- Which permissions does each platform connection request?
- Who can connect and revoke accounts?
- Can access differ by user, region or agency?
Guide · Governance
AI marketing governance for paid media: a guide for mid-market and enterprise teams
Most AI marketing governance frameworks were written for content: brand voice, copy approvals, disclosure. Paid media needs more, because the AI moves money and changes what the ad platforms learn. Govern by action class, not by tool.
Short answer
AI marketing governance is the set of rules, roles and records that decide what AI may do in your marketing without a person. For paid media it covers which account changes run automatically, which need approval, the spend limits and cooldowns around them, and what customer data may be sent to ad platforms.
Why is governing AI in paid media different from content governance?
Short answer
Content governance asks whether an asset is on brand and accurate before it publishes. Paid-media AI also changes budgets, bids, audiences and the conversion signals platforms learn from, often after launch and at speed. So governance must cover money, reversibility and data flows, not just approval of words and images.
Published governance frameworks for marketing AI mostly cover brand voice, disclosure, content review and training. Those still matter when AI writes ads. But an AI that manages Google, Meta and LinkedIn accounts adds three risks content policies don’t address:
- Money moves. A reallocation between regions or platforms is a finance event, not a creative one.
- Learning is disrupted. Significant edits can send campaigns back into learning, so even correct changes cost something if made too often.
- Signals change what platforms optimise toward. Switching the CRM stage sent as a conversion changes who every campaign looks for. It is the highest-impact change an AI or an agency can make, and the least visible.
Autonomy for a single team’s decisions is covered in agentic AI marketing. This guide covers the organisation-wide programme: policy, roles, procurement and multi-team rollout.
What changes between SMB, mid-market and enterprise?
| B2B SMB | Mid-market | Enterprise | |
|---|---|---|---|
| Approvers | Founder or head of marketing | Head of marketing, finance for large moves | Regional and business-unit heads, finance, legal, IT |
| CRM maturity | Spreadsheet or basic CRM | CRM with defined stages | Several CRMs or regions with different stage definitions |
| Entities and regions | One | One to three | Many, often under different data laws |
| Agencies | Sometimes one | Often one per channel | Several, with overlapping access |
| Record needs | Know what changed | Change log reviewed monthly | Audit trail for finance, security and regulators |
| Sensible first scope | One platform, approvals on everything but hygiene | One business line, classes A–C automated | One region or business unit under a written policy |
Original framework
How should you classify AI actions in your ad accounts?
Short answer
Group actions into classes by reversibility, spend impact, evidence strength and scope, then set one control for each class: automatic, by permission, or approval by a named role. Classes survive tool changes, agency changes and new platforms, where tool-by-tool rules have to be rewritten every time a vendor ships a feature.
The Paid-Media AI Change Register is a template to adapt. The classes, limits and approvers below are starting points for your own policy, not any vendor’s internal definitions.
| Class | Covers | Default control | Boundary and cooldown | Approver | Evidence to log |
|---|---|---|---|---|---|
| A. Observe and report | Diagnostics, anomaly alerts, tracking checks | Automatic | None | None | The data behind the alert |
| B. Reversible hygiene | Negatives from a reviewed list, pausing disapproved ads | Automatic | Daily count limit | None; weekly review | Search terms or policy notice |
| C. Budget and bids within limits | Changes inside one campaign or between sibling campaigns | Automatic within limits | A set percentage per period, with a cooldown | Performance lead reviews | Qualified outcomes since last change |
| D. Structural | New campaigns, audience expansion, match types, AI Max settings, new landing pages | Approval | Not during learning | Performance lead | Rationale and expected effect |
| E. Claims and creative | Ads in regulated or technical categories | Approval | Claim checked against a source | Legal or product expert | Source for each claim |
| F. Conversion and data | Which CRM stage is sent, new fields, new platforms receiving data | Approval | Change window agreed with RevOps | RevOps and legal | Data map before and after |
| G. Channel exits and large reallocations | Moves between platforms or regions above a threshold | Approval | Review date set | Regional head; finance informed | Cost per qualified lead or opportunity by platform |
Every class records how a change is undone and who does it. That matters with any tool: in Zephra, every change is recorded in an audit log and there is no automatic rollback, so a correction is a new instruction that follows the same permissions and approvals.
What approvals should AI budget changes need?
Short answer
Classify changes by reversibility, spend impact and evidence. Let small, reversible changes within a set budget percentage run automatically and be logged. Route moves between campaigns or platforms, new audiences, conversion-action changes and regulated claims to a named approver. Add cooldowns so platforms aren’t pushed out of learning.
