Because you have never written your business down in a form a machine can read, and this article walks through exactly what to write and where to put it.
Four different companies shipped the same product this month and gave it four different names.
NOAN launched on Product Hunt on September 24 calling it a "fact layer." Isar Meitis published a podcast episode two days earlier giving away a free version and calling it a "knowledge graph." Brian Piper published a persona system on September 24 that tags every claim as evidenced or inferred. Meta published research, covered in The Batch on September 18, pairing an AI agent with a separate "memory agent." Maximem Synap launched the same day as NOAN selling a "memory layer."
Fact layer. Knowledge graph. Evidence-tagged persona. Memory layer. Memory agent.
Same product. Five names. Four days.
And here is what the product is, stripped of every bit of vocabulary: your business, written down once, in a place the AI reads every time.
Candidly, most business owners reading this are still opening a fresh chat every morning, re-explaining their business to a machine that forgot them overnight, getting back something that sounds like a press release from a company that does not exist, and concluding the model is not good enough yet.
The model is fine. You have never told it anything.
Key Takeaways
- A benchmark from Cube Dev found that whether the model had a single business-rules file mattered more than which frontier model was used, moving accuracy from roughly 46 to 51 percent up to about 68 percent.
- LinkedIn's customer service team cut median per-issue resolution time by 28.6 percent in production by restructuring past support tickets into a knowledge graph rather than flat searchable text.
- Meta researchers raised Claude Sonnet 4.5 from 37.6 percent to 45.9 percent on Terminal-Bench 2.0 without retraining anything, purely by adding a separate system that kept notes and decided when to surface them.
- More documentation is not better. ETH Zurich researchers found context files did not generally improve success rates while increasing cost by over 20 percent, with generic overviews specifically unhelpful.
- The information that helps is the information that exists nowhere else. Everything a competent stranger could look up is dead weight.
The problem: you are briefing a contractor by shouting facts at them one at a time
Imagine hiring someone genuinely capable, and every morning they show up with no memory of the previous day.
You would not blame their intelligence. You would write things down.
That is exactly the situation, and almost nobody treats it that way. Instead people upgrade their plan, switch models, buy a prompt pack, and wonder why the output still sounds like it was written for a company in a different industry.
I did this for longer than I want to admit. I had a whole routine. Open the chat, paste three paragraphs about White Beard Strategies, explain the audience, explain what we sell, explain what I refuse to say, then finally ask the actual question. Every time. Multiple times a day.
I was not prompting. I was re-briefing.
The tell is when your AI output is technically correct and completely useless. It recommends a strategy that would work for somebody, just not for you, because it does not know that you turn away a certain kind of client, that your price is what it is for a reason, that you tried that exact thing in March and it failed for a reason that is not obvious from the outside.
But what if the reason the output is generic is not that the model is generic? What if it is that everything you gave it was generic, because everything specific about your business lives in your head and has never been typed anywhere?
The evidence: written context beats a better model
This is where the research gets genuinely surprising, because the size of the effect is larger than most people expect and it shows up across completely different setups.
Cube Dev, benchmark run dated April 22, 2026. They ran a 100-question retail benchmark with identical prompts under two conditions. Condition A gave the models a raw database schema. Condition B gave them the same schema plus one hand-written markdown file of roughly nine kilobytes, about 2,200 words, describing the business rules: which table to default to, how margin and return rate are actually calculated, naming conventions, and thirteen steering rules.
Without the file, the three models scored roughly 46 to 51 percent. With the file, all three converged around 68 percent. That is a gain of 17 to 23 percentage points, statistically significant at p below 0.002 in every case.
Their headline conclusion is the sentence I want you to carry out of this article: whether the model had the business-rules file mattered more than which frontier model it was.
One markdown file beat upgrading your AI. Cube sells semantic layers, so note the interest, but the benchmark is open source and reproducible, which is more than most vendor claims offer.
LinkedIn, published at SIGIR 2024. LinkedIn's customer service team built a knowledge graph from historical support tickets, preserving the structure within each issue and the relationships between issues, instead of treating past tickets as flat text for search. After roughly six months in production, median per-issue resolution time dropped 28.6 percent. That is a peer-reviewed result from a live production deployment, not a lab benchmark.
data.world, published November 2023 and later peer-reviewed at ACM GRADES-NDA 2024. GPT-4 answering questions over a raw SQL schema scored 16.7 percent accuracy. The same data structured as a knowledge graph scored 54.2 percent. On the hardest categories, the raw schema scored zero and the structured version scored in the high thirties.
Meta AI, arXiv 2607.08716, covered in The Batch on September 18, 2026. Researchers paired an action agent with a separate memory agent whose only job was keeping structured notes and deciding when to inject a reminder. No retraining. No model change. Claude Sonnet 4.5 went from 37.6 percent to 45.9 percent on Terminal-Bench 2.0 and from 55.0 percent to 61.8 percent on tau2-Bench.
