Who Picks a Marketing Team's AI Tools and Who Sets the Rules
Centralize AI marketing tools or let channel owners choose? A federated model with a RACI, vendor admin controls and a policy mapped to the NIST AI RMF.
On this page
- Three ways to split the decision
- What stays central and what channel owners decide
- Who enforces the rules: legal, compliance or marketing ops?
- What the admin consoles let a central team control
- An AI governance policy for marketing, mapped to NIST
- A worked example (hypothetical)
- Where the line moves next
- Sources
Centralize the rules and let channel owners choose the tools. When a marketing leader asks whether to centralize AI marketing tools or let channel owners choose their own tools, the arrangement that holds up for most teams is federated: a central group owns data rules, vendor contracts, security review and the use policy, while each channel owner picks from an approved list and answers for what their tools produce. Full centralization pushes every channel's request through one queue. Full channel ownership leaves nobody holding the contract when a tool turns out to train on customer data or a connector reaches a folder it was never meant to see.
The admin consoles of the three big workplace assistants make the split concrete. Each documents controls that only a workspace admin can set, and Google's admin help says Gemini features in Workspace services are on by default, so on a Google Workspace Enterprise Standard or Enterprise Plus tenant the first decision has been made for you unless an admin changed it.
Three ways to split the decision
Marketing ops team centralized AI tool governance and decentralized channel teams sit at two ends of a range, with a federated model between them. The table compares all three.
| Central | Federated | Channel owned | |
|---|---|---|---|
| Who picks tools | Marketing ops or an AI lead buys one stack for everyone | Central team keeps an approved list; channel owners choose from it and request additions | Each channel lead buys what they like |
| Who holds the contract | Central team | Central team | Whoever signed up |
| Speed for a new use case | Slow, since every request waits in one queue | Quick for listed tools, slower for new ones | Quick |
| Data and security exposure | Lowest | Low when the risk tiers below are enforced | Highest, because nobody has a full list of tools or the data they touch |
| Best fit | Small teams with one ops owner, or regulated sectors | Teams with several channels and a marketing ops function | Short experiments on public data |
Federated suits most teams because the knowledge sits in two places. The paid media lead knows whether an ad copy assistant fits the way their team builds campaigns. The central team knows what the vendor's data terms say and whether single sign-on is switched on. Asking each side to decide what it knows best moves faster than asking either side to decide everything.
What stays central and what channel owners decide
Four things stay central because a mistake in any of them lands on the whole company:
- Data rules: which classes of data (public, internal, customer personal data) may go into which tool, and which connectors may be switched on.
- Contracts: order forms, data processing terms, model training terms and retention.
- Security review: single sign-on, admin roles, audit logs and retention settings, checked before a tool joins the approved list.
- Use policy: disclosure, human review before anything ships, rules for product and pricing claims, and incident reporting.
Channel owners decide everything inside those lines: which approved tool fits their workflow, the prompts, templates and custom assistants they build, the review step for their own output, and when a tool stops earning its seat. For that last call, set the time saved against the license cost in the AI ROI calculator before the renewal date comes round.
A risk tier on each use case tells everyone which side of the line a request falls on. The three tiers below are our own working scheme, offered as a hypothetical starting point:
- Tier 1 covers public inputs only, such as rewriting published copy or brainstorming headlines. A channel owner can use any approved tool without further review.
- Tier 2 covers internal non-public inputs, such as campaign plans or performance exports. Approved tools only, signed in through single sign-on, with the data rules applied.
- Tier 3 covers customer personal data, or a tool that acts on its own by sending, bidding or publishing. It needs a security review, a legal review and a named human approver for each action type. This is where a chatbot becomes agentic AI, and the permission tiers and audit trails in our governance operating model for marketing agents apply.
Who enforces the rules: legal, compliance or marketing ops?
Who should own AI governance enforcement on a marketing team depends on which part of enforcement you mean, because legal, compliance and marketing ops each own a different piece. Legal and compliance write the rules on claims, consent and contracts. IT and security run the reviews. Marketing ops enforces day to day, since it keeps the tool inventory and usually holds the admin seats for marketing's own tools. The CMO stays accountable for marketing's use of AI, which matches the call from NIST for executive leadership to take "responsibility for decisions about risks associated with AI system development and deployment" (Govern 2.3).
The matrix below is our suggested starting point. Govern 2.1 asks that roles and lines of communication be "documented and are clear to individuals and teams throughout the organization" and leaves the assignment to you. R means responsible, A accountable, C consulted and I informed.
| Activity | CMO | Marketing ops or AI lead | Channel owner | Legal and compliance | IT and security |
|---|---|---|---|---|---|
| Approve the AI use policy | A | R | C | C | C |
| Set data rules for AI tools | I | C | I | A | R |
| Security review of a new tool | I | C | C | C | A, R |
| Negotiate contract and data terms | I | R | I | A | C |
| Configure admin console settings | I | A | I | C | R |
| Choose a tool for a channel workflow from the approved list | I | C | A, R | I | I |
| Review AI-assisted claims before publishing | I | C | A, R | C | I |
| Keep the AI tool inventory | I | A, R | C | I | C |
| Handle an incident | I | A | R | C | R |
| Retire a tool and remove its data | I | A | R | C | R |
What the admin consoles let a central team control
ChatGPT Enterprise, Microsoft Copilot (formerly Microsoft 365 Copilot) and Gemini for Workspace each document controls that a workspace admin sets for everyone. The table quotes each vendor's own page as we read it on November 6, 2026.