- Spend boundaries work best as a percentage of a campaign’s budget over a period, plus an absolute ceiling for moves between platforms or regions.
- Cooldowns space out changes to the same campaign. Google lists new bid strategies and setting changes among things that start a learning period (Google Ads Help); Meta says significant edits can reset learning (Meta Business Help).
- Safety stops halt automatic actions when spend or cost per result moves outside expected bounds, often a sign of broken tracking rather than real performance.
- Turnaround targets. An approval that takes ten days turns a good Act decision into a stale one. Set a response time per class.
- Wait is logged too. A decision not to change anything, with its reason and a date to check again, is part of the record.
Zephra supports these controls: low-risk changes can run automatically, everything else follows the permissions you grant or waits for approval, with spend boundaries, cooldowns, learning-phase protection and safety stops. How it works shows the approval flow.
Who should own AI governance in marketing?
Short answer
Marketing leadership owns the policy and spend boundaries. RevOps owns conversion definitions and CRM data. IT or security owns access and vendor review. Legal owns consent and claims. Finance is informed of large reallocations. The performance team owns day-to-day decisions within the policy, including when to wait.
| Decision | CMO | Performance lead | RevOps | Regional head | Finance | IT / security | Legal | Agency |
|---|---|---|---|---|---|---|---|---|
| The policy itself | A | R | C | C | C | C | C | I |
| Class C–D approvals | I | A | I | C | – | – | – | R |
| Class E claims | I | R | – | C | – | – | A | C |
| Class F conversion definitions | I | C | A | C | – | C | C | I |
| Class G reallocations | I | R | C | A | I | – | – | I |
| Vendor approval | A | C | C | – | C | R | C | – |
| Incident review | A | R | C | C | I | C | C | C |
Is sending CRM data to Google, Meta and LinkedIn a privacy risk?
Short answer
It’s a data-sharing decision that needs a lawful basis and minimisation. Platforms accept hashed identifiers and conversion events, not your whole CRM. Decide which stages and fields leave, record consent on lead forms, check India’s DPDP rules, UAE PDPL or US state laws with counsel, and restrict who can change it.
Sending lead stages back is what lets platforms find more qualified buyers. Google takes offline conversions and enhanced conversions for leads through the Data Manager API since 15 June 2026 (Google Ads Help); Meta and LinkedIn use their Conversions APIs (Meta for Developers, Microsoft Learn). Contact details are normalised and hashed before sending, but hashed data is still personal data under most privacy laws.
Govern it as class F: a written data map of which stages and fields go to which platform, sign-off from RevOps and legal, and consent wording on every form that feeds it. Zephra, where configured, sends CRM lead stages to Google through the Data Manager API and to Meta and LinkedIn through their Conversions APIs; security explains how, and the sub-processor list names the third parties involved. For the mechanics, see CRM lead stages for ad platforms.
Procurement
What should we ask an AI advertising vendor?
Short answer
Ask which actions run automatically and which need approval; what spend limits and safety stops exist; which OAuth permissions each platform requires and how to revoke them; what CRM data leaves and how it’s hashed; whether your data trains shared models; and whether every change is logged with its evidence.
Generic AI security questionnaires rarely ask about ad-account permissions, spend boundaries or conversion data. Add these six groups to yours.
- Which actions run automatically, by permission, or after approval?
- What spend boundaries, cooldowns and safety stops exist?
- Are changes held while campaigns are learning?
- Is every change logged with when, what and why?
- Are decisions to wait recorded?
- Can history be exported?
- Which CRM fields leave, to which platforms, and how are they hashed?
- Who are the sub-processors, and where is data processed?
- Is our data used to train shared models? How is it deleted?
- How are duplicate conversions prevented?
- Can we see whether events were accepted or rejected?
- Is Google’s Data Manager API supported?
- Do ad accounts and data stay ours?
- What happens at offboarding?
- What are the terms for agencies and multiple clients?
Red flags: “fully autonomous” as the headline, no change log, vague answers on data, and no clear way to revoke access. Zephra’s answers on access and data are on its security page; agency and multi-client use is sold on Enterprise (pricing). The capability scorecard for evaluating platforms is on AI marketing platform.
How do you roll out AI across regions, business units and agencies?
- Fix measurement. Duplicate or missing conversions will be optimised at speed. See conversion tracking.
- Define the conversion. Agree one qualified stage, defined the same way in every region that will be compared.
- Start with recommend-then-approve in one region or business unit, not all at once.
- Set guardrails before scale. Boundaries, cooldowns and approvers are written down before any class runs automatically.
