The news
Marketing AI Institute's Mike Kaput covered a real example of a team at SmarterX using OpenAI's Codex — typically thought of as a developer tool — to analyze a 144,000-row, 1,000-column dataset to understand how a piece of content connected to revenue. No formulas. No manual column sorting. The agent was handed a goal and ran its own multi-step investigation.
Our take
This is the right use case framed the wrong way. The article positions this as a story about clever tool repurposing — "surprise, Codex isn't just for developers." But the more important pattern here is what it reveals about the actual bottleneck in marketing analytics.
The problem was never that marketers lacked the analytical curiosity to ask "what content drove revenue?" The problem is that answering that question has historically required either a data analyst on staff, a BI tool someone else owns, or enough patience to hand-wrangle a spreadsheet that crashes your laptop on open. Most marketing teams hit that wall and either settle for a proxy metric or wait for someone in RevOps to get around to it.
What Codex — and tools like Claude Code — actually change is the labor cost of the question. You can hand a messy CRM export to an agent, define the goal in plain language, and get back a reasoned, multi-step analysis without writing a single formula. The agent identifies relevant fields, flags noisy or duplicative columns, runs validation checks, and narrows scope. That's not magic — that's what a good junior analyst does. Except now it runs on a Sunday night and costs you a few dollars in API calls.
The catch: this only works if the underlying data is exportable, reasonably clean, and — critically — you know what question you're actually asking. "Find what's connected to revenue" is a useful prompt because someone already understood their attribution problem well enough to frame it. Teams that haven't mapped their funnel, tagged their sources consistently, or defined what "revenue-connected" means in their CRM will hand an agent garbage and get back a confident-sounding mess.
The tool is ready. The data discipline is where most teams will stall.
So now what?
- Start with a question you already know the shape of. Pick an attribution or performance question you've been avoiding because the data is too messy to touch manually — that's your pilot.
- Use a fully anonymized export. The SmarterX approach is the right one: strip PII before handing anything to an external model.
- Ask the agent to narrate its logic before trusting its output. Have it explain which fields it used and why — that step catches hallucinated column interpretations before they become decisions.
The marketers who get value from these tools first won't be the ones with the cleanest data — they'll be the ones who've gotten specific enough about their questions to actually delegate them.