Search termsAct
Clear waste, strong evidence
A search term has produced repeated RFQs that sales marked as students or job seekers. The agent adds it as a negative inside the reviewed list and logs the leads it based that on.
Guide · AI agents in paid media
Agentic AI marketing: how AI agents run B2B paid advertising, and what to let them decide
An AI agent in paid media is only as trustworthy as the evidence behind each decision. Give autonomy to decisions, not to agents. In B2B, the best agent is often the one that knows when to investigate or wait.
Short answer
Agentic AI marketing uses AI agents that pursue a goal, use tools such as ad platform APIs, act, and check the result without being prompted at each step. In paid advertising, agents can diagnose accounts, change budgets, bids, keywords and creative, and explain why, within the permissions and limits people set.
Definition
How is agentic AI different from marketing automation?
Short answer
Automation follows fixed rules you write: if CPA exceeds a number, pause. An agent works toward a goal, gathers evidence from several sources, chooses among actions and checks outcomes. That flexibility is why agents need permissions, spend boundaries, cooldowns and a decision log that shows what they saw and did.
An agent runs a loop: observe the account, decide what the evidence supports, act through a tool, then check whether the change did what it was meant to. The loop is what makes it an agent. The quality of each step is what makes it safe.
| Rules-based automation | A single generative tool | An AI agent | |
|---|---|---|---|
| What it does | Runs one fixed rule, such as “pause if CPA is above $300” | Does one task on request, such as drafting headlines | Works toward a goal across several tasks and tools |
| Adapts to new evidence | No; it fires whether or not the rule still makes sense | Only inside the prompt it was given | Yes; it reassesses before each action |
| Main risk | Acting on a rule that no longer fits | Inaccurate output that nobody checks | Acting quickly on weak or wrong evidence |
| Control you need | Rule review | Human review before use | Permissions, spend boundaries, cooldowns and a decision log |
The useful question isn’t whether a vendor calls it agentic, but what it may change, on what evidence, and what record it leaves. The wider picture is in what is AI marketing.
How do AI agents access Google, Meta and LinkedIn ad accounts?
Short answer
Through each platform’s official management API, authorised by a person who signs in with OAuth and grants permissions. Outcome data travels the other way: CRM stages go back to the platforms through Google’s Data Manager API and the Meta and LinkedIn Conversions APIs. Some general AI assistants also connect through MCP connectors.
- Management APIs. The Google Ads API, Meta Marketing API and LinkedIn Marketing API let software read performance and edit campaigns. Access is capped by the permissions granted and the connecting person’s role in the account.
- Conversion feedback. Agents that judge campaigns on lead quality need outcomes back in the platforms. Since 15 June 2026, Google offline conversion imports and enhanced conversions for leads uploads go through the Data Manager API (Google Ads Help). Meta and LinkedIn take CRM events through their Conversions APIs (Meta for Developers, Microsoft Learn).
- MCP connectors. The Model Context Protocol is an open standard for connecting AI applications to outside systems. An ad-account connector is a permissioned integration like any other.
See CRM lead stages for ad platforms, and security for Zephra’s connections.
What do platform-native AI agents already do?
Short answer
Each platform now runs agent-like optimisation inside its own walls. Google’s AI Max widens search matching, writes ad text and picks landing pages. Meta can generate text variations and expand images. LinkedIn offers Accelerate ad sets. Each optimises toward the conversion you send it, inside one platform, and none decides between platforms.
- Google AI Max for Search adds search term matching, text customisation and final URL expansion to a Search campaign (Google Ads Help). Campaigns using automatically created assets or the campaign-level broad match setting are auto-upgraded from September 2026, and Dynamic Search Ads follow from February 2027 (Google). See AI Max for Search.
- Meta lists text generation, image expansion and background generation as generative AI creative features. Advertisers opt in and are responsible for previewing the output before ads publish (Meta for Developers). Delivery-side automation is covered in Advantage+ explained.
- LinkedIn offers Accelerate ad sets in Campaign Manager (LinkedIn Help). Check LinkedIn’s current help pages for what each ad set automates before you rely on it.
None of them can see that LinkedIn produced cheaper opportunities than Meta last quarter, or that sales rejected last week’s form fills. Deciding where money goes across platforms, on business outcomes, is a separate job.
Why do B2B ads agents need a different evidence standard?
Short answer
Because B2B outcomes are few and arrive late. A company might see 60 qualified RFQs and 4 purchase orders a month, with stages updating weeks after the click. An agent that acts on form fills alone optimises toward the wrong thing; one that waits for revenue never acts. It needs an early stage that predicts value.
- Sparse outcomes. A few opportunities per campaign a month can’t separate a real change from noise.
- Delay. This week’s numbers are an incomplete funnel: quotes and orders arrive weeks or months later.
- Buying committees. The person who fills in the form often isn’t the one who signs.
