On this pageJump to a section
"The company's revenue grew by 3% over the previous quarter."
Which company? Which quarter? A person reading that sentence inside a report knows. A machine that pulled the sentence out of the report by itself does not. Anthropic used almost exactly this example when it explained why AI systems that search documents so often retrieve the wrong thing: when text gets split into chunks, the chunk loses the context that gave it meaning.2
Now picture the notes field in your CRM. "Wants to start in Q1, budget concern, follow up." Which offer? Which Q1? Who said it? If a person has to guess, a machine will guess too, and sound sure about it.
What is MCP, and why does it matter to a small business?
A new standard, announced last week, for connecting AI to your tools.
The short version
The Model Context Protocol, or MCP, is described by Anthropic as an open standard for connecting AI assistants to the systems where data lives. It lets developers expose their data through "servers" and build AI applications, called "clients," that connect to them.1 The problem it targets is that each new data source has needed its own custom integration, which makes connected systems hard to scale.1
Anthropic released the specification and software development kits, local MCP server support in its Claude Desktop apps and an open-source set of servers for tools such as Google Drive, Slack, GitHub, Git, Postgres and Puppeteer.1 Block and Apollo have integrated it, and several developer tools are working with it.1
What it does not do
It does not clean your data. A standard that makes it easier for an AI to reach your system also makes it easier for the AI to reach the part of your system you have been meaning to fix since 2021. I am mentioning MCP briefly here because it is new. The point of this piece is the work underneath it.
Why does AI make bad data more expensive?
Because it turns small inconsistencies into confident answers.
The chunk problem
Here is the mechanism in Anthropic's research on retrieval. Documents get split into small chunks so a system can find relevant pieces. Anthropic says traditional approaches strip away surrounding context when they do this, which means relevant chunks often go unretrieved.2 Its fix was to attach a short, chunk-specific explanation of context to each piece before indexing it.
In Anthropic's tests, adding that context reduced failed retrievals by 49%, and combining it with reranking reduced them by 67%.2
Treat those figures as Anthropic's result on its own test data, not a guarantee for your business. The lesson travels well, though: pieces of information that carry their own context are easier to find and harder to misuse.
Your business is full of chunks
Every CRM note, support email, call summary and Google Doc is a chunk. If each one states who, what, which offer and when, an AI tool has something to work with. If each one is a fragment, you are asking the tool to read your mind.
Sam Altman, in a clip from the Lex Fridman Podcast posted on March 21, 2024, was blunt about trusting outputs: "any of these models hallucinate some of the time and if it's mission critical, you gotta check it."3 Checking is easier when your records are clean enough to check against. That is the case for fixing the data before you automate anything.
What are the five checks?
Run them in this order.
1. One source of truth
For your offers, prices, policies and processes, pick one document and one place, the kind of written SOP I described in your first AI hire is a writer. Date it. Name an owner. Retire every other copy, or at least stamp it "old." AI that finds three prices will quote one of them.
2. Consistent fields
Look at the values in your CRM dropdowns and tags. If "Facebook," "FB," "Meta ads" and "meta" all mean the same source, merge them. Define your pipeline stages in one sentence each and use them the same way every time (I wrote about stages that tell the truth).
3. Notes with context
Teach yourself and your team a short note format: who, which offer, what they said, what happens next, by when. It takes ten seconds and it makes every later search, human or machine, accurate.
4. Labeled documents
Give each file a clear title, a date and a version. "Proposal final v3 FINAL" tells nobody anything. "Coaching proposal template, 2024-11" does.
5. Clear permissions
Decide what an AI tool may read and what it may not. Client health details, banking information, passwords and anything under a confidentiality agreement stay out of reach until you have a reason, a policy and a person responsible. Easier access is not the same as safe access, and the sandbox logic from my computer use test plan applies here too.
What does a labeled record look like?
Here is the same CRM note written two ways. Neither is real client data, only an illustration of format.
| Note nobody can use | Note with context |
|---|---|
| "Interested, budget worry, call back" | "2024-12-03. Lead from a Meta ad. Asked about the 12-week program. Worried about monthly cost. Send payment options. Call Thursday." |
| No date, no offer, no source | Date, source, offer, objection and next step |
| Needs the author to explain it | Understood by anyone who opens the record |
The second note takes about ten extra seconds. It also means that a person covering for you, a report you run next month and any AI tool reading the record all arrive at the same understanding. Context travels with the record or it does not travel at all.
How do you keep it clean?
A cleanup that is never maintained lasts about a quarter.
A monthly fifteen minutes
Once a month, review three things: new tag or source values that do not match your list, documents created without a date and any offer or price mentioned in a document that differs from your source of truth. Fix them while the list is short.
Make one person responsible
In a small business that person is probably you. Say it out loud anyway. Data hygiene without an owner is a wish.
How do you fix this without stopping the business?
Slowly, on purpose, in a fixed block of time.
- Spend one hour listing every place customer and offer information lives: CRM, email, docs, chat, spreadsheets.
- Choose the single source of truth for offers, prices and policies. Move or delete the rest.
- Merge duplicate tags, sources and stages. Write the definitions down.
- Adopt the one-line note format for all new notes. Do not rewrite old ones yet.
- Label your ten most-used documents with a title, date and version.
- Write a one-page rule for what AI tools may access.
What not to do
Do not connect an AI tool to everything on day one. Start with one low-risk, read-only source, such as a folder of public-facing offer documents, and see what it gets right and wrong. The mistakes it makes will show you which cleanup matters most.
AI is a multiplier. Multiply a clean system and you get speed. Multiply a messy one and you get fast mistakes.
I would not hand an assistant my records until I could explain them to a stranger. That standard is a good one for you too, and if you want help getting there, that is part of what I do.
Keep building.Ekai
Questions people ask
What is the Model Context Protocol?
It is an open standard Anthropic released on November 25, 2024 for connecting AI assistants to the systems where data lives. It aims to replace one-off custom integrations with a single protocol.
Why does my data need to be clean for AI?
An AI tool that reads your systems can only be as accurate as what it reads. Inconsistent field values, undated notes and conflicting documents produce confident-sounding wrong answers.
What is the first thing I should clean up?
Your offers and prices. Pick one document as the source of truth, with a date and an owner, and retire every older copy.
Do I need to buy new software to get ready?
No. Most of the work is naming, labeling, deleting duplicates and writing things down. You can do it in the tools you already use.
Sources
- 1
- 2
- 3
