Skip to content

The AI assistant

The answer is alreadyin your software.

Which invoices are open, where is that contract, which email is still waiting for a reply: you ask in your own words, the assistant reads your software and documents and answers with its source. It reads; it writes nothing.

The problem

The question waiting for the monthly review

Every week the same questions go to the accountant, the office or the IT person. The answer already sits in a program, but someone has to open it, know where to look, or find the time.

Meanwhile the decision waits: the reminder that goes out too late, the contract nobody can find, the forgotten request.

  • Which supplier invoices are still open this month?
  • Where is the latest version of that contract?
  • How many patient requests are waiting for an answer, and since when?

A real example

A real question, a real answer

“Which supplier invoices are still open this month?” Someone in the team asks their assistant. It reads the accounts at the moment of the question, with that person's rights and nothing more.

Today

  • The questionasked in your own words
  • The accountsread at the moment of the question, read-only
DeepFlow
  • A list of documentsnumber, date, amount
  • The source of each linethe document number, to check it
  • Nothing changedany write attempt is refused

The method

How we know it answers correctly

You don't judge an AI on a demo. Before opening access, you give us twenty or so questions your team really asks. We put them to the assistant, compare its answers with what you know, and record every mistake.

As long as it gets common questions wrong, we don't open it. And even once open, it always cites its source: you never take its word for it.

What we connect

What we connect, and what we don't

The assistant only sees what you open to it, and only to read.

01

The accounts

Horus, Pennylane or your software: open invoices, trial balance, general ledger. This is running at a clinic with several companies.

02

Documents and emails

SharePoint, OneDrive, a file server or a mailbox: contracts, procedures, letters, with each person's rights.

03

Requests

Your request-tracking tool: what is waiting, since when, with whom.

04

What we don't connect

Patient and staff data without a written framework agreed with you before the first line, with full respect for medical confidentiality. A payment, never.

Your data

Your data doesn't go just anywhere

Your data stays in your software. The assistant reads it at the moment of the question, through a connector hosted by DeepIT, and each person signs in with their own account: they only see what they already see.

Before opening, a written framework sets out what is connected, who has access, which AI service is used and where it is hosted. You approve it, and you can cut everything off at any time. The details: your data under control.

  • Read-only, twice: the connector refuses any write, and access to the software is also limited to reading.
  • One account per person, never a shared password.
  • Answers that cite their source: document number, file, date.
  • An access you cut off whenever you want.

Getting started

One source first, the rest later

  1. 01

    We choose one source

    Just one to begin with: the accounts, or the documents. Not everything at once.

  2. 02

    We set the framework

    Which data, for whom, hosted where, and what never leaves. Written and approved before we start.

  3. 03

    We test on your real questions

    Twenty or so real questions. We see where the assistant goes wrong and fix it.

  4. 04

    We open access one person at a time

    Each person has their own account and only sees what they already see in their software.

Your questions about the assistant

Can the assistant change my data?

No. Access is limited to reading, and the connector refuses any write before it even reaches your software. A payment, never.

Our team already uses ChatGPT: what should we do?

Start by knowing what they paste into it. Copying an invoice, a contract or a patient record into a public tool takes that data out of your control. An assistant connected to your software reads at the source, with each person's account, without copy and paste, and says where each answer comes from.

Where does my data go?

It stays in your software. The assistant reads it at the moment of the question. The written framework states which AI service is used and where it is hosted: you approve it before opening.

Does it make mistakes?

Yes, sometimes. That is why we test it on your real questions before opening it, and why every answer cites its source, so you can check it. An assistant decides nothing in your place.

Which question do you ask every week?

Tell us which, and we'll tell you whether your data can answer it.