Skip to content
← All articles
A Client-Facing AI Policy Template for Agencies

A Client-Facing AI Policy Template for Agencies

An AI policy is not a compliance artifact. It is a sales document, and it works because the alternative is the agency that says “we use AI where it makes sense” and hopes the topic moves on. A named scope, a sign-off rule and a priced opt-out answer the question before procurement asks it.

One page, attachable to a proposal, not legal advice.

In short

  • Write the policy around three lists: what AI touches, what it never touches, and who signs off. Everything else is decoration.
  • A rule that is not enforced by a tool is a promise. Map each clause to a mechanism, or delete the clause.
  • Price the opt-out instead of refusing it. The number does the negotiating for you, and most clients do not take it.

Why the document closes deals

Three things changed in the last two years of client conversations. Buyers now assume AI is involved somewhere, they have started asking which parts, and they have a legal or procurement person who wants an answer in writing. An agency that answers in writing looks organised. An agency that answers verbally looks like a risk they have to manage.

The policy also protects you internally, which matters more than the pitch. It settles the argument you have every month about whether the intern may let the assistant write the first comment. Once the boundary is written and attached, that argument stops being a judgment call each time.

The template

Copy this, replace the bracketed parts, keep it to one page. The sections are ordered so a client reads scope, then safety, then cost.

1. Scope. This policy covers how [Agency] uses automated and AI-assisted tooling to produce, review and publish social media content for [Client]. It is a description of our practice, not a warranty, and it is attached as a schedule to our agreement.

2. What AI touches. Draft options from a client-approved brief. Variants of an approved post per network. Caption rewrites for tone or length. Automated checks for claims, missing disclosures and platform limits. Transcription of calls and source material. Nothing in this list publishes on its own.

3. What AI never touches. Strategy and channel decisions. Replies in a live conversation with a customer. Crisis and complaint handling. Anything under NDA, or unreleased products, pricing and customer data. Legal, medical and financial claims. Approving its own work.

4. Sign-off. Every post is approved by a named person at [Agency] before it publishes, and the approval is recorded against that version of the post. If [Client] nominates a reviewer, the same record applies to their decision. A post that is waiting for approval cannot be rescheduled or pushed out around the approval step; it publishes only after the approval is resolved.

5. Confidentiality. Client material stays in the workspace assigned to the client. Prompts are limited to the content needed to produce the draft, and the list in section 3 never enters a prompt. [Client] material is not used to train models we control.

6. Disclosure. We tell [Client] which parts of a deliverable were machine-drafted on request, and we follow the platform’s own rules on labelling automated content where they apply.

7. Opt-out. [Client] may exclude any or all channels from AI-assisted drafting with 14 days notice. The excluded channels continue on manual drafting at [the rate in the agreement], and the difference reflects the production time it adds.

8. Evidence. On request we provide the approval record for any published post: the version approved, the approver, the timestamp and the final published text.

9. Review. We review this policy every six months and give 30 days notice of any change that affects [Client].

Make each clause enforceable

The failure mode of these documents is that they promise practice the agency cannot evidence. Before you send the policy, walk each clause and ask what would prove it. This is the mapping I would accept from a scheduling layer:

ClauseThe mechanism that proves it
Nothing publishes without a named approverApproval is a state on the post, recorded with the actor, the time and the version, and readable outside the UI
A post cannot be routed around the reviewA post in the review state refuses status and date changes from every other endpoint, and only an approval action resolves it. New rounds drop unused approval links
Certain topics never publish unattendedRules that match on words, links, video or posting time, with an effect of require approval or block
Agents can only publish inside agreed limitsAutomated posts are separated from human posts by actor type, so a rule can target the agent and spare the team
One client’s work stays separateA workspace per client with its own channels, history and credentials
Speed has a brakeA single switch that pauses publishing across the workspace, and a distinct error that tells your automation it was paused rather than broken
The signed-off wording is the published wordingThe published text and the approved version are both retrievable, and compared on request

That last row is the one agencies skip and the one a client’s lawyer will ask about. If the only artifact you can produce is a screenshot of Slack, the clause in section 4 is decorative.

