Four circuit-breakers every autonomous marketing workflow needs

There's a third option, and finance has used it for decades. A circuit-breaker is a limit that sits outside the agent, trips on a condition written down in advance, and can be reset only by a named person. The agent acts freely until it trips one.

01Agents are arriving faster than limits

KPMG's third-quarter 2026 survey found 62% of organizations building, deploying, or developing AI agents, up from 53% the quarter before. Yet 43% have usage or token budgets, and the share building controls into the agents themselves fell to 30%.

Other surveys point the same way. Deloitte reported that only 21% of respondents have a mature governance model for agentic AI, and McKinsey found about two in ten have reached the scaling phase with agents. Kana, which sells an agentic marketing platform, reported that 76% of leaders say their governance is ready for supervised AI decisions. The question and audience differ, so that isn't a contradiction. It does show that self-assessment depends on who asks.

A caution on all of this: these are self-reported surveys, mostly of large US enterprises, and several come from firms that sell advice or software. They show direction, not precision.

02What markets learned about stopping

The New York Stock Exchange halts trading when the market falls fast. An SEC filing describes three thresholds: declines of 7%, 13%, and 20%. The first two pause trading for 15 minutes. The third ends the day.

The useful part is the grading: a small move gets a short pause, a big one gets a long stop. Marketing can borrow that shape, with an alert first, then a pause, then a hold until a person resets it. The mapping is our analogy, not a standard.

The cost of having no breaker is documented. In 2012, Knight Capital's trading software sent more than 4 million orders while trying to fill 212 customer orders, and the firm lost more than $460 million in about 45 minutes, according to the SEC. The SEC cited missing financial risk controls and weak deployment and incident procedures. That wasn't marketing.

We looked for a verified 2026 marketing incident with real spend loss and a named company, and didn't find one. What we found is concern. Bayer's Glenniss Richards told Digiday, “I want a person overseeing the bot.”

03A daily budget isn't a breaker

A budget in an ad platform is a ceiling the platform sets, not one you do. Google says a campaign may spend up to twice its average daily budget on a given day, and up to 30.4 times the daily average in a month. Pacing rules also change: a change to how daily budgets interact with ad schedules took effect on June 1, 2026.

So your own limit has to sit tighter than the platform's, and it shouldn't move when theirs does. That's the core of budget pacing for an agent: you compare spend to your plan, not to the platform's envelope.

04Four circuit-breakers

These four are a practical kind of AI guardrails. They cover the failures we think are most likely in marketing automation, and they come from our own judgment, not from a survey. Set the thresholds from your own history.

1. The pacing breakerStops an agent whose spend outruns the plan. It compares the rate of spend to what you agreed for that hour or day, and to the orders or leads coming in.Trips when: spend runs well ahead of plan with no matching rise in results. Then: pause bid and budget changes and notify the owner. Reset by: the media lead.
2. The margin breakerStops promotions that still look good on revenue but no longer clear a profit floor, for example when costs or inventory shift.Trips when: margin on promoted items falls below the floor finance set. Then: suppress the promotion. Reset by: finance and the brand owner together.
3. The content breakerChecks generated copy and creative against brand and compliance rules before anything ships, and holds items that fail.Trips when: an asset fails a rule or has no approver. Then: hold it in review. Reset by: a named approver.
4. The stale-data breakerPauses automated changes when the data the agent relies on is late, broken, or incomplete.Trips when: a feed or pipeline misses its window. Then: freeze optimization and flag the missing data. Reset by: the data owner, once the feed is confirmed.

Every breaker needs the same three things written down: what trips it, what happens, and who resets it. A breaker with no named person tends to get reset by whoever is annoyed first, and then it isn't a breaker.

The stale-data breaker deserves a word. Google notes that conversions can be reported up to 90 days after the click, so recent performance can look weaker than it is. An agent that optimizes on incomplete numbers can pull back from campaigns that are working. The breaker isn't about waiting 90 days. It's about knowing which data is complete before the agent acts. Flagging a feed that misses its window is a simple form of anomaly detection.

05What an agent should not do alone

KPMG found 49% of leaders have defined high-risk uses where autonomous agent decisions aren't permitted. Write yours down. That list is where human-in-the-loop design starts, and deciding which steps an agent owns is AI workflow design. A reasonable starting list: budget changes above a set size, publishing to a new audience, committing to a new partner, and anything touching pricing or legal claims.

06Where Alli fits

Alli Alerts is built around several of these checks. Investment Guardrails let teams define triggers for budget caps, no-spend alerts, and hourly pacing. Global Hygiene Monitoring checks URLs and product feeds, which covers part of the stale-data problem. Custom Alerts are built with SQL views, so a team can watch other measures in its own data. Notifications go to Slack, where teams already work.

Creative Studio keeps people in charge of storytelling, brand quality, and final approval, while automation handles resizing, formatting, and translation. Governance rules and approvals are set by named roles, which is what '+g('governed AI','governed-ai')+' means in practice.

Alerts notify; a person decides. Whether an alert pauses anything automatically depends on how a team sets up the workflow around it.

07When breakers aren't the answer

Breakers limit damage. They don't make a bad strategy good, and too many of them make an agent useless. If every action trips something, the thresholds are wrong or the agent has the wrong job. Start with the two or three failures that would cost you most.

Pick the one action your agent could take tomorrow that you would regret. Write down what stops it.

08Terms in this article

Agentic AI in marketing

AI agents that research, plan, produce, and monitor work, with people directing and approving.

AI guardrails

Controls that define what AI can access and do, what gets logged, and what needs sign-off.

Human-in-the-loop AI

AI that keeps people involved at key points and approves consequential decisions.

Budget pacing

Tracking spend against budget and timeline so campaigns neither run out early nor underspend.

Anomaly detection

Statistical flagging of metrics that move outside their normal pattern.

Governed AI

AI that works on governed data within set permissions, with audit trails and human approval.

09Frequently asked questions

Sources

1. KPMG, AI's Value Story Sharpens as Organizations Gain Confidence in Governance, Accountability and Workforce Adoption, September 24, 2026 (n=314). KPMG sells AI advisory services.
2. Deloitte, Agentic AI is scaling faster than guardrails, April 24, 2026 (n=3,235). Deloitte sells consulting.
3. McKinsey, The state of AI in 2026: On the road to ROI, August 25, 2026 (n=1,719). McKinsey sells consulting.
4. Kana, New Research: 70% of Enterprises Already Run AI Agents in Marketing, But Ownership Remains Unclear, July 2026 (n=225). Kana sells an agentic marketing platform.
5. SEC, Release No. 34-90134 (NYSE circuit breaker rules), October 8, 2020.
6. SEC, SEC Charges Knight Capital With Violations of Market Access Rule, October 16, 2013.
7. Google Ads Help, About overdelivery and your average daily budget. Undated; Google sells ads.
8. Google Ads Help, Changes to Google Ads budget pacing for Ad Scheduling. Effective June 1, 2026.
9. Google Ads Help, Find out how long it takes for your customers to convert. Undated.
10. Digiday, Marketers put up guardrails as AI agents reshape programmatic buying, May 14, 2026.

Last updated November 6, 2026. Statistics are quoted from the sources above and were checked on that date. Survey results are self-reported unless noted.