AI Risk: 2026 Anomaly Detection for Media Buying

AI agents are making media buying way more efficient, but they’re also creating new and complex risks. We’re handing over massive budget control to algorithms, which means we have to get much better at spotting when performance starts going sideways. The big question is, how do you let these autonomous systems run without them spiraling into hugely expensive mistakes?

Let’s do a teardown on a direct-to-consumer (DTC) apparel brand’s Q3 2025 campaign, “Urban Nomad.” The goal was to launch a new line of sustainable outerwear and hit a 3.0x Return on Ad Spend (ROAS). The campaign ran for eight weeks, from July 1 to August 26, 2025, with a $450,000 budget split between Meta and Google. A secondary goal was keeping Cost Per Lead (CPL) for email sign-ups under $15. We used AI agents to handle all the real-time bid optimization, budget allocation, and audience targeting, letting it learn as it went.

The whole strategy was built on letting the AI manage targeting and bidding dynamically. On Meta, the agent was constantly tweaking audience segments using real-time engagement data, building lookalikes from our recent purchasers and high-LTV website visitors. For Google Ads, we let it run wild with Performance Max campaigns, where it optimized bids on the product feed and tested keyword expansion based on what was converting. For creative, we had a mix of polished videos showing the gear in action in the city and static carousels that focused on product features and sustainability. We gave the AI 20 unique creative assets to play with across both platforms, and its job was to figure out which ones worked and shift the spend accordingly.

For the first three weeks, things looked good. We were hitting a 2.8x ROAS, and the CPL was hovering around $16.50. We’d served 18 million impressions and tracked 45,000 conversions between sales and email sign-ups, with the CTR sitting at a respectable 1.8%. The AI agents were clearly working, moving budget between platforms based on what they saw. For example, after seeing a quick spike in conversions from Instagram Reels, the agent reallocated 8% of the Meta budget to that placement within two days. This was the hands-off, fast-moving optimization we were paying for.

But in week four, the first red flag popped up. Over a 36-hour window, the Cost Per Conversion (CPC) on our Google Ads shot up by 35%, going from a $25 average to $33.75, while the number of conversions didn’t budge. The AI tried to react by lowering bids, but the cost just stayed high. This wasn’t a conversion problem, it was a cost problem. Our internal anomaly detection system, which we have set to trigger when any metric moves more than 2.5 standard deviations from the 7-day rolling average, lit up. Of course, eMarketer’s 2025 forecast shows ad spending is always going up, so more competition can raise costs, but a 35% jump out of nowhere means something is broken.

We had to dig into the decision logs ourselves. It turned out the Performance Max agent had decided to get “creative” and started bidding aggressively on a huge list of broad, non-brand keywords. Think “winter jacket” and “warm coat”, terms with tons of volume but terrible intent. It was an attempt to find new audiences that went way too far, burning money on useless clicks. The agent’s learning model saw a temporary bump in search volume for these generic terms and mistook it for a massive new conversion opportunity. The anomaly system did its job and flagged the cost spike, but only a human could look at the logs and see the stupid mistake the AI had made.

Our immediate fix was to build out a negative keyword list at the account level to rein in the Performance Max campaign. While PMax doesn’t give you perfect control, this was the best lever we had. We also tweaked the AI agent’s core parameters, telling it to focus on conversion *value* instead of just conversion *volume* when it was exploring keywords, and we tightened its lookback window for performance data. Specifically, we changed the weighting algorithm to penalize any keyword with a conversion rate under 0.5% over a 72-hour period. This is the part of the job that reminds you the “set it and forget it” promise of AI is still marketing fluff. You absolutely have to keep an eye on these things and tune the parameters. That part’s not optional.

Then in week six, a second, sneakier problem showed up on Meta. Our CPL for email sign-ups started climbing, going from $12 to $18, a 50% increase, even though our main ROAS number for purchases looked fine. This was tough to spot initially because the overall campaign health seemed fine, masking the fact that one of our key goals was failing. Our anomaly detection, which was focused on the primary ROAS goal, didn’t even flag it as a critical issue. We only found it when we started segmenting performance by conversion type. A recent IAB report on the State of Data in 2025 talks about this, and our campaign was a perfect example of why you need to watch multiple metrics.

Looking at the Meta data, we saw the AI had fallen in love with one of our video creatives, “Mountain Trailblazer.” It started showing it almost exclusively to audiences that had a history of high video completion rates but who rarely, if ever, converted on an email sign-up. The creative itself was more of a brand piece and didn’t have a strong email CTA. The AI, chasing a vanity metric like video views, matched a piece of creative to an audience in a way that actively hurt our CPL goal. The video’s high view-through rate looked great in the dashboard, but it was a misleading signal.

To fix this, we did a quick creative swap and re-jiggered the AI’s priorities. We pushed a new set of static images into the mix that were designed for one thing: getting email sign-ups, with clear value props and big CTAs. Then we went into the Meta Campaign Budget Optimization (CBO) settings and adjusted the agent’s objective function to put a much higher weight on email sign-up conversions within those specific ad sets. This forced the agent to deliver ads to people likely to sign up, even if they were less likely to watch a full video. It worked. Within four days, the email CPL dropped back down to $13.50.

By the time the eight-week campaign wrapped up, the “Urban Nomad” line hit a final ROAS of 3.1x and an average CPL of $14.20, so we beat both our goals. Total impressions hit 32 million with 85,000 conversions. The average CTR landed at 2.1%. The budget ended up being 60% on Meta and 40% on Google. With a total spend of $450,000 generating $1,395,000 in revenue, the 3.1x ROAS was solid. The campaign proved that AI agents are powerful but they’re not mind readers. They need constant supervision and a human hand on the tiller.

The system worked because we had strong anomaly detection in place and a team that knew how to interpret what the AI was doing and why. Setting up the AI agents from the start let us scale fast and make complex bid adjustments a human team could never keep up with. That real-time feedback loop between the data and the agent’s algorithm was, for the most part, very effective. And the AI’s ability to find winning creative and shift budget to it automatically is a huge advantage. For example, that “Mountain Trailblazer” video, while bad for CPL, actually drove a great ROAS for purchases. Just turning it off would have been a mistake, the right move was to redirect it, not kill it.

Our biggest mistake was assuming the AI could be totally self-sufficient. These things are always balancing exploration (finding new things) and exploitation (cashing in on what works), and if you don’t set the right boundaries, they can get lost exploring. The agent blowing the budget on generic keywords on Google or chasing video views on Meta are perfect examples of that. We also learned that our anomaly detection needed to be more granular. A good top-line ROAS can easily hide a fire in one of your specific conversion funnels.

For our next steps, we’re building out the anomaly detection to use multi-metric thresholds and some predictive modeling. The goal is to predict these problems before they happen based on early signals, like a slight dip in conversion rate for one audience, instead of just reacting after the CPC has already spiked. We’re also looking at AI frameworks that have a better “human-in-the-loop” process, where the AI can suggest a big strategic shift but needs a human to approve it before it can pull the trigger on the budget. This kind of hybrid model, with the AI doing the heavy lifting and a human providing strategic oversight, feels like the only sustainable way to manage complex media buying in 2026.

Proper AI agent risk management isn’t a feature, it’s a process. It requires being proactive and data-obsessed, finding the right balance between letting the machine optimize and having a human ready to intervene before a small error becomes a costly disaster.