A few weeks ago, an executive on my team saw a scary number in our paid search data and ran it through an AI reporting agent to figure out what was going on. The agent went into our AdWords account, crunched everything, and spit out a ten page report explaining the collapse of our paid program in great detail.
The problem was that almost none of it was right. It claimed our best performing competitor ad groups were completely missing; it was showing data from campaigns that were paused a year ago; it even contradicted itself multiple times. And the worst part, it didn't take into account any BoF data from our CRM. The AI did exactly what it was told to do, which was generate an authoritative looking analysis. It just had no idea what any of the numbers actually meant in the context of our business.
This is the trap with AI in marketing right now. It is genuinely good at crunching huge amounts of data and surfacing trends. But unless you do real upfront work to teach it what to look at, and unless someone on the other end actually knows how to interpret what it finds, you end up with a confident, polished, ten page document full of crap. And nonsense that looks this clean is worse than no report at all, because it triggers panic and action.
So I built my own skill to do this better, before alarm bells ring. And the real important part: I can analyze and add context before sharing it out.
What the skill does
It runs weekly inside Claude Code, pulls our marketing data, compares it against historical baselines and seasonality, and flags anomalies across the full funnel. Traffic by channel, conversion at every stage (visit to lead, lead to MQL, MQL to SQL, SQL to opp, opp to closed won), lead quality and source mix, velocity and time in stage, CPL and CAC trends, top page performance, form completion, email engagement into MQL creation.
Each flag comes with a short note on what to investigate. Things like:
- "Paid traffic is up 18 percent but MQLs are flat. Form completion rate dropped two points last Thursday. Check the form."
- "Organic conversion rate on the pricing page is down 30 percent week over week, but sessions are stable. Page change or a tracking break?"
- "MQL volume looks healthy but the source mix shifted hard toward one campaign. Check lead quality from that source before celebrating."
- "Email click rates held, but MQLs from email dropped. Likely a downstream issue, not a creative one."
- "CPL is up across paid social with no volume gain. Auction pressure or audience fatigue."
That is the kind of read I used to do by hand when something already felt wrong. Now it happens every Monday whether something feels wrong or not.
Where the human part comes in
The skill is fast and tireless. But it's also missing a ton of context that actually matters. It doesn't know we ran a webinar last Tuesday that boosted organic traffic, but they were mostly customers. It doesn't know sales retooled the SQL acceptance criteria six weeks ago. It doesn't know that we don't want to build campaigns and landing pages against our integration partner, even if there's a lot of search volume there.
The value I add is knowing what the signal actually means. Years of running this funnel, memory of every campaign and how the data behaved when we ran it, a relationship with sales. The skill finds the thread. I decide to pull it or ignore it as just noise.
Verification is the actual work
Early versions of this skill confidently reported numbers that were not in the source data. Invented campaigns. Compared metrics calculated on different definitions and presented the comparison as meaningful. Same failure mode as the ten page AdWords report, just smaller blast radius because I was the only one reading it.
So a big chunk of building this was verification. Cross checking every number against the raw source, pinning definitions, making it say "I can't compare these cleanly" instead of pretending it can. I still spot check outputs every week. When I find a drift, I tighten the prompt or the data handling. The skill I use today is materially different from the one I shipped six weeks ago. That iteration is the job, not a setup step.
The point
There are so many channels, campaigns and touchpoints where the funnel can spring a leak. It's impossible for one person to monitor them all. So I let AI look for anomalies across every data point I have, then I use my expertise and human intuition to find out what's really happening.
AI is a force multiplier for marketers who already know what they're doing but a chaos machine for someone without an underlying knowledge of the details of your marketing engine. Treat it like a very fast junior analyst. It can pull, compare, and flag. It cannot interpret, and it should not be trusted to do so.
If you're building something like this, start with the funnel stages you understand best, verify every output until you stop finding errors, and then keep verifying anyway. The goal is not to outsource the thinking. It's to surface issues before they become visible on the executive level and tell you where to start digging.