AI document search finds a file from a question in plain words, by reading what the file says rather than matching its name. For a firm with client files, the method that holds up is to find the client first, then let an AI agent read the documents on that client's record and answer with the file.
AI document search: find a client file by asking
AI document search reads what a file says, not its name. See how full-text, OCR and semantic search differ, and why to look for the client before the file.
The Standards team
In short
What is AI document search?
AI document search is software that finds a file from a question such as “the signed fee agreement for the Hollis household” by reading the text inside your documents, not only their file names. Google's AI Overview for “AI document search”, read in October 2026, describes it as finding files “by meaning and content rather than exact filenames or keywords”. It stacks four layers. Full-text search matches the words a document contains. OCR, optical character recognition, turns a scanned PDF or a phone photo into text those searches can read. Semantic search compares meanings, so a question about a “power bill” can surface an “energy statement”, the example Filex AI gives in its guide updated on 19 October 2026. An AI agent then opens the candidates, reads them and answers with the file attached. A tool that skips a layer misses a class of documents: without OCR, every scanned ID and signed form stays invisible, and without a link to the client, you get another household's file.
Full-text, OCR and semantic search: what each one finds
Full-text search finds the words a document contains, OCR makes scans and photos readable, semantic search finds documents close in meaning to your question, and an AI agent reads the file to answer it. For a firm that keeps client files, the layer you need depends on how those files arrive. Typed letters and exports answer to full-text search, while signed forms, scanned IDs and phone photos of statements stay invisible until OCR gives them text.
Layer | What it finds | What it misses | When you need it |
|---|---|---|---|
File name search | Files whose name holds your words | Everything named | When your team names files with discipline |
Full-text search | Files containing your exact words, often with typo tolerance | Scans with no text layer, and documents that use other words | Typed documents: letters, exports, Markdown notes |
OCR | The text of scanned PDFs and photos, made readable | Nothing by itself: it feeds the other layers | Any firm that receives signed or photographed documents |
Semantic search | Documents close in meaning to your question | Exact identifiers such as a policy number, which keyword search finds better | Questions phrased in your words, not the document's |
AI agent reading | The answer inside the document, with the file | Documents it was not pointed to, or cannot open | Questions that need a fact from the file, not only the file |
Find the client first, then the file
For a firm with client files, the reliable path to a document runs through the client: two households can both hold a “2025 tax return”, and a semantic match on the wrong one is worse than no match. Standards attaches every document to one record, such as a contact, or a household or a contract if your firm keeps them, and files it in that record's drive. So the Standards assistant looks for the client before it looks for the file, with a record search that tolerates typos, and its answer names both the client and the file, which you can open. The method works whether the file was named with care or arrived as IMG_2231.jpg, because the assistant reads the text of the scan rather than its name. The tools Google shows for “AI document search” in October 2026 search a library of files you upload or connect: thedrive.ai and docAnalyzer in the results, Filex AI in the AI Overview.
The assistant answers from your own records, the idea behind a company brain. The Standards documentation on files and documents describes how a document attaches to its record and its drive.
How to find a client document from a plain-language question
You find a client document in Standards by asking the assistant in your own words, from any page of the workspace. The assistant does the searching: it finds the client, opens that record's drive and reads the extracted text of each scan before it answers.
You attach each document to its client's record, by dropping the files on it, up to 50 MiB each, or let the assistant save an email's attachments onto the right client from a connected mailbox.
You ask the assistant a question in your own words, such as “find the latest proof of address for the Okafor household and tell me its date”.
The assistant searches your records for the client, by name, email or any text field, and picks the household or contact that matches.
The assistant reads that record's drive, its folders and files, and selects the documents whose titles or types fit the question.
The assistant reads the text of each candidate: Standards extracts it by OCR for PDFs and JPEG, PNG or WebP images, and reads plain-text files as they are, on demand if a file was never processed.
The assistant answers with the fact you asked for and links the document, so you open the file and check the page it quoted.
Your developers can also read the text Standards extracts from a document's files outside the assistant, from the API or the command line, as the documents and files command reference shows.