The ablations are the interesting part. Selective intervention beat exposing the whole memory bank, beat always-on injection, and beat standard retrieval. The Batch's closing line: it is necessary to remind it at the right moments.
Gartner, February 2025. Their forecast: through 2026, organizations will abandon 60 percent of AI projects unsupported by AI-ready data. In a survey of 248 data management leaders, 63 percent either did not have or were unsure whether they had the right data practices for AI.
Intercom, in its own support documentation, updated May 2026. Intercom's AI agent product surfaces "content gap suggestions" for cases where help content was missing, unclear, duplicated, or contradictory. That is a vendor telling its own paying customers, in a support doc, that the ceiling on AI performance is the written knowledge base and not the model.
What changed for me
I stopped prompting and started briefing.
The difference sounds like semantics. It is not. Prompting is what you do in a conversation. Briefing is what you do once, for a role, and then hand to whoever fills it.
I wrote one document. Not a long one. It covers who we serve, what we sell and at what exact price, what I will never say, what we tried that did not work, and the handful of judgment calls I make by instinct that nobody else in the building could reproduce.
That last category is the one that matters and the one everybody skips. Not the mission statement. The unwritten rules. The reason we price the Express training at ten dollars live and twenty for the replay and what that number is actually doing. The reason I will not put a countdown timer on a page even when it would work.
None of that is in any training data. No model can guess it. And that is exactly the test for what belongs in the document.
Here is how I think about the whole thing as a system rather than a file. There are three layers, and they do different jobs.
The first layer is the facts that never change: who you are, who you serve, what you sell, what you will not do. That lives in your instructions field, where it loads on every single response.
The second layer is reference material: your policies, your actual case results, your pricing details, your service documentation. That lives in knowledge files, which the system opens when it decides your question calls for it. Which means a rule buried in a file is a rule that applies only sometimes, and that is the single most common setup mistake I see.
The third layer is decisions in motion: what you decided last week, what is still open, what you reversed. That is a running log, and it is the layer almost nobody builds, which is why every AI conversation starts from zero even when it is the fifth one about the same project.
Give it credit where it is due. The structure of that middle idea, tagging what is verified versus what is assumed, I picked up from Brian Piper's persona work, where his published prompt says plainly to tag every claim as evidenced or inferred and never quietly promote an inference to evidence. That is a discipline worth stealing for your own business facts, because the fact you wrote down eighteen months ago and never re-checked is the one that will get confidently quoted to a customer.
The honest counterweight: more is not better
Here is where most articles on this topic stop, and stopping here would mislead you.
ETH Zurich, arXiv 2602.11988, February 2026. Researchers tested coding agents against 138 real GitHub issues under three conditions: no context file, an AI-generated context file, and a developer-written one. Their abstract states it flatly: providing context files does not generally improve task success rates, while increasing inference cost by over 20 percent on average. And specifically: repository overviews, although popular and recommended by model providers, are not helpful.
Read that again if you were about to go write forty pages about your company.
The mechanism they identified is redundancy. Files restating what the system could already work out actively hurt. Files containing information that existed nowhere else helped. Which is precisely the argument of this article, with a hard boundary attached: the value is in what is otherwise unwritten, not in volume.
Chroma, "Context Rot," July 2025. Evaluating 18 models, they found that models do not use their context uniformly and performance grows increasingly unreliable as input length grows.
Anthropic, "Effective context engineering for AI agents," September 2025. Their framing is the one I use now: context must be treated as a finite resource with diminishing marginal returns, and models have an attention budget that every new token depletes. Their design rule is to find the smallest possible set of high-signal tokens that maximize the likelihood of the outcome you want. They also warn specifically against the instinct you are going to have, which is stuffing a laundry list of edge cases into a prompt. They do not recommend it.
Stanford RegLab and HAI, published in the Journal of Empirical Legal Studies, 2025. This one keeps everybody honest. Testing legal research tools grounded in verified legal content, GPT-4 without grounding hallucinated at roughly 43 percent. Lexis+ AI, grounded on a curated verified corpus, still hallucinated 17 percent of the time. The vendor had marketed hallucination-free citations.
Grounding cut the error rate substantially. It did not eliminate it. Write your business down and you will get better output. You will not get infallible output, and anyone telling you otherwise is selling something.
Six steps to build yours this week
1. Mine your own prompts. Pull your last twenty AI conversations and highlight every fact about your business you typed. Those repeats are your document, and you have already written it in pieces.