How to introduce it in the first month

Do not send the policy as a surprise attachment in month seven. Sequence it so it reads as process, not as a confession.

In week one, send it to your two most reasonable clients with one line: this is how we work now, tell me if anything is off. Fix what they flag before anyone else sees it. In week two, attach it to every new proposal and add the opt-out line to your rate card, because a policy with a price is credible and a policy without one is decorative. In week three, check that the mechanisms behind it exist in your tooling: approval state on the post, per-actor rules for anything automated, a pause switch, and an export that includes who approved what. In week four, put the review date on your calendar, six months out, and note what would trigger an early revision, such as a new platform or a new AI feature in your stack.

That sequence also produces the answer for the client who asks whether you are making this up as you go. You are not, and you have the dates.

Where the policy will actually be tested

Three moments, in my experience, decide whether the document survives.

The 4:55pm request. A client forwards a press mention and wants it posted now. If the policy says every post has a named approver, this is now a two-minute exchange where the client approves their own request and it goes out with a record. Without the policy, it goes out under your login and nobody can say who asked for it.

The agent with a live key. Someone connects an automation with an API key and it starts scheduling. The policy promised human approval, and the tool either enforces that or the clause is a lie. This is where per-actor rules earn their place: require approval for anything an agent creates, and let human scheduling stay fast.

The day the client leaves. They ask for proof of what was published and who signed off. An export with approvals and timestamps is a clean handover. A mailbox is an argument. The offboarding checklist covers the rest of that conversation.

What I would cut

Most agency policies I read are four pages of principles that nobody enforces and no client reads twice. Cut anything that cannot be evidenced: “we use AI responsibly”, “we monitor outputs continuously”, “we follow best practice”. Cut the tool inventory too, because it dates within a quarter and invites a question about the tool you dropped.

Keep three lists, one sign-off sentence, one confidentiality list, one priced opt-out and one review cadence. That is a page a client can actually approve in a proposal meeting.

One more cut worth making on purpose: leave out the tool names you use internally. The client is buying the practice, not your stack, and a named vendor in a schedule invites a due diligence conversation you did not ask for. Name the controls instead, and keep the vendors in your own operations notes.

Two related posts go deeper on the pieces clients ask about most: telling clients you use AI before they ask and automation that does not read as slop. If the enforcement side is the open question, the agency scheduling setup shows where approval state, per-actor rules and the publishing pause live, and the pricing page is where the opt-out conversation turns into client capacity.

One last note, since this is a template on the internet: I am not a lawyer, and this is a description of how to write a policy, not legal advice for your jurisdiction.

Frequently asked questions

Do we need a written AI policy if the client has not asked?

The pitch is where it earns its keep. A named scope, a sign-off step and an opt-out with a price turns an awkward question into a competence signal. The agency with a document wins against the agency that answers with a shrug, and the document costs an afternoon.

What belongs in the policy and what belongs in the contract?

The policy states how work is produced and reviewed. The contract carries the obligations, the price and the termination terms. Attach the policy as a schedule to the contract so a change of tooling does not require a renegotiation, and say in the policy that it is a description of practice, not a warranty.

How granular should the sign-off promise be?

One level: every published post is approved by a named person before it publishes. Anything more elaborate breaks in a busy week, and a broken clause is worse than a modest one. If the client wants velocity on a specific channel, write an exception with a named approver rather than lowering the rule.

Where does confidentiality actually break?

In prompts. Anything pasted into a draft request leaves your tooling. So the policy should name what may never enter a prompt: unreleased product names, pricing not yet public, customer lists, anything under NDA. Keep it to a short list a junior can memorise.

Let your AI post
to social media for you.

Connect your accounts, add PostSider to your AI and publish your first post in a few minutes.

7-day free trial · 30+ networks · No credit card · Cancel anytime