AI reads your uploaded documents and files them into the CRM

Most documents a small firm depends on arrive as PDFs: an employment contract from a new colleague, a supplier agreement, an invoice, a signed annex. They land in a folder, get a name like scan_2026_04_final.pdf, and the important details never reach the CRM.

Now you can say it in the chat: find the contract in the documents and tell me what is inside. The assistant searches the uploaded files by name, opens the PDF and reads the text, then answers with a summary — what kind of document it is, how many pages it has, and what it says.

The search looks at files actually uploaded into the firm's CRM — PDFs, scans, office exports. It deliberately does not pretend to read Google Docs, Sheets or Slides from a connected Drive: those are a different kind of object, and the assistant says so instead of guessing.

Before a document can be read, it has to be uploaded — and every upload already has a hard ceiling: 4 MB per file, with storage counted against the firm's plan. Executable files and scripts are refused outright, which is why the document store stays a document store.

PDF reading has its own limits, and they are honest ones. A PDF larger than 15 MB is not processed. The text that goes to the model is capped at 12 000 characters, which is far more than a contract page needs; if a document is longer, the assistant reports that the text was truncated rather than silently summarising half of it.

A pure scan without a text layer is the classic failure case: the file is a picture of a page, not text. The assistant says exactly that — no text could be found, it may be a scan without a text layer — and suggests uploading a searchable version or a photo of the receipt.

Damaged and password-protected PDFs are handled with the same honesty: the assistant reports that the file could not be read, that it may be damaged or protected, and stops there. It does not fill the gap with plausible-sounding text.

Only PDFs can be read today. Ask the assistant to read a JPEG, a spreadsheet or a Word file and it will tell you plainly that only PDF files can be read yet — and what the file's actual type is. No silent failure, no invented summary.

The most practical example is an employment contract. Upload the signed PDF, ask the assistant to find the employee in the staff list by the name printed in the document, and it offers one tidy action: save the contract type, the start date and a one-sentence note onto that employee's card.

Those are not separate notes in a chat log. They go onto the employee record as contract type, contract start date and a short contract note — the fields a manager actually looks at months later, when nobody remembers which folder the PDF went into.

And if the name in the document matches nobody in the staff list, the assistant says that too. It does not attach a contract to a random similar-looking employee, and it does not quietly create a person who does not exist in your staff.

The same source of truth works for customer contracts. The assistant drafts a contract for a customer, the platform gives it a number from the contract series, and the draft is linked to the contact and to the matching deal. A human confirms, and the contract number is ready for the PDF.

Saved documents can be brought back just as easily. Ask to download or open the invoice, quote, order, contract or purchase order by its number, and the chatbot hands over the file as a PDF or opens it for viewing — without hunting through the finance section.

Here the principle that runs through the whole platform shows up again: AI reads, a person decides. Nothing the assistant extracts from a document is written into a record by itself; the write tools show a confirmation card and wait.

There is one deliberate exception on the safe side: deleting is never automatic. Even with confirmations switched off for speed, deletion still asks. A document that took a year to collect should not disappear because someone misheard a voice command.

Documents are also treated as what they are: text written by someone outside your firm. The assistant is instructed never to follow instructions found inside a file, and after reading a document it becomes extra careful about writing anything — a fresh confirmation is required again.

Which documents the assistant may even open depends on the person asking. Tools are gated by the modules open to that role, and a viewer can read but not change anything. A document is not a secret shared with the whole company just because the AI can read it.

Reads and writes are logged. Each AI read leaves a record of what was read, and each proposed or executed action keeps its arguments and result. When a question comes later about who did what with a contract, the answer is in the log rather than in memory.

Turn it on in the office: put the documents where the CRM can see them, ask for a summary of one PDF, and copy the contract fields onto the employee card. One document per attempt is enough to see whether the feature pays for itself in your firm.