A deeper outcome is not automatically a better signal: a purchase order that arrives four months later, four times a month, can’t steer anything. Which lead stage to optimise for sets out how to choose. Whatever stage you pick, the agent’s rule should say how many outcomes it needs since the last change before it acts. Google, for its part, recommends evaluating Target CPA over 30 days with at least 30 conversions (Google Ads Help).
Original framework
What should an AI agent be allowed to change without approval?
Short answer
Changes that are reversible, low in spend impact, narrow in scope and backed by enough recent outcomes, such as adding a clearly irrelevant search term as a negative from a reviewed list. Budget shifts between platforms, audience expansion, AI Max settings, new claims and conversion definitions should need a named approver.
One trust level per agent fails, because the same agent makes a harmless edit at 9am and a risky one at 10am. Set autonomy per decision instead.
| Rung | The agent… | Typical paid-media use |
|---|---|---|
| 1. Observe and explain | Reports anomalies and the evidence behind them | Event delivery failures, CPM spikes, conversion counts that jump overnight |
| 2. Recommend | Proposes a change with evidence attached | A budget move between platforms |
| 3. Act after approval | Executes once a named person approves | New audiences, AI Max settings, new claims in ads |
| 4. Act within boundaries | Executes and logs automatically, inside spend limits and cooldowns | Small, reversible changes with strong evidence |
| 5. Set goals | Changes targets or conversion definitions | Not recommended for B2B spend |
The Decision Autonomy Score
Score each type of decision on four factors, 0 to 2 each. It is a heuristic for setting permissions, not a platform rule.
- Reversibility: 2 if it can be undone in minutes without affecting learning; 0 if undoing it resets learning or can’t be done.
- Spend at risk: 2 if well under your spend boundary; 0 if it moves a large share of budget.
- Evidence strength: 2 if enough outcomes at your chosen stage have arrived since the last change; 0 if the decision rests on clicks or form fills alone.
- Blast radius: 2 for one ad group; 1 for one campaign; 0 for the whole account or a platform.
A total of 7–8 fits rung 4. A total of 4–6 fits rung 3. A total of 3 or less belongs on rungs 1–2, and the agent’s output should usually be Investigate or Wait.
| Decision | Reversible | Spend | Evidence | Radius | Total | Rung |
|---|---|---|---|---|---|---|
| Add a negative keyword for a search term with repeated unqualified RFQs | 2 | 2 | 2 | 2 | 8 | 4 |
| Shift 10% of one campaign’s budget to a sibling campaign | 2 | 1 | 1 | 1 | 5 | 3 |
| Turn on Advantage+ audience expansion | 1 | 1 | 1 | 1 | 4 | 3 |
| Move budget from Meta to LinkedIn | 1 | 0 | 1 | 0 | 2 | 2 |
| Change the conversion action a campaign bids to | 0 | 0 | 1 | 1 | 2 | 1–2; human decides |
In Zephra, low-risk changes can run automatically and everything else follows the permissions you grant or waits for approval, within spend boundaries, cooldowns, learning-phase protection and safety stops; see how it works.
Decision triggers
Why would an AI ads agent decide to wait?
Short answer
Because acting on weak evidence costs money. If a campaign is still learning, a tracking change just happened, or too few qualified outcomes have arrived since the last edit, another change resets learning and hides the effect. A good agent logs the reason and a date to check again.
“Do nothing” is a decision, and it should be as visible in the log as any change.
Cross-platform budgetReview
A move a person should own
Cost per qualified lead on LinkedIn has been lower than on Meta for two months. The agent proposes a budget shift with the numbers attached, and a named person approves or adjusts it.
Lead qualityInvestigate
Cheaper leads, fewer qualified
Cost per lead falls but the qualified share falls with it, or conversions jump right after a tag change. The agent checks placements, search terms, form answers and duplicate events before proposing anything.
LearningWait
Not enough has happened yet
The ad set is still learning, or fewer outcomes have arrived than the rule needs. Meta notes that significant edits can put an ad set back into learning (Meta Business Help). The agent holds and sets a date to check again.
Worked numbers
Can an AI agent run Google and LinkedIn ads for B2B lead generation?
Short answer
Yes, within limits. Agents handle reversible changes well when they can see lead-quality outcomes, not just form fills. Because B2B outcomes are few and arrive late, give autonomy per decision: automatic for small, reversible changes with strong evidence; approval for budget moves between platforms, new audiences and conversion changes.
| Before audience expansion | After audience expansion | |
|---|---|---|
| Spend | $12,000 | $12,000 |
| Leads | 100 | 150 |
| Cost per lead | $120 | $80 |
| Sales-qualified leads | 18 (18%) | 10 (about 7%) |
| Cost per SQL | about $667 | $1,200 |
SQL numbers arrive 14 to 21 days later. An agent reading cost per lead sees a campaign 33% cheaper and scales it. An agent reading qualified stages sees cost per SQL up 80%: it investigates placements, form answers and job titles, then puts a proposal to restrict expansion up for review, because audience settings sit on rung 3.
