AI agent for data entry
An agent that reads orders, invoices, forms and delivery notes from email, scans and portals, checks every value against your records and enters the clean ones into your system, leaving doubtful ones to a person.
- 38 orders read from email and messagesDone
- 36 orders entered into the order systemDone
- 2 orders have an unknown customer codeAsks a person
- Entry log saved for reviewDone
An example. Your agent is built around your own work.
How it works
-
Collect the documents
Files arrive from a shared mailbox, a watched folder, a web form or a portal download. Each one is queued with its sender and the time it came in, so nothing is lost and nothing is read twice.
-
Read the fields
The agent pulls out the values you need, such as customer, item codes, quantities, dates, tax and totals, and puts them in a fixed layout. The page and position each value came from are kept next to it.
-
Check before writing
Rules compare the values with what you already hold: known customer, duplicate document number, totals that add up, dates in range. A document that fails a check is not entered.
-
Enter and log
Clean records go into your ERP, accounting software or spreadsheet. Doubtful ones appear on a review screen, with the original beside the extracted values, for a person to correct or approve.
New to AI agents? How they work, in plain words
Real examples, setup steps and tool connections for this agent
An example conversation
An example of how a conversation runs. The grey lines show what the agent does in your systems meanwhile.
- CustomerMorning. The Tuesday orders are in the orders@ inbox, mostly PDFs and a few Excel files. Please enter them in Odoo.
- Read 38 messages from the orders mailbox, 41 attachments (33 PDF, 6 Excel, 2 photos of handwritten orders). Extracted customer, item codes, quantities and delivery dates from each.
- Matched 36 customers to existing records. Checked item codes against the product list: 3 codes not found, 1 order number already entered on Friday.
- Created 33 sales orders in Odoo as drafts. Each has the source file attached and a link to the email.
- Agent33 orders are entered. Four are waiting for you: one is a duplicate of Friday's order 20917, and three have item codes I cannot find in the product list.
- CustomerWhich codes?
- AgentHartmann sent "KB-220", "KB-221" and "KB-2200". The closest products are KB-220 and KB-2200 with a different pack size. I have not guessed. Shall I ask Hartmann to confirm?
- CustomerYes, ask them. And skip the duplicate.
- Marked order 20917 duplicate and left it out. Drafted a confirmation question to Hartmann for approval and put it in the sales inbox.
- AgentDone. The question to Hartmann is waiting in the sales inbox for you to send. I will enter those three lines as soon as they answer.
How it is built
- Input from wherever the documents are
- The agent reads a shared mailbox through its official connection (Gmail, Microsoft 365 or plain IMAP), watches a folder, or receives form submissions. Scanned pages and photos are turned into text with document reading (OCR), and the original file is always kept.
- The language model reads, rules decide
- A language model finds the values in messy documents whose layout changes from sender to sender. It does not decide whether a value is acceptable. Fixed rules do that: totals against line items, tax number format, known customer, duplicate document number, allowed date range.
- A fixed layout, with the source kept
- Every document is turned into the same set of fields. Next to each value the agent keeps the page it came from, so a reviewer can see in one glance where a number was read and why it looks wrong.
- Low confidence goes to a person
- A document that fails a rule, or has a value the agent is unsure about, is not written anywhere. It goes to a review screen with the original beside the extracted values. A person corrects it once, and approved records continue.
- Writing uses the system's own doors
- Records are entered through the target system's API or its import template, for example in Odoo, QuickBooks, Xero, Salesforce or a spreadsheet. The agent stores the document's identity and the new record's ID, so a retry after a timeout cannot create a second entry.
- Systems without an API
- If the target has no API or import, the agent can work in a browser on a dedicated limited account, filling the same screens a person would, and stopping on any page it does not recognise.
- Logging and where it runs
- Every run records the document, the values read, the checks that failed, who corrected what, and what was written. It runs in the cloud or on your own servers. We track corrections per sender, because a sender who changes their layout shows up there first.
What you can connect to it
- Gmail or Outlook mailbox
- Reads incoming orders, invoices and forms with their attachments, and files the messages once they are processed.
- ERP (Odoo, SAP, Microsoft Dynamics, ERPNext and others)
- Looks up customers, products and open orders, then creates sales orders, purchase orders or stock receipts through the ERP's API or import.
- Accounting software (QuickBooks, Xero and others)
- Creates bills and invoices with the document attached, and reads vendor records to catch duplicates.
- CRM (Salesforce, HubSpot and others)
- Creates or updates contacts, accounts and records from forms and applications.
- Excel and Google Sheets
- Reads source tables, appends clean rows as one row per source document, and can serve as the review sheet for doubtful entries.
- SharePoint, OneDrive or Google Drive
- Picks up files dropped into a folder and stores the originals next to the entries made from them.
- Web forms and portals
- Receives form submissions directly