AI assistant for marketing: define its decision boundary

An AI marketing assistant is useful when its task and permissions are explicit. “Manage our marketing” is not a boundary. “Sort anonymized incoming questions into themes and draft a suggested reply for human review” is. Before connecting an assistant to a CMS, inbox or store, write down what it may read, propose, change and publish. The more public or irreversible an action is, the stronger the review should be.

Draw the decision boundary

Make a table with four columns: task, allowed input, allowed output and required approval. For a first pilot, let the assistant classify customer questions, identify repeated themes and produce draft answers. A person checks facts and sends or publishes the final text. Do not grant publishing permissions because the interface makes it easy. A draft-only role makes errors visible before customers see them.

Define excluded decisions explicitly: setting a discount, promising delivery, changing a product specification, deleting a page, contacting a customer or disclosing private information. If some of these are later needed, add them one at a time with their own evidence and approval rule. “Use your judgement” is too vague for a system that can act at scale.

Design the source packet

Give the assistant a short, approved set of sources: current FAQ, product facts, price policy and tone guide. Each source has an owner and update date. A response should cite the source behind a factual claim or say that no approved answer exists. An assistant that confidently completes a sentence from memory may sound helpful while contradicting today's offer.

Check the source packet for contradictions. If one document says the guide has 12 chapters and another says 13, the model should not choose silently. It should flag the conflict for the owner. Remove unnecessary personal data from examples. Limit access to the minimum accounts and documents needed for the task. Document who can change the source packet; otherwise stale facts will keep returning even when prompts improve.

Write escalation rules for difficult cases

An assistant should stop or hand off when it sees a complaint, refund request, safety issue, legal question, unusual price, sensitive personal detail or unclear customer intent. A useful rule is specific: “If the requested variant is missing from the approved catalog, do not infer compatibility; create a ticket for the product owner.” Include a response the person can use while waiting, such as acknowledging the question without promising an outcome.

Test those rules with examples before launch. Include a routine question, an ambiguous question, a request to reveal private information, a hostile message and a source document containing instructions aimed at the assistant. The last case checks whether the system treats outside text as evidence rather than as a command. Record which cases were handled, handed off or answered incorrectly. A successful demonstration on easy questions is not enough.

Keep human review meaningful

A reviewer needs to see the customer's actual question, the proposed answer, the cited source and any uncertainty. A bare “approve” button with no evidence invites rubber-stamping. Give the reviewer a way to edit, reject and send the case to another owner. For public marketing content, preview the final page and check title, body, links and metadata. For a customer reply, confirm the address and any personalized facts before sending.

Human review has a cost. Measure it. If the assistant creates ten weak drafts that each take longer to fix than writing a response, narrow the task to classification or source retrieval. If it handles only easy cases, say so. The NIST AI Risk Management Framework emphasizes managing risks across the AI system; in practice, review and handoff are parts of that system, not decorative policy lines.

Log actions and make reversal possible

Keep a minimal record: input category, date, source version, proposed output, reviewer, decision and published URL or sent message ID. Avoid storing more personal data than the workflow needs. When an error occurs, the log shows which pages or customers may be affected. A rollback plan might mean restoring a previous page revision, pausing the integration and reviewing recent outputs. Test that plan before increasing permissions.

Use a separate test environment or draft status where possible. Limit the volume of actions per hour while the assistant is new. A single incorrect public claim is easier to contain than a hundred simultaneous edits. Agree who can turn the workflow off and how that person is alerted when a serious error appears.

Evaluate the whole workflow

Establish a manual baseline: how many questions arrive, how long they take to classify, how often a person needs to investigate and what a good response looks like. Test the assistant on a representative set, including difficult cases. Measure correct classifications, unsupported claims, handoff accuracy, review minutes and customer outcomes. Compare total time, not model response time. A fast first draft that creates more corrections may be a loss.

For a fictional agency, an assistant labels 60 anonymized inquiries weekly as audit, redesign or general question. It must not quote prices. A reviewer checks every label during the pilot. Twelve ambiguous inquiries go to a person. After three weeks, the team finds that classification is mostly useful but drafted replies often imply a fixed turnaround that the agency has not approved. The sensible change is to keep classification and disable reply drafting until the source policy is settled.

Connect the assistant to the customer journey

The assistant should help a customer complete a real step, not merely create more messages. If it groups questions, use the themes to improve a help page. If it proposes content briefs, verify that the resulting pages answer distinct questions. If it drafts product copy, follow the product fact sheet and variant checks. Google's guidance on AI-assisted content expects accuracy and added value; an automated volume of near-identical pages does not meet that test.

Review the assistant's boundary whenever the offer, team or connected tools change. A new data source can change the privacy and accuracy risk. A new permission can turn a draft helper into a publisher. Record each change and rerun the difficult-case tests. A good assistant is one whose useful actions, limits and failures can be explained by the people responsible for it.