Where to start with AI-powered performance monitoring and optimization

At the November MarTech Conference, Sara Owens, SVP, analytics and data science at ARS X Machina convened a pragmatic panel on AI-powered performance with Anthony Tedesco, global search and programmatic lead, Cisco; Christina Inge, CEO of Thoughtlight; and Reza Moalandin, co-founder, SALT Agency. The conversation stayed refreshingly grounded: less hype, more playbooks.

Perfect data vs. messy data + AI?

Owens opened with a curveball: Would you rather have perfect data and no AI, or messy data with AI?

• Inge didn’t hesitate: Perfect data. “Messy data with AI just makes bad decisions faster.”
• Moalandin took the contrarian slot: give him messy data and AI and he’ll mine value.
• Tedesco sided with clean data, especially for performance marketing: “Data integrity is step one.”

A quick poll set the baseline: About 40% of attendees already use AI to monitor/optimize campaigns; about 60% don’t — yet. Among the “yes” group, LLMs led usage, followed by machine learning, with a notable early cohort experimenting with agentic AI.

The real risk isn’t just ‘hallucination’ — it’s over-optimization

Owens asked where AI can go wrong. Moalandin drew a sharp line between data-led and data-informed decisions. If you blindly let the model chase a metric, you can “optimize to the wrong thing faster.”

Inge added the flip side: Analysis paralysis. Waiting for “perfect” can be as harmful as bad automation. Inaction is a choice — and it usually costs you performance and market share.

If everyone optimizes, is there a ceiling?

Yes and no. Inge argued that baseline optimization becomes table stakes, but differentiation doesn’t disappear — it shifts. The advantage goes to brands that optimize into their niche: Get radically specific about the audience you serve (e.g., allergy-sensitive, gluten-free, sustainability-minded shoppers in CPG) and tune offers, messaging and channels to that micro-need. Optimization then deepens your moat instead of flattening it.

Where to start if you’re at zero

Speaking for a Fortune 100 brand with a cautious stance, Tedesco recommended anomaly detection before automation:

• Build lightweight monitors that alert you to spend spikes, conversion drops or 404s/redirects — well before you let software make changes on your behalf.
• Use LLMs as code co-pilots to stand up simple scripts and alerts (he credited a teammate who did exactly that without a heavy engineering background).
• Plug alerts into your daily workflow (Slack, dashboards, email) and shorten your time-to-intervention first; then consider controlled, rules-based automations.

A tidy human-in-the-loop pattern

All three panelists pressed the same point: keep humans in the loop at multiple checkpoints — strategy, data prep, creative review and deployment.

• Localization/content: Tedesco said Cisco uses automation to draft and scale content, but final market checks stay human to avoid off-brand or off-culture missteps.
• Agentic cautionary tale: Inge shared a data-cleaning automation that accidentally removed valid contacts. The fix was simple — re-insert a human QA pass — but the lesson stuck: Don’t outsource decisions about people to software.

A clever way to reduce support load (and optimize SEO)

Moalandin outlined a practical loop many brands can copy:

1. Transcribe call center audio.
2. Feed transcripts to an LLM to cluster recurring questions.
3. Generate draft answer pages; have editors verify and publish.
4. Ensure those pages are discoverable (SEO) and promoted in help flows.

Result: A measurable drop in inbound support and faster, consistent answers when a sudden spike hits (think: a product bug or a pop-culture mention that triggers new concerns).

Two signals marketers will wish they’d tracked earlier

• Customer lifetime value (CLV), tied back to channel and message. Inge: too many teams still judge success by short-term vanity metrics. With today’s models, you can finally connect campaigns to long-term value.
• “Ghost browsing” by LLMs. Moalandin noted that answer engines fetch and parse your pages without firing client-side analytics. If you’re not inspecting server logs (or using edge tools like Cloudflare’s AI reports), you’re missing how often models consume and quote your content — a rising pillar in AEO (answer-engine optimization).

When automated decision loops actually help

Automate where the math beats the manual:

• High-volume, multivariate problems (e.g., modern bid strategies) where algorithms evaluate far more signals than a human ever could.
• Repeatable, rules-based tasks you can write down as steps before you build (Tedesco’s test: “Can you explain the logic in a one-pager?”). Search-query mining, negative keyword suggestions, or QA checks are good candidates.

Stealable tricks the panel swears by

• Turn bulk builds into push-button work. Tedesco’s team now auto-generates campaign structures and naming using LLM-assisted “bulk sheet” creation, reclaiming hours from trafficking chores.
• Compare models, not just prompts. Moalandin routes the same task through multiple LLMs and evaluates the spread—great for validation and idea generation.
• Create for questions, not keywords. Inge recommends shifting content planning toward the questions people ask (traditional search, social, YouTube, and answer engines). Use SEO tools to surface those questions, then let generative tools help you draft—after you pick the right questions.

The bottom line

AI can give you “eyes everywhere,” but it’s your judgment that keeps performance pointed at the right target. Start with data integrity and anomaly detection. Add measured automation where the economics are obvious. Track the signals that actually matter (CLV today, AEO exposure tomorrow). And keep a human hand on the tiller—especially when the optimization is about people, not just pixels.