Indian industrial manufacturer on Google Search, ₹3,00,000 a month
The account normally produces 240 RFQs (₹1,250 each), 60 quote-ready (₹5,000 each), 18 quotes and 4 purchase orders (₹75,000 each). After the September 2026 AI Max upgrade, RFQs rise. The agent:
- Investigate the AI Max search terms behind the new RFQs.
- Act on terms that are clearly irrelevant, such as job and training searches, within the negative-keyword boundary.
- Wait on any budget change until the new RFQs have been through technical qualification, because more RFQs are not yet more quote-ready enquiries.
The same numbers are worked through in cost per qualified lead.
In real estate the equivalent early signal is a site visit booked; in education, an application started. In both, generated claims about registration, accreditation or placements need a human approver.
What can go wrong when agents run ad accounts?
Most failures are ordinary measurement and process problems, carried out faster.
- Optimising a duplicated conversion. A duplicate tag doubles conversions overnight and the agent scales the wrong campaign. Check deduplication first; see conversion tracking.
- Over-editing. Google lists new bid strategies and setting changes among the things that start a learning period (Google Ads Help). Cooldowns exist for this.
- Two optimisers fighting. Let platform AI own in-auction bids; let the agent own inputs and allocation.
- Stale CRM stages. A funnel updated fortnightly looks worse than it is, and good campaigns get cut.
- Consent and claims. Withdrawn consent must stop uploads and follow-up; generated ads can state things you can’t support.
- Unclear accountability. If nobody owns an agent’s area, nobody reads its log.
Practise undoing a low-stakes change before trusting an agent with spend, and add cross-platform coordination last. In Zephra every change is recorded in an audit log; there is no automatic rollback, so you tell Zephra in chat what should be done instead.
How should SMB, mid-market and enterprise teams start?
- B2B SMB: rungs 1–3, with a spreadsheet of lead outcomes updated weekly.
- Mid-market: rung 4 for reversible actions such as negatives, once a quarter passes with none reversed.
- Enterprise: written policy first, then rung 3 in one region; see AI marketing governance.
India, UAE and US notes
- India. Consent under the Digital Personal Data Protection Act can change after a campaign starts, so check it again before automated follow-up or uploads. WhatsApp follow-up delays CRM stage updates, so more decisions should be Wait. See lead generation compliance.
- UAE. GCC B2B audiences are small, so outcomes build slowly and more decisions sit at rungs 2–3. Expect more vendor pitches for agentic tools; ask each one where its decisions sit on the ladder. See GCC B2B paid ads.
- US. Buying committees make SQL and opportunity stages the useful signals. If an agent can trigger calls or texts to leads, check TCPA and state privacy rules first; see lead generation compliance.
Questions to ask before giving an agent access
- Which actions run automatically, which need permission and which need approval?
- What spend boundaries, cooldowns and safety stops apply, and who sets them?
- Does it hold changes while campaigns are learning?
- What evidence does it record for each change, and each decision to wait?
- Which stage does it judge campaigns on, and can that be a qualified or opportunity stage?
- What permissions does it request on each platform, and how do we revoke them?
The full procurement questionnaire is in the governance guide.
What to measure next
- Share of agent decisions by type: Act, Review, Investigate, Wait.
- Changes reversed within 14 days, by action type.
- Time from recommendation to approval.
- Cost per qualified lead or opportunity, before and after each rung change.
- Days campaigns spend in learning each month.
Questions, answered
Agentic AI marketing FAQ
01Will AI agents replace performance marketing managers?
The work shifts rather than disappears: choosing the signals agents judge on, setting boundaries, diagnosing problems the data can’t explain, and deciding what the business should sell. See AI vs agency.
02Is it safe to connect an agent to an ad account through MCP?
Treat it like any integration: check who built it, what permissions it requests, what it logs and how to revoke it.
03Do AI agents need a CRM?
No. A spreadsheet of lead stages, defined consistently and updated on time, works to start.
04How many conversions does an agent need before acting?
There is no universal number. Set a rule for each stage and decision type. For bidding, Google recommends evaluating Target CPA over 30 days with at least 30 conversions, which is a useful reference point.
05Can an agency use Zephra for several clients?
Yes. Agencies buy Zephra on the Enterprise plan; see pricing.
06What happens if an agent makes a bad change?
Find it in the change log, correct it, and ask why the evidence supported it; often it is a tracking or stage problem. Then tighten the permission or cooldown for that action type. In Zephra, you tell it in chat what should be done instead.
Before any agent touches your account
See what the evidence in your ad accounts supports today.
The free ad account assessment reviews wasted spend, campaign structure, performance constraints and measurement gaps across Google, Meta and LinkedIn. It takes 10–15 minutes, needs no CRM and makes no campaign changes.