Researchers at Stanford, together with colleagues from other institutions, tested something simple: they took a question with a known right answer, placed that answer at different positions inside a long text, and measured whether the AI model could find it. The result, published in 2024 in the academic journal Transactions of the Association for Computational Linguistics, became known as "Lost in the Middle."
The finding: AI models do well when the right information sits at the beginning or the end of the material they're given. In the middle of a long text, performance drops — even in models built for long texts, and even when the right answer is sitting right there, just surrounded by other information.
That changes the question most companies ask about AI. It's not just "is the model good enough" — it's "is the material the company is feeding the AI organized in a way it can actually use." And in most cases, it isn't.
The full manual instead of the right passage
When someone asks a corporate AI what the company's refund policy is and the answer comes back wrong, the first instinct is to blame the model. But in practice, it's common for the AI to have received the entire HR manual — eighty pages covering benefits, vacation, health plan, and refunds — and been left to find one specific paragraph inside it on its own.
The "Lost in the Middle" finding explains why that approach fails: the more irrelevant material surrounds the right answer, the higher the chance the AI "loses" that answer somewhere in the middle, no matter how much reading capacity it has. Today it's easy to confuse reading capacity with reading quality: an AI model may be able to process a giant document in one go, inside what's usually called a "context window" — how much the AI can read in a single conversation — and still read poorly whatever sits right in the middle of it.
The problem gets worse when the company hands over conflicting documents at the same time: the old policy still sitting in the same space as the new one, a draft next to the approved version, the email that "replaced" a process without replacing the original file. The AI has no way to know which one is current — because, at the source, no one decided that for it.
A common example: the sales team updates the price table in a new spreadsheet, but the old one stays saved in the same folder, under a similar name. A customer asks the price of a plan, and the AI — with no way to know which file is right — may answer with the wrong number. No one programmed the AI to lie; it simply received two possible answers, in the same place, with no signal for which one was correct.
Handing the AI everything is not the same as handing it the right thing
The most common fix, once someone notices the AI is getting things wrong, is to hand over even more material — "put it all in there, let it search." It's intuitive, and it's the opposite of what helps. Anthropic, the maker of the Claude model, calls this effect context rot: as the volume of text handed over at once grows, the AI's precision at locating and using the right information drops — not because the material disappears, but because each piece gets less proportional attention within a larger set.
In other words: more material is not more precision. It's more chance the right answer lands exactly in the middle zone that "Lost in the Middle" showed to be the weakest.
Anthropic's own documentation for people building on top of an AI model recommends the opposite of "put it all in": treat the material you hand over as a limited resource with diminishing returns, and pass along the minimal set of information that solves the task — not the maximum available. That holds whether it's a one-off question or an agent working through a series of steps over the course of a day.
What has to be in place
Fixing this isn't about teaching people to ask better questions — it's about organizing the material before any question ever arrives.
A document with an owner. Every policy, manual, or process has a person or area responsible for keeping it current. Without an owner, no one notices when a document goes stale.
A current version, kept apart from the historical ones. The old version can keep existing — for audit, for history — but marked as not current, outside what the AI uses to answer by default. The research firm IDC describes the opposite problem as data decay: content that keeps existing but grows outdated, inconsistent, and less and less reliable, mixed in with what's still valid.
Scope by area. Finance's manual doesn't need to show up in the same search as a sales question. Limiting scope by area cuts down the amount of material competing for the AI's attention when it answers — less material in the middle, the same problem "Lost in the Middle" described, only solved through how the knowledge is organized rather than the size of the text.
An answer that cites the source it used. When an answer can point to which document it came from, whoever receives it can double-check — and whoever maintains the knowledge finds out quickly when a document is being misread or has gone out of date.
A review cycle with an expiration date. A document with no review date tends to be forgotten until someone notices, too late, that it's wrong. Setting how often each category of document needs review prevents that quiet buildup.
This is how Skyller was designed: company knowledge with an owner, a current version, and scope by area, so the answer draws from the right source — not from a pile of documents competing with each other.
The payoff of organizing before asking
The first payoff is obvious: fewer wrong answers. But the more lasting effect shows up somewhere else — in the trust of whoever is using it. Someone who got two contradictory answers from the same AI two weeks apart learns to doubt every answer after that, even the correct ones. Winning that trust back costs far more than organizing the material from the start.
Organizing knowledge up front also speeds up new hires. Someone who just joined has no idea which of the five versions of a process is current — but if the AI already knows, they reach the right answer on day one, without having to learn the company's history on their own first.
There's also a maintenance payoff. When knowledge has an owner and a current version, updating a process becomes a single event — swap the document — instead of a hunt across every corner where an old copy might have been left behind.
And there's a cross-department payoff that tends to go unnoticed. When sales, support, and finance all consult the same knowledge base, with the same current version, the three of them stop giving different answers to the same customer question. That avoids exactly the kind of inconsistency that erodes trust fast: a customer who hears one thing from support and another from finance learns to distrust the whole company, not just one department.
Three questions to bring to your next meeting
Before blaming the AI model for the next wrong answer, it's worth bringing these three questions to whoever manages the company's knowledge:
- How many versions of the same document live in the same place today? If the answer is "more than one" and nothing distinguishes which is current, the problem is organization, not AI.
- Is there someone responsible for each category of document? Without an owner, going stale becomes the norm — and no one notices until a wrong answer travels far.
- Can the AI's answer point to where the information came from? If it can't, no one has a fast way to confirm whether that answer is still correct.