- Expand with the same register. New regions adopt the register with a regional data annex rather than a new policy.
Agencies work under the register too, with permissions matching their classes. The hiring agencies guides cover the wider relationship. Accounts on Google Search should also plan for the AI Max auto-upgrade, which began in September 2026 for campaigns using automatically created assets or the broad match setting (Google); treat it as a class D change to review. See AI Max for Search.
Worked numbers
A two-region industrial group under one policy
A mid-market industrial group is headquartered in Pune with an entity in Dubai. India spends ₹10,00,000 a month across Google, LinkedIn and Meta; the UAE spends AED 60,000 a month. An agency runs Meta. The policy lets class C changes run automatically up to 15% of a campaign’s weekly budget with a 7-day cooldown, sends moves between platforms above ₹1,00,000 or AED 10,000 to the regional head with finance informed, and requires RevOps and legal for class F.
Class G · IndiaReview
₹1,50,000 from Meta to LinkedIn
Over 8 weeks, LinkedIn’s cost per qualified RFQ is ₹9,000 against Meta’s ₹21,000. The proposal is above the ₹1,00,000 threshold, so it goes to the regional head, who approves ₹1,00,000 with a review in 4 weeks. Finance is informed.
Class C · UAEWait
Six qualified RFQs
A UAE LinkedIn campaign has 6 qualified RFQs since its last change, below the rule for a budget decision. Nothing changes; the reason and a check date are logged.
Class F · AgencyInvestigate
An optimisation event changed without sign-off
The agency switches Meta’s optimisation event to a new CRM stage without RevOps. The monthly change review catches it; the event is set back, and the agency’s permissions are narrowed until the change goes through approval.
A US SaaS company spending $80,000 a month with a quarterly board budget would use the same register, with class G thresholds set by finance’s change control rather than by region.
What goes wrong?
- Budget runaway from bad tracking. Duplicate conversions make a campaign look twice as good, and automation feeds it.
- Agencies changing conversion actions without the owner of the CRM stages knowing.
- Approvals so slow that decisions go stale.
- Governance written only for content, leaving budget and data changes uncovered.
- Pilots across every region at once, so nobody can tell what worked.
- Stage definitions that differ by region, making cost per qualified lead incomparable.
India, UAE and US notes
General information, not legal advice; confirm with counsel.
- India. The Digital Personal Data Protection Act 2023 applies to lead data, with the DPDP Rules phasing in (PIB). Consent notices on lead forms should cover sending stages to ad platforms, and consent should be checked again before automated outreach.
- UAE. Federal Decree-Law 45 of 2021 applies outside DIFC and ADGM, which have their own regimes (UAE Government). A group with free-zone entities may need more than one data annex.
- US. State privacy laws such as California’s (CPPA) affect what can be shared; the TCPA applies if automation triggers calls or texts to leads.
More detail is in lead generation compliance.
What to measure next
- Share of changes by class and path: automatic, by permission, approved.
- Approval turnaround by class.
- Changes reversed, and incidents, per quarter.
- Cost per qualified lead or opportunity by region, before and after rollout.
Enterprise readiness checklist
- Change Register adopted, with approvers named per class.
- Qualified stage defined consistently across regions.
- Conversion tracking checked for duplicates.
- Data map and consent wording signed off by RevOps and legal.
- Vendor questionnaire completed and reviewed by IT or security.
- Pilot scoped to one region or business unit, with a review date.
Questions, answered
AI marketing governance FAQ
01Do we need a CRM to govern AI in paid media?
No. A spreadsheet of lead stages works, provided stages are defined the same way in every region you compare.
02How do agencies fit into governance?
Give them permissions under the same Change Register, and make class F changes such as conversion definitions off-limits without RevOps sign-off.
03Does governance slow results down?
Only if approvals are slow. Automate the reversible classes and set turnaround targets for the rest.
04Should AI ever act without approval?
Yes, for reversible changes within spend boundaries and cooldowns, logged with their evidence. That is what makes approvals for the larger changes manageable.
05What about AI assistants connecting to ad accounts through MCP?
Treat each connector as a vendor: review its permissions, logging and revocation through the same questionnaire.
06How often should the policy be reviewed?
Quarterly, after any incident, and whenever a platform changes how its automation works, such as Google’s AI Max upgrade.
07Can one policy cover Indian and UAE entities?
Yes, with one Change Register and a regional data annex for each legal regime.
Planning a governed rollout
Walk through approvals, permissions and data flows with us.
Book a demo to see how changes, approvals and the audit log work, or review how access and data are handled on the security page. For the pilot region, the free ad account assessment makes no campaign changes and needs no CRM.