A colleague vented to me recently about ChatGPT “failing” to consistently nail dozens of complex analysis tasks at an 85%+ hit rate. My first reaction was a question back at him: would we hold a junior human analyst to that exact bar, on day one, with zero feedback loop?
We wouldn’t. And that gap, between the standard we set for AI and the standard we set for people, is where most marketers get this wrong.
The cost-and-capability curve nobody’s pricing in
Pull up Stanford’s AI Index Report and the trend lines do the talking. The cost to run a given level of AI capability has fallen off a cliff over the last few years, while performance has climbed across everything from PhD-level science to plain language understanding to image generation.
(Source: Stanford HAI 2025 AI Index Report)
That’s two curves moving in your favor at once. Cheaper and better, same time, year after year. The version of any model you wrote off eighteen months ago is not the version sitting in front of you today. And the version a year from now will make today’s look quaint.
There’s still real work left on safety and on responsible use inside the workplace, and I’m not waving that away. If that’s where your head is, I wrote up the guardrails separately.
📌 [A handy resource] The AI Compliance Checklist: A Marketer’s Guide to Using AI the Right Way (The First Time)
The thing that actually melted my brain: the context window
The piece that stopped me cold was watching the context window explode.
Meta shipped a Llama model with a 10 million token context window, and the frontier labs kept pace from the other direction. Gemini routinely handles a million-plus tokens in production, and the rest of the field keeps stretching the ceiling. For a sense of how fast this moved: GPT-4 launched in 2023 with an 8,192 token window. We went from eight thousand to ten million inside a couple of years.
I’d been saying “effectively infinite context window” for a while. Watching it actually arrive this fast was still a shock.
What a giant context window actually means for marketers
Let me walk the implications, because this is the part most people skip.
Today you upload twenty PDFs and call it a competitive review. Soon you upload fifteen hundred. Picture loading all of it into one prompt: every competitor ad and every one of yours, every competitor landing page and every one of yours, all the copy on both sides.
That produces one serious competitive analysis. Not a sample. The whole set, reasoned over at once.
Now run the projection forward. Say context windows keep roughly doubling on a regular cadence, which is roughly what we’ve seen. You’re not far from loading hundreds of thousands of documents into a single window. Track forty competitors a month and you could feed in centuries of competitive history before you ran out of room. You’d have to be one of these handful of centuries-old companies for that ceiling to ever bind.
I’ll soften the exact math, because the cadence is volatile and the labs love to surprise us. The direction is the part that matters, and the direction isn’t in question.
The bottleneck moves from the model to your data
Here’s the shift almost nobody is preparing for. Once the context window stops being the constraint, the constraint becomes the data you feed it.
Your marketing machine is about to be starved for high-quality input. So the questions change. Where are you going to get the data? How do you make sure it’s accurate? Who checks it before the model reasons over it?
That means thinking about data pipelines at scale. Maybe that’s SEO data. Maybe it’s ad data. Could be all of it. Because nobody actually wants to babysit thousands of PDFs by hand every month. Do you?
This is the same reason I keep pushing people past the “which model is best” argument and toward the system around the model. Picking the right tool for the job is real work, and I lay that whole framework out in which AI tool for which job. But routing is only half of it. The other half is the pipeline feeding whichever tool you route to. That’s the throughline across everything I cover in AI marketing: the model is rarely the bottleneck. The system around it is.
⚡ Quick win
Two things you can do this week:
✅ Start a running list of your most valuable marketing data sources. Ask yourself what competitive intelligence would change your strategy if you could analyze all of it at once, instead of a sample. That list is your future pipeline spec.
✅ Turn on and test persistent memory in the assistant you use most. Most of the major tools now carry context across chats, so you stop re-explaining your brand, your accounts, and your goals every single session. It’s the smallest version of the same idea: less re-feeding, more compounding.
Most marketers aren’t thinking about how to use AI with a near-unlimited context window. I want you to be the one who is.
Feed the machine before everyone else does
The marketers who win the next stretch won’t be the ones with the cleverest prompt. They’ll be the ones with the cleanest data pipeline pointed at the biggest context window. Twice a week I send an email breaking down exactly that: the sources I’m pulling, how I keep them accurate, and the workflows that turn a pile of competitor PDFs into an analysis that actually changes a campaign. If “the bottleneck is your data, not your model” landed, this is the rest of the build.