AI Business Rules

AI Business Rules is the seller Page for deciding what your storefront accepts and how review happens. It brings together Brief Acceptance, which sets the content rules for advertisers, categories, and creatives; Creative Policy; approval gates; and approval routing. Brief Acceptance is the content half of Govern: it says what is acceptable, while your approval settings say whether a human reviews eligible work. Brief Acceptance is enforced during product discovery, including for a Listing storefront before an eligible brief is forwarded to its connected sales agent. A storefront using an Interchange-managed manual media-buy approval workflow evaluates the submitted buy again before Approval Mechanics determines whether it can proceed. Provider-policy-owned social adapter storefronts pass transactions through to the social platform without an Interchange-managed approval queue. The current cohort is LinkedIn, Meta, Pinterest, Reddit, Snap, and TikTok. The social platform owns acceptance in those paths, so these storefronts do not need an Interchange acceptance policy for readiness or transaction admission. The provider’s own review rules still apply. Adapter-routed storefronts outside this cohort — including Amazon, Google Ads, and Spotify — retain their configured Interchange approval behavior and need a policy when that workflow is enabled. You can author it either way:

• Open the portable AI Business Rules page from Murph, Claude, ChatGPT, or another MCP Apps host to create or edit the full policy inline.
• Tell Murph the exact rules you want. Murph shows you the complete markdown for confirmation, then writes it through the same typed configuration contract.

Creating or activating policy versions requires an account administrator. Other storefront members can review the active policy and immutable history without receiving a hidden mutation path through the Page or MCP API. You don’t pick from a fixed menu of categories. You describe your rules, and the agent reads what you wrote. Every save creates an immutable version and can activate it atomically, so version history never contains a half-applied update.

​

Edit and restore versions

The AI Business Rules page always reads the current active Brief Acceptance policy and its version history. Save and activate creates a new version without overwriting the old one. If versions exist but none is active, the history remains visible so an administrator can preview, reactivate, or reuse a version as a new draft. The page performs those operations directly; it does not turn a button click into a new chat prompt or depend on a particular host model.

​

Policy vs. approval mode

These are two separate dials. Keeping them straight is the most common point of confusion: Brief Acceptance has two enforcement points. During product discovery, a confident conflict with the seller’s written policy returns no products; an unavailable or inconclusive qualifier fails open so discovery remains available. At create_media_buy, manual mode evaluates the submitted buy again: clearly on-policy buys auto-forward, while ambiguous results, evaluator errors, and deterministic hard findings queue for an operator. Live auto-rejection requires a separate seller opt-in or versioned rollout. auto mode preserves the established auto-forward contract and does not add a second create-time evaluator gate.

​

How the agent reads your policy

The manual create-time pre-screen reads your written policy against the buyer’s submission. It is not a fixed category list — a rule like “we don’t accept fast-food advertisers” is honored even though “fast food” isn’t a built-in category, because the agent derives the subjects you wrote and matches them against the submission. It also recognizes common evasion (look-alike characters, light obfuscation) so a rule isn’t sidestepped by spelling. When it cannot judge a submission with confidence, it errs toward holding it for you. When no policy document is available, the AI Business Rules Page starts with a conservative draft: no category is pre-cleared, every advertiser, category, and creative needs review, and illegal, malicious, deceptive, discriminatory, exploitative, phishing, impersonation, or destination-concealing submissions are flagged for operator rejection. This draft has no effect until an account administrator reviews, saves, and activates it. While that starter remains active, its review-all contract also makes the deterministic pre-screen hold every submission for an operator; it is not just instructional copy.

​

The three verdicts

The manual-mode create-time pre-screen sorts a managed-storefront media buy into one of three buckets:

• Clearly on policy → eligible and auto-forwarded without operator review.
• Needs a look → queued for an eligibility decision. Evidence is thin, the policy is ambiguous, or a rule says this category needs review.
• Clearly against an explicit rule you wrote → queued with a strong operator-facing rejection signal, without exposing private policy text to the buyer.

An enabled per-buyer auto-approve override bypasses this create-time screen and the human queue for that buyer’s media buys. Creative review remains separate.

​

The second opinion

On top of the deterministic manual-mode pre-screen, an AI layer gives a second opinion. It can only downgrade a proposed clear result to “needs a look” if it spots a policy conflict the rules missed (for example, a competitive- exclusivity clause you wrote in prose). It can never approve a buy on its own, never reject one, and never overrides a hold. It also fails closed within the manual approval path: if the AI layer is unavailable, errors, or is uncertain, the buy stays with the operator rather than being cleared by the evaluator.

​

How creatives are evaluated against your policy

Creative review applies the same content rules to a buyer’s submitted creative, plus a rendered-image check that looks at the actual pixels for brand cues, prohibited categories, and third-party marks. As with media buys, the AI layer can route a creative to human review but cannot approve one over a clear policy violation the deterministic checks already found. For locked assets — a product shot, logo, or book cover you’ve marked as exact-fidelity — there’s a stronger, deterministic guarantee that the AI layer cannot override:

• A creative whose final render goes through the platform’s locked-asset compositor preserves your supplied asset exactly, and clears that check.
• A final-stage creative that carries your locked asset but did not go through that deterministic path — for example, one that re-generated or edited your product from a reference — is blocked from auto-finalizing and held back for a human to inspect. A buyer’s generated creative cannot quietly substitute a look-alike stand-in for your locked product at final approval.

These fidelity checks are deterministic and run independently of the AI layer, so a crafted creative or brief can’t talk the reviewer into clearing them — they err toward human review whenever exact fidelity isn’t deterministically guaranteed.

​

What the agent never does on its own

• It never hard-rejects a submitted media buy in the current rollout. Deterministic Brief Acceptance conflicts can block product discovery and supply a strong operator-facing queue signal; AI judgments can only escalate.
• It never changes an approval gate for you. As it watches your decisions it may suggest opting a consistently-clean buyer into per-buyer auto-approve — but you author the carve-out. See the learned suggestion in seller analytics.
• Within manual create-time evaluation, it never promotes an uncertain result to eligible. If the AI layer can’t corroborate, the buy stays with the operator. Automatic storefronts do not run this second create-time evaluator.