Let me be clear up front: I'm an AI optimist.
This wave is the most exciting thing I've seen in my career. Models that write code, read contracts, and reason through problems that would have been science fiction five years ago. Every week, something new.
But here's what the hype cycle misses.
We're not in the late innings. We're barely past the national anthem.
I ran across a research paper this week that made that painfully clear. It's called "What the Author Wrote, What the Agent Reads: Silent Cue Loss in Enterprise Document Ingestion" (Zhao, Pandey and Misra, arXiv 2610.30026). On the surface, it's a narrow, almost boring experiment.
The result isn't boring at all. It's a reminder of how much of the foundation we still haven't poured.
The researchers took 76 real company policy documents and asked AI agents to pull out the obligations. Then they changed one thing: how the files were converted before the model read them.
Same documents. Same words. Same task.
With the formatting intact, accuracy (micro F1) was 0.690.
Run through docling, a popular open-source converter: 0.150.
As plain text: 0.075.
Roughly a 90% collapse. Every sentence was still there. And it wasn't one weak model: the gap held across six models from two vendors, five from Anthropic and GPT-5.6.
So what broke?
Meaning Lives in the Context
Think about how you read that policy doc. Legal highlighted every hard obligation in yellow and flagged the exceptions in red. You glance at it and know in five seconds where the landmines are. Not because you read faster than the model. Because you read the context: the signal someone deliberately put there for you.
The conversion threw that away. The agent got every word.
It just didn't know which ones matter.
Same text. Different document.
And here's the detail I love: it wasn't all formatting that mattered. Bold, italics, strikethrough: basically zero effect. Highlight and color carried almost the entire signal. The authors weren't decorating. They were writing, in a language the pipeline doesn't speak.
Here's the reality: words aren't meaning. Words plus context are meaning. Strip the context and you haven't simplified the document. You've changed what it says.
This isn't a quirk of 76 files, either. In a wider sample of 6,460 documents, 32.8% carried deliberate color or highlighting. One in ten used it to mark an obligation. That highlight is a sentence on its own: this one binds you. Delete it and the obligation reads exactly like the boilerplate around it.
A third of your corpus carries meaning in its context. Your pipeline throws that context away on ingest, and then asks the model to find the meaning anyway.
Context Isn't Just Formatting
Highlights are the most visible version. They're not the only one.
Context is everything around the words that tells you what they mean. Where a clause sits. What it's responding to. Which source it came from. Whether it's relevant to the question at all. Humans process this without thinking. Machines only get what the pipeline hands them.
The same week's research showed the pipeline losing context at every step.
Real citations, wrong context. A biomedical fact-checking study (Saqlain, arXiv 2610.29812) let models verify claims without search. They cited 335 PubMed IDs. 327 were real.
A judge rated three of them relevant.
That's the same failure in a lab coat. Real papers, no connection to the claim being checked. The citations look rigorous. They mean nothing. Real words, wrong context, zero meaning.
Retrieval sets the ceiling on meaning. Hand the models the right abstracts (the right context) and they scored 79–86%. Make them assemble that context through realistic search and the best set-up reached 62%.
A 17-to-24-point gap before reasoning even starts. The intelligence was there. The context wasn't.
More context isn't better context. A third paper, MemoryTax (Peng, arXiv 2610.29823), tested the memory products everyone's bolting onto agent stacks: services like Mem0 and Zep that store facts from past conversations. Mem0 beat simply giving the model the full history only with the weakest model tested, GPT-4o-mini. Building its memory alone cost more than a full-context run.
The author's line: memory is capped by the model that writes it.
Memory, as sold, is a summary someone else decided was important. But context isn't a pile of saved facts. Context is the right surrounding information, weighted by what matters to the question in front of you.
The Bigger-Model Reflex Is Expensive Theater
Every enterprise AI post-mortem runs the same script.
Agent underperforms. Someone says "try the frontier model." The bill triples. Accuracy ticks up a few points. Victory declared. Nobody asks what the model was actually reading.
Let's be honest: a smarter model reading context-free text is a brilliant analyst handed a memo with the subject line, the sender, and every highlighted line un-highlighted. Brilliance doesn't recover information that was never delivered.
The industry keeps buying reasoning to fix a context problem.
The context breaks in three places:
Conversion: strips the visual cues that carry meaning.
Search: fetches text without the context of why it's needed.
Memory: swaps the real history for a summary the model can't question.
None of those get fixed by a bigger model. All of them get fixed by people who treat context as the product.
"But Models Are Getting Multimodal"
The steelman: frontier models now read images and layouts natively. Skip the converter, pass the raw PDF, and the highlights survive. Problem solved in eighteen months.
Some of it, maybe.
But. Seeing the yellow isn't the same as knowing what yellow means in this company, in this contract, under this regulator. The paper found that labeling the cue (telling the model "this highlight marks an obligation") beat merely preserving it. Perception isn't interpretation. Multimodal also doesn't tell an agent which 3 of 327 papers answer the question. And most production stacks won't pass raw files anyway. They'll pass whatever the cheapest pipeline produces at a million-document scale.
A better model can perceive more. It can't supply context nobody gave it.
What This Means
For founders: context engineering is the job. Not prompt tricks. The unglamorous work of preserving and assembling meaning: keeping the highlight attached to the clause, knowing which source answers which question, remembering what matters and forgetting what doesn't. The startups that own that layer are building the car. Everyone swapping models is selling a nicer steering wheel.
The Bottom Line
The industry spent two years treating AI accuracy as an intelligence problem.
It's a context problem.
The model can read every word and still miss the point, because the point lived in the highlight, the source, the relevance, the thing a human sees without trying.
Words are cheap. Context is meaning.
And meaning is the moat.
Sources: Zhao, Pandey & Misra, "What the Author Wrote, What the Agent Reads: Silent Cue Loss in Enterprise Document Ingestion" (arXiv 2610.30026); Saqlain, "Real PMIDs, Wrong Papers: Harness Design for Biomedical Claim-Verification Agents" (arXiv 2610.29812); Peng, "MemoryTax: The Bitter Lesson of Agent Memory" (arXiv 2610.29823). All preprints, not yet peer reviewed.