| Control | ChatGPT Enterprise | Microsoft Copilot | Gemini for Workspace |
|---|---|---|---|
| Training on your data | Business data "is not used to train our models by default" | Prompts, responses and Microsoft Graph data "aren't used to train foundation LLMs" | "Workspace does not use customer data for training models without customer's prior permission or instruction" |
| Who gets the features | Admins "can manage who can use agents, connected tools, and actions across the workspace" | Admins choose which agents are allowed, in the Integrated apps section of the Microsoft 365 admin center | On or off per service, by organizational unit or group; "Group settings override organizational units" |
| Access to company content | We found no statement on the page we read | Surfaces only data a user has "at least view permissions" for, and honors the usage rights set by sensitivity labels | "If the user doesn't have access to a document or email, Gemini will not retrieve that content." |
| Retention and audit | "supports encryption, retention, and compliance controls" | Purview retention policies and Content search for Copilot interactions | Vault retention, audit logs, and client-side encryption that can restrict Gemini's access |
Three findings matter for the central team. First, Google says "The default setting for Gemini features in Workspace services is on," and, for Enterprise Standard and Enterprise Plus, lists separate switches for Gmail, Calendar, Drive and the Docs editors, Meet, Chat and Google Workspace Studio, so someone has to decide which organizational units get which service. The same page notes that a user in Gmail can still ask Gemini about a Drive file when Gemini is off for Drive, so the switches limit features while sharing permissions limit data. The Gemini app is a separate setting that, on the same page, "has no effect on other AI features in Google Workspace services."
Second, both Microsoft and Google ground answers in permissions the user already holds. Microsoft says Copilot "only surfaces organizational data to which individual users have at least view permissions," and Google says "The AI follows your existing Google Workspace permissions." An overshared folder is therefore within the assistant's reach for everyone it is shared with, so a sharing cleanup belongs in the security review before rollout.
Third, model choice is partly an admin decision now. Microsoft's page says admins can decide whether to use Anthropic and OpenAI models in Copilot, and notes that Anthropic models "are currently excluded from the EU Data Boundary." For a team with EU customer data, that setting belongs in the data rules.
For ChatGPT Enterprise, the product page from OpenAI lists "Compliance API visibility for conversations and agent activity" and admin control over who can use agents, connected tools and actions. We found no role or single sign-on settings described on that page, so confirm those in your own admin console before you write the data rules.
An AI governance policy for marketing, mapped to NIST
The best practice for an AI governance policy for marketing use in 2026 is to borrow the structure of a framework your legal and security teams already know, and the Govern function of the NIST AI Risk Management Framework fits well. NIST describes govern as "a cross-cutting function that is infused throughout AI risk management and enables the other functions of the process." Its table for the function lists six categories and 19 subcategories, and NIST says some organizations "may choose to select from among the categories and subcategories", so a marketing policy can map to the ones that bite.
| Policy section | What it covers for marketing | NIST Govern |
|---|---|---|
| Scope and legal requirements | Laws and platform rules that apply to AI-made ads, emails and pages, and who tracks changes | 1.1 |
| Risk tiers | The three tiers and the review each one needs | 1.3 |
| Review cadence | When tiers, tools and the policy itself are reviewed, and by whom | 1.5 |
| Approved tool inventory | Every AI tool with its tier, owner, contract and allowed data | 1.6 |
| Retirement | Exporting and deleting data when a tool is dropped | 1.7 |
| Roles | The RACI above | 2.1 |
| Training | What a channel owner must know before using a Tier 2 or Tier 3 tool | 2.2 |
| Executive sign-off | CMO approval for Tier 3 uses | 2.3 |
| Human oversight | Which outputs need a named approver and which can be sampled | 3.2 |
| Testing and incidents | How to report a bad output, a data leak or a wrong claim | 4.3 |
| Outside feedback | How customer complaints about AI content reach the owner | 5.1 |
| Vendors and rights | Contract terms, training terms, and rights in generated assets | 6.1 |
| Vendor failure | What happens when a tool goes down or changes its terms | 6.2 |
Govern 1.6 reads "Mechanisms are in place to inventory AI systems and are resourced according to organizational risk priorities." The inventory is the cheapest place to begin, because every other row in the table refers back to it.
A worked example (hypothetical)
Take a hypothetical B2B team on Microsoft 365 with leads for paid media, lifecycle, content and SEO, plus a marketing ops manager who keeps the approved list. The content lead wants Copilot to tighten published blog posts. That is Tier 1, Copilot is already on the list, and no request is needed.
The lifecycle lead wants a separate tool that drafts nurture emails from CRM segments. Customer personal data makes it Tier 3. Marketing ops finds it missing from the inventory and opens a request; IT checks single sign-on, retention and training terms; legal reviews the data processing terms and the claims rules; the CMO signs off on the use. Once it is approved, the lifecycle lead owns the prompts, the send approval and the brand rules, much like the controls in our rule sheet for an email-drafting agent.
The paid media lead wants an assistant that changes bids on its own. That is Tier 3 as well, and the handoff contracts and kill switches in the governance layer for multi-agent stacks belong in its review. Across all three requests, nobody waits on the central team for Tier 1 work and nobody signs a vendor contract alone.
Where the line moves next
As tools move from drafting to acting, more requests land in Tier 3, and the central group's job shifts from approving tools to approving actions: what an assistant may send, spend or publish without a person in the loop. The federated split still holds; the central list simply gets longer and more specific. If you want help building the approved list, the tiers and the agent controls, our agentic AI automation team does this work with marketing and IT in the same room.
Sources
- NIST, AI RMF Core, section 5.1 Govern
- OpenAI, ChatGPT Enterprise
- Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot
- Google Workspace Help, Manage access to Gemini features in Workspace services
- Google Workspace Help, Generative AI in Google Workspace Privacy Hub
