Here's a stat that should stop you cold: only 21% of marketers say their CRM data is actually ready for AI.
That's it. One in five.
So while everybody's racing to bolt AI agents onto their marketing stack, most of them are feeding those agents a mess. And here's the kicker — AI doesn't clean up the mess. It amplifies it.
I've noticed this pattern everywhere lately, not just in marketing. Give a powerful tool bad inputs, and it doesn't quietly fail. It confidently does the wrong thing, fast, at scale. That's exactly what's happening in B2B marketing right now, and there's a name for the fix: context engineering.
The Problem AI Didn't Create, It Just Exposed
Marketing teams have lived with fuzzy data for years. What counts as a qualified lead? Which channel gets credit for a sale? Why did Salesforce say one thing while the marketing automation platform said something else entirely?
Humans found workarounds. RevOps built spreadsheets. Sales asked questions in Slack. Someone always knew which numbers to trust and which to quietly ignore before the quarterly review.
AI doesn't know any of that. It has no institutional memory, no gut feeling about which field is reliable. As Snowflake's Leo Rodriguez put it, in the old world a data scientist carried that context in their head. In the AI world, you have to spell it out — through governed data, clear definitions, and connected records. If you don't, the AI just runs with whatever's in front of it, inconsistencies and all.
And the inconsistencies are real. One study found CMOs estimate 45% of the data driving their marketing decisions is incomplete, outdated, or just wrong. Meanwhile, 91% of marketers say data readiness is critical to AI adoption. So almost everyone agrees it matters. Almost nobody has fixed it.
Why Coding Agents Succeed and Marketing Agents Struggle
ZoomInfo's Henry Schuck has a good way of explaining why this is so much harder in marketing than in, say, software development. A coding agent works inside a tidy, self-contained repository. A go-to-market agent has to stitch together context from CRM records, emails, buying signals, and conversations scattered across a dozen systems — often duplicated, often missing entirely.
That's not an intelligence problem. It's a context problem.
I find that distinction useful outside of marketing too. I've run my own businesses for years, and I've learned the hard way that a smart tool fed scattered, contradictory information will make confident, scattered, contradictory decisions. Doesn't matter how good the tool is.
Structuring Context vs. Building It
There's a lot of noise right now about better prompts and better models. Fine. But that's downstream work.
Atlan makes a sharper distinction: structuring context is about which information reaches the AI and how it's prioritized. Building context is the unglamorous upstream job of defining, governing, and maintaining that information in the first place. Most companies jump straight to the first step and skip the second. It's like polishing the delivery truck without checking what's actually in the boxes.
The lead-to-opportunity journey — from first pageview to closed deal — is where this shows up most. By the time a deal closes, the story is scattered across Salesforce, marketing automation, and a handful of other systems, each holding a piece. Attribution tries to reconstruct what happened after the fact instead of carrying the real story forward. That takes five things working together: solid record architecture, data integrity, evidence behind prioritization, clear routing rules, and attribution that tracks who was actually involved, early.
The Real Question: Can You Explain Why?
Here's the one that matters most. If an AI agent flags an account as ready to buy, can anyone trace that back to real signals and trusted records? If it reroutes a lead, do people understand the logic behind it?
That's not about turning every calculation into a paragraph. It's about making sure the reasoning isn't a black box.
Every martech vendor is adding AI features right now. Eventually, that stops being a differentiator — everyone will have one. What won't be commoditized is the foundation underneath: clean records, agreed definitions, visible lineage, data people and machines can both trust.
Fix the data first. Build real context around it. Then AI becomes something worth relying on — not just something new to worry about.
Source: CMSWire
Still catching up? Start with Your Smart Oven Might Be Plotting Something