What an AI agent does with the file once it finds it
Once a Standards agent has found a client document, it can copy what it reads into the client's record, such as the expiry date of a passport or the policy number on a signed contract. It can also check the document against what the record already holds. Filing is part of the job too: the agent creates folders in the record's drive, moves or renames a document, and attaches a file already in Standards. From a connected mailbox, it saves an email's attachments as files on the right client. What it changes without asking depends on the mode you choose. Two limits hold in both modes: the agent asks you before it deletes anything or sends an email out, and a client record you keep private stays out of your colleagues' searches and out of their agents' reach. A deleted document goes to the trash with all its files, and you can restore it.
Mode | Before a change to your data |
|---|---|
Manual | “Asks for your approval before each change to your data” |
Auto | “Changes your data without asking”, except deletions and emails sent out |
What Standards searches today, and what it does not
Standards does not yet search the text inside documents across the whole workspace, and its meaning-based search covers mail, not documents. Each place you search in Standards, as built in October 2026, matches something different, so a question about a scan goes to the assistant and a question about an email goes to the mailbox.
Where you search | What it matches | By meaning? |
|---|---|---|
The search bar (⌘K) | Records by their fields, in a few letters, typos included | No |
A record's drive (the loupe in the Documents tab) | Folder and file titles, up to five levels deep and 200 results | No |
The mailbox | Subject, body, people and attachment names | Yes, fused with words, over the 20,000 most recent emails, unless the workspace turned it off |
The assistant | Records, then the text it reads inside each document it opens | It reads; it does not rank documents by meaning |
The text of a scan is extracted for the assistant to read; it does not feed the search bar. The mail search documentation details how words and meaning combine for email.
Example: a wealth advisory firm finds a client's proof of address
Harbor Lane Advisors is a fictional twelve-person wealth advisory firm, used here as an example. A compliance review in October 2026 asks for the current proof of address of every client renewing in the fourth quarter. An adviser asks the Standards assistant: “for the Moreaux household, find the most recent proof of address and tell me the date on it”. The assistant finds the Moreau household record despite the misspelling, reads its drive, and opens the two documents filed under identity. One is a utility bill dated April 2025, scanned from paper; the other is a phone photo of a bank statement dated September 2026, saved as a JPEG. The assistant reads both through their extracted text, answers with the bank statement and its date, and links the file. In Manual mode, it then proposes to write the date into the household's “proof of address date” field, and the adviser approves. The Standards page for wealth advisers covers the rest of the firm's client work.
Questions
Before you switch tools.
Yes. Standards does it from the client: you ask the assistant in your words, it finds the client's record, reads the documents in that record's drive, including scans through their extracted text, and answers with the file linked. A misspelled client name still finds the record, since the record search tolerates typos.
For a firm whose documents belong to clients, Standards finds the client first, then reads the documents on that client's record and answers with the file, so the answer comes from the client you named. Tools such as thedrive.ai, Filex AI or docAnalyzer search a library of files you upload or connect, by what they say.
Yes, once OCR has turned the image into text. Standards extracts the text of PDFs and of JPEG, PNG and WebP images by OCR, and of plain-text files such as CSV or Markdown directly. The assistant reads that text when it opens a document. Word files, Outlook messages, zip archives and HEIC or TIFF images get no extracted text, so save Word files and messages as PDF, and photos as JPEG, if you want them read.
Normal file search matches names and exact words. AI document search reads the content: OCR makes scans readable, semantic search matches meaning, and an agent opens the files and answers the question. In Standards, mail search fuses words and meaning, while documents are found through their client and read by the assistant.
With Claude, yes, through Claude Code, and with OpenAI through Codex; the Standards documentation does not cover ChatGPT itself. The documentation names three AI tools that connect to a workspace over MCP: Claude Code, Codex and Cursor. Connected, they search your records and read documents, scans included, with the rights of the key you give them, and a read-only key keeps them from changing anything. A private record stays invisible to anyone you have not shared it with, agents included.
It is safe when the tool keeps each client's documents behind that client's access and says what it sends to an AI provider. In Standards, a private record's drive answers as if it did not exist to anyone without access. Every PDF and image goes to Mistral for OCR, and its text to the assistant's AI model. For meaning-based mail search, each email's text goes to an embedding provider through the Vercel AI Gateway, which the workspace can switch off.
Read next
On the same subject.
Ask for a client's document in your own words
On a first call, Standards finds one of your clients, opens the documents on that client's record, scans included, and hands you the file.