AI in Practice: Avoid Real Pitfalls & Make Better Google Ads Decisions

In EM’s latest Lunch & Learn, Mark Harnett shared hands-on lessons from automating parts of his Google Ads workflows with Claude Code. If you’re not this deep into AI, these lessons also translate to a broader use. Mark is an ex-engineer turned marketer turned builder who’s been scaling paid search for 15 years. We heard what actually broke along the way and what it takes to make these tools trustworthy.

Once a problem is fixed, it stays fixed

One key lesson Mark has learned is that AI-built automations get stronger over time. “When I build something and it breaks, I fix it once, and then it stays fixed,” he explained. Human errors, including typos in complex naming conventions and inconsistent judgment calls, are not an issue anymore. It’s a real payoff, even though the upfront cost of finding and fixing problems is real.

Trust is the hard part and it has to be engineered

Mark didn’t shy away from sharing his failures. He described presenting a slide to a client where Claude had fabricated an entire row of data because it couldn’t locate the real numbers. Guess who caught it? His CMO, in a meeting Mark wasn’t even in.

So Mark built verification into the system. Three techniques stood out:

1. Spot-check and interrogate. Ask the model directly: where did this number come from? Have it show its work.
2. Move from probabilistic to deterministic. When accuracy is critical, Mark routes the question out of the LLM and into code (Python via Claude Code), which doesn’t hallucinate the way language models can.
3. Build tests that stick. Every mistake becomes a permanent check. Tell the AI to “make sure you don’t make this mistake again.” Those fixes propagate across his other client projects automatically.

Time-saving workflows; significant maintenance tax

Mark shared several systems he’s built:

• A weekly client reporting deck that used to take 4 hours now runs in about 40 minutes, with him reviewing and “interrogating” the AI-built conclusions afterward.
• A negative-keyword review tool that pulls all search terms, recommends an action with rationale, and lets him approve or override, learning from his corrections over time.
• A Statistical Process Control system (borrowed from manufacturing) that flags when any of ~25 performance metrics per client moves outside normal range, alerting him in Slack before issues sit unnoticed for weeks.
• An automated ad-trafficking system that pulls creative assets from a folder and posts them across platforms in the correct formats and campaign-naming conventions.

Mark was equally direct about the costs. A “time-saving” LinkedIn uploader once broke and cost him six hours to fix. He estimates he’s still “underwater” and acknowledges the maintenance tax of building and debugging these systems. But he’s confident that the gap will close as the tools mature.

Practical advice for getting started

For anyone earlier on the AI adoption curve, Mark’s advice was refreshingly low-tech: “Claude’s really smart. It’ll tell you how to do it. You just have to keep asking.” What he recommends:

• Start with read-only access before granting write permissions
• Pick one AI tool and commit to start building your skills in using it
• Resist the urge to automate everything at once; focus on what matters in the next few weeks
• At the top level of AI instructions, build in a “devil’s advocate” check for moments when you and the AI agree too quickly; something might actually be wrong

Mark’s mantra around AI is that persistence matters more than expertise. The tools are smart enough to fix their own mistakes and find a path forward, but only if you keep pushing. What AI can do now is automate the repetitive parts of the job so your time goes toward the work that actually requires strategic thinking, interpretation, and knowing when a number looks wrong.

Watch clips from the EM Lunch & Learn.