulated data, credentials, and proprietary client strategy?

A useful policy has a simple rule table:

Data type	Allowed in approved AI tools?	Conditions
Public website copy	Yes	cite source if used externally
Internal agency template	Yes	remove client identifiers if reused
Client confidential documents	Restricted	only approved tools/settings and client permission if required
Personal data	Restricted/No	follow contract, law, and minimization rules
Credentials, secrets, API keys	No	never paste into AI systems

Without this, staff default to convenience.

2. Tool-approval gap

Agencies often have five AI tools in use before anyone maintains a list.

The first version does not need procurement theater. It needs a lightweight approved-tools register:

• tool name;
• owner;
• approved use cases;
• data allowed/not allowed;
• client opt-out impact;
• retention/training setting notes;
• review date.

If a tool is experimental, label it experimental and limit client-data exposure.

3. Human-review gap

"AI assisted" should not mean "unowned." Define which outputs require human review before client delivery:

• claims about performance, revenue, compliance, health, finance, or law;
• factual research and citations;
• client strategy recommendations;
• code, automations, and workflow changes;
• paid ad copy and landing-page claims;
• public brand content.

A practical rule: AI can accelerate drafts, but a named human owns the final client-facing output.

4. Disclosure gap

Disclosure rules vary by contract, client expectation, medium, and risk level. The agency should not improvise them project by project.

Create three standard disclosure levels:

1. Internal-only AI assistance — no client-facing disclosure unless contract requires it.
2. Material AI assistance in deliverable creation — disclose in project notes, SOW language, or delivery memo.
3. AI-generated/AI-operated client-facing experience — disclose directly where users interact with it.

The goal is not over-sharing every prompt. The goal is avoiding surprise.

5. Claims and proof gap

Agencies should be especially careful when AI is used to produce marketing claims. FTC business guidance has repeatedly warned companies not to overstate what AI can do or make unsupported performance claims.

A safer agency workflow asks:

• What exact claim are we making?
• What evidence supports it?
• Is the claim about the client's product, our service, or the AI tool?
• Could a reasonable client/user misunderstand the level of automation or review?
• Who approved the final wording?

This is where a policy turns into revenue protection: fewer risky promises, fewer rework cycles, fewer client escalations.

6. Client-contract gap

Many agencies already have confidentiality, IP, subcontractor, data-processing, or security commitments in client agreements. AI usage can accidentally touch all of them.

The policy should include a contract-screening step:

• Does the client prohibit third-party tools without approval?
• Does the client require written consent before using AI/subprocessors?
• Are there data-location, retention, or security terms?
• Is the deliverable subject to regulated review?
• Does the SOW promise all work is original, human-created, or client-exclusive?

This is not legal advice; it is a prompt to stop before staff create avoidable contract tension.

7. Staff rollout gap

A policy nobody reads is not a control.

A workable rollout has:

• a one-page rule summary;
• examples for each team function;
• a Slack/Notion decision tree;
• an approved-tool list;
• a client exception workflow;
• a quarterly refresh owner.

For small teams, the best policy is short enough to paste into onboarding.

A compact policy gap scorecard

Area	Green	Yellow	Red
Client data	staff know what data is allowed	rules exist but are scattered	no clear restrictions
Tool approval	approved register exists	informal list only	staff use any tool
Review	named human owns final work	review depends on team	AI output ships unchecked
Disclosure	standard levels by use case	ad hoc disclosure	no disclosure norm
Claims	proof check before publication	some claim review	unsupported AI/ROI claims
Contracts	AI-screening question in intake	only major clients checked	contract terms ignored
Rollout	one-page SOP + owner	policy doc only	no staff training

Any red row is a good candidate for a 48-hour cleanup.

What a first audit should produce

A useful first pass should not be a 40-page governance deck. It should produce:

• a ranked policy gap map;
• a one-page staff AI use policy;
• an approved/prohibited data table;
• a tool approval register template;
• disclosure snippets for SOWs, delivery notes, and client FAQs;
• a rollout checklist for account, creative, ops, and leadership teams.

That is enough to make client conversations calmer and staff behavior more consistent.

If you want an outside pass

Memetic Forge runs a fixed-scope Client AI Policy Gap Audit for small agencies using AI in client work.

Typical first pass:

• review current AI usage, client-facing promises, and tool list;
• map policy gaps across data handling, review, disclosure, claims, contracts, and rollout;
• deliver a one-page policy, staff checklist, tool register, and client-safe disclosure snippets.

The first audit is scoped for 48 hours and does not require client confidential data. Sanitized examples, current SOPs, and a walkthrough are enough.

If useful, email [e-mail] with the subject AI policy gap audit and the type of agency you run.

Source notes

• NIST AI Risk Management Framework page, accessed 2026-06-29: NIST describes AI RMF as voluntary guidance for managing risks to individuals, organizations, and society and incorporating trustworthiness considerations into AI design, development, use, and evaluation.
• NIST page also notes the July 26, 2024 Generative AI Profile, which helps organizations identify unique risks posed by generative AI and choose risk-management actions aligned with goals and priorities.
• FTC AI advertising/business guidance is treated here as a conservative operating principle: avoid unsupported AI-performance claims and keep proof behind client-facing statements.