A group of researchers connected to Stanford tested something simple: give AI models the exact same question, varying only where the correct information sat inside the reference material provided. The result, published in 2023 and still cited as a reference on the topic, was direct. Accuracy was high when the relevant information appeared at the beginning or end of the material — and dropped sharply when it sat in the middle, even in models specifically designed to handle long texts.
Nothing was wrong with the model in either case. What changed was only the position and quality of the material it received to answer from. It is the same logic any manager has already lived through when asking an analyst for an answer: given the right material, the answer comes out right. Given outdated, ambiguous, or incomplete material, the same competent analyst delivers a wrong answer — just a confident, well-written one.
For anyone evaluating a broader rollout of AI inside the company, that finding flips the question most people ask first. It is not "which AI model is best for my company." It is: what is this tool actually going to fetch to answer with, and does anyone make sure that is the right version?
The perfect answer, built on the wrong document
The problem gets clearer with a common example. Someone in finance asks a corporate AI assistant what the company's travel reimbursement policy is. The AI finds two similar files on the network: one emailed around two years ago, and the version that is actually current, approved three months ago and stored in a different folder. If the search pulls up the wrong file, the answer comes out fluent, well formatted, and confident-sounding — and completely outdated.
No one in that chain made a visible mistake. The AI model processed the text it received competently. The problem happened one step earlier: at the moment the system decided which document to fetch between the two. And that is exactly what the Stanford study measures — when the reference material is poor, ambiguous, or badly positioned, no amount of extra model intelligence makes up for it.
This kind of failure is more dangerous than an obvious mistake, precisely because it doesn't look like one. A visibly odd answer makes people suspicious enough to double-check. A well-written, coherent, confident-sounding answer — built on the wrong document — sails right through, because everything about it looks trustworthy except the fact that it is wrong.
Switching models does not fix the problem
The most common reaction, when an AI answer turns out wrong, is to assume a more advanced model was missing. It is an understandable instinct, but one the Stanford study itself already ruled out: the researchers tested several models, including the most advanced ones available at the time, and all of them showed the same drop when the relevant information was poorly positioned in the reference material. Switching models attacks the wrong part of the problem.
The root usually sits somewhere else, and the numbers confirm the scale of it. Gartner estimates that poor data quality costs organizations an average of $12.9 million a year — a cost that existed before AI arrived in companies and that AI simply makes more visible, because now someone questions that archive in plain language and expects a reliable answer on the spot. Gartner itself predicts that, through 2026, 60% of AI projects unsupported by data and documents ready for the job will simply be abandoned.
There is also a feature of the corporate archive that makes everything worse: according to MIT Sloan, between 80% and 90% of a typical company's information is unstructured — text documents, emails, presentations, meeting notes — and only 18% of organizations surveyed in a widely cited Deloitte study were able to put that kind of content to good use. That is exactly the kind of scattered, unstandardized material any corporate AI assistant needs to consult to answer something specific to the company.
The pattern across all four figures is the same: the AI model is almost never the bottleneck. The bottleneck is knowing, with confidence, which document is current, who owns it, and where it should live.
What has to be in place
A corporate knowledge environment ready for AI rests on concrete, verifiable mechanisms — not on "use a better model."
A document with a defined owner. Every important policy, manual, or procedure needs a person or department responsible for keeping it current. Without an owner, no one notices when content goes stale.
A clear current version, not a pile of similar-looking files. The system the AI queries needs to be able to tell the official current version apart from old drafts, copies, and emailed attachments — instead of treating everything as equally valid.
Scope by department. Not every document should be accessible to every question. An internal legal policy should not surface in an answer for the sales team if it is not relevant and authorized for that context.
Two-step review and approval for critical documents. Before a piece of material becomes an official source the AI draws on, someone other than the author checks and approves it — configurable and logged, applied not to every trivial answer but to what is genuinely sensitive.
Answers that can point back to where they came from. When an answer is grounded in the company's internal knowledge, it should ideally be able to cite the source used — letting the person check the original document before acting on the answer.
This is how Skyller was designed: the company's knowledge comes in with sources, an owner, and a defined scope, and critical documents can go through two-step review and approval before becoming a reference for answers.
From individual guesswork to a trustworthy archive
When the knowledge archive is confusing, every person develops their own way of coping with it. Someone learns, over time, that folder X has the reliable files and folder Y is always out of date. Someone else calls a longtime colleague before trusting any search result. That knowledge about where to trust things is never written down — it only lives in the head of whoever already got burned by the wrong document once.
A corporate archive organized for AI removes that individual burden. The person no longer has to know, from memory, which folder to trust: the system has already answered that question before responding, because it only considers a current, owned source within the authorized scope. The gain isn't just speed — it's consistency: two different people, asking the same thing on different days, get the same correct answer, instead of two answers that depend on whichever file each one happened to find first.
That also changes the most expensive side effect of all: the wrong decision made with confidence. A fluent AI answer built on outdated material tends to get accepted without a second check — precisely because it looks good. Reducing that risk at the source, by fixing the archive instead of just trusting the model, is cheaper than correcting the decision after it has already been made.
A starting checklist
Before expanding AI use over the company's internal knowledge, it is worth reviewing this checklist with whoever manages documents and processes:
- List the company's five most consulted documents and ask who owns each one. If the answer takes too long or varies from person to person, that is already a sign the AI will inherit the same confusion.
- Check for duplicate or outdated versions circulating alongside the official one. Old emails, copies in personal folders, and meeting attachments are the most common sources of confidently wrong answers.
- Define, by department, what can and cannot be used as an answer source. Not all internal content should be available for every question, even when the person asking has good intentions.
- Choose the critical documents that require two-step review before feeding AI answers. Not everything needs that step — only what causes real damage if it turns out to be wrong.