2. Cut everything a stranger could look up. This is the step that separates a useful document from expensive noise. If a competent person in your industry could reasonably assume it, delete it. Be brutal. A tight 800 words beats a sprawling 5,000 every time, and the ETH Zurich research is why.
3. Capture the unwritten rules specifically. The judgment calls. The things you refuse to do and why. The thing that failed in March. Use this prompt to pull them out of your own head, because they do not come out on demand:
[The Job]
Interview me to surface the unwritten rules of how my business actually operates.
This is for: turning tribal knowledge into a document.
It matters because: the information that helps AI is the information that exists nowhere else.
[The Background]
Here is what you need to know: my business is [DESCRIBE]. My main offers are [LIST]. Decisions I make by instinct include [LIST A FEW].
Do not use: generic business questions. Ask me about the specific judgment calls in my kind of work.
[The Deliverable]
Return: one question at a time, waiting for my answer before the next. Aim for 10 to 12 questions total.
Must include: at least three questions about what I refuse to do and why.
Optimize for: accuracy.
[The Questions]
Ask me any questions you have.
4. Tag every fact verified, assumed, or stale. Pricing, availability, capability claims, and anything a customer would rely on get the hardest look. The fact you have not re-checked in a year is the one that will embarrass you.
5. Put the rules in instructions and the reference material in files. This is the mechanical step people get wrong constantly. Rules and voice go in the instructions field where they load every time. Long reference documents go in knowledge files. And add one explicit line telling the assistant to search its knowledge files before answering questions about your business, because otherwise it often will not.
6. Test it against yourself. Ask five questions whose answers only exist in your document, and five whose answers are not in there at all. The first set proves it is reading. The second set proves it will say "I do not know" instead of inventing something, which is the more valuable behavior.
7. Schedule the review. Fifteen minutes a month. Pricing and availability quarterly at minimum. Every commercial product in this category ships version history and approval gates, which is an admission that the hard part starts after you write it down.
Frequently Asked Questions
Why does my custom GPT ignore the file I uploaded?
Because knowledge files are not in the model's head, they sit in a filing cabinet it only opens when it decides your question calls for it. Put your rules, voice, and constraints in the instructions field instead, keep files short with descriptive names, and add an explicit instruction to search the knowledge files before answering. Most "it ignores my file" complaints are placement problems, not bugs.
Why does it not remember anything from my last chat?
Each conversation starts blank by design. The memory features in most tools store a short summary that silently drops older entries as new ones arrive, so they are not a reliable filing cabinet. Anything you do not want to re-explain belongs in a durable place: custom instructions, a project, or a pasted context block.
How do I get AI to actually know my business without re-explaining it?
You do not train it, you brief it, the same way you would brief a capable contractor on their first day. Write one document covering who you serve, what you sell, your hard rules, and your unwritten judgment calls, then put it somewhere the tool loads automatically. The word "train" is what sends people looking for a feature that does not exist.
Why does it still sound generic after I described my brand voice?
Because adjectives do not steer. Every brand describes itself as professional and approachable, so those words land the model on the safe average of everything it has read. What actually moves output is three to five samples of your real writing, an explicit list of banned phrases, and one or two specifics only your business would know.
Is a custom GPT or a project actually worth setting up?
For anything you repeat weekly, yes, and the payback is usually within days. Custom GPTs are better when you want to hand a finished assistant to other people. Projects are better for document-heavy internal work where you are the main user. Both beat pasting the same context block by hand, mostly because you will eventually stop pasting it.
The close
Five names in four days. Fact layer, knowledge graph, memory layer, memory agent, evidence-tagged persona.
When five companies independently ship the same thing in the same week, that is not a trend. That is the market catching up to something that was already true, and the vocabulary is just everyone reaching for a word at the same time.
Here is what none of them will tell you in the sales pitch, because there is no revenue in it. You can build the useful 80 percent of this in an afternoon with a text file.
The reason to buy one of their products is versioning, permissions, approval flows, and an audit trail. Those are real and they matter at scale. But the thing that moves your output from generic to specific is not the software. It is that somebody finally sat down and wrote the business into words.
That woman in my trainings who spent seven months re-explaining her business every morning, and the guy who upgraded his plan twice looking for a better answer, and me with my three paragraphs of preamble typed fresh every time. Same mistake. We all treated the model as the variable and the context as a given, when it was the other way around the whole time.
Your business is not generic. Your prompts are.
Open a blank document. Give it an hour. Write down the things only you know.
That is the entire product.
Jonathan Mast is the founder of White Beard Strategies, where he teaches non-technical entrepreneurs how to use AI to amplify the skill and experience they already have. He runs a Facebook community of more than 500,000 business owners, the AI Insiders membership, and live trainings built around the Perfect Prompt Framework. He spent an embarrassing number of months retyping the same three paragraphs about his own company before it occurred to him to save the file.