AI governance for marketing operations

We govern the AI-assisted systems we operate for marketing work: content drafting, reporting and analysis, data preparation, internal automation, and workflow integrations for intake follow-up and measurement. Each has a named owner, an approved use, and a review point.

We do not operate clinical systems, and nothing on this page concerns diagnosis, treatment, or patient care. Where a client uses AI inside its own systems, the client's governance applies to those systems. Ours applies to what we build and run.

Governance here means deciding where automation is appropriate and where a person must review the output or approve an action. The controls below expand the AI governance summary on our homepage.

Four approvals sit between an idea and an automated action in production. Each one has a named approver and a record.

• Use-case approval. Before a workflow is built, its purpose, inputs, outputs, data sources, and limits are written down and approved by the engagement owner. Anything that touches client data also needs the client's approval of that use.
• Pre-deployment testing sign-off. The workflow runs against sample or approved data, acceptance criteria are checked, and a reviewer signs off before it is enabled.
• Change approval. Changes to a model, prompt, configuration, data source, or integration are approved before they are applied, with the previous version retained.
• Publish and send approval. Anything that reaches the public or a prospective patient, such as page copy, ads, or follow-up messages, is approved by a person before it is published or sent.

Automation can prepare, draft, sort, and summarize. People decide what is published, what is sent, and what changes.

Every system we operate has a named owner who is responsible for who can reach it. Access is granted by role, limited to what the role needs, and reviewed on a schedule.

Credentials for client-owned platforms are issued by the client, scoped to the work, and revocable by the client. We do not share credentials between people, and we do not reuse a credential across engagements.

When a role changes or a person leaves an engagement, their access is removed as part of the handoff. The access list is available to the client on request, as described on our data sovereignty page.

Each workflow has acceptance criteria defined with the use case: what a correct output looks like, which inputs it must handle, and which it must refuse. Testing uses sample data or data the client has approved for that purpose.

We look for wrong or fabricated outputs, inputs that produce unsafe or off-policy results, data reaching a system it should not reach, and conflicts with advertising platform policies. Results are recorded with the sign-off.

During an initial period after deployment, a person reviews outputs at a higher rate before the workflow moves to routine sampling. The length of that period is set in the use-case record.

Configurations are versioned. A change log records what changed, who approved it, when it was applied, and what the previous version was.

Changes that alter what data a workflow uses, where it runs, or what it can send require re-testing against the acceptance criteria. Changes that affect a client's workflow are communicated to the client owner before they take effect.

Staged rollout is the default. A change runs on a limited scope first, and the results are reviewed before it applies everywhere.

Automated workflows are sampled on a schedule, and the sampling rate depends on what the workflow does. Anything that reaches the public or a prospective patient is sampled more often than internal reporting.

We watch for output drift, error rates, failed runs, and unexpected data in inputs or outputs. Alerts go to the system owner, and the review is recorded.

Monitoring is part of the weekly operating checks we propose for each engagement, with the cadence defined in the engagement.

If a workflow produces a harmful, incorrect, or off-policy result, the first step is to stop it. The owner pauses the automation and reverts to the last approved configuration.

The client owner is notified, the incident is recorded with its cause and effect, and the fix is tested against the acceptance criteria before the workflow is re-enabled.

Where an incident involves client data, the client's own incident process governs the notification and response. We provide the record and support the review.

A person is responsible for every automated action. That person is named in the use-case record, and the engagement owner is accountable for the engagement as a whole.

Escalation goes from the system owner to the engagement owner to agency leadership. Clients know who to contact and what happens next.

The governance record, including use-case approvals, test results, the change log, monitoring reviews, and incidents, is available to the client for the systems in their engagement.

We use the NIST AI Risk Management Framework as a reference for how risks are identified, measured, and managed. Advertising platform policies and the client's own compliance program set the rules that automated work must stay within.

These controls support responsible execution. They do not mean technology alone resolves every security or compliance obligation, and they are not a certification. HIPAA and related obligations belong to the client's program, and we work within the client's review process.

Governance applies to the workloads that run on our owned infrastructure and to the integrations that connect them to client systems. Data sovereignty covers where data resides and who can use it. Customizable AI shows a scoped workflow with its approval and review points. Our AI systems policy states the commitments behind these practices.

If you are evaluating a marketing partner's use of AI, ask who approves a use case, who signs off testing, who can change a configuration, and who approves what gets published or sent. An audit of your current marketing systems is a practical way to see where automation already runs and whether those four answers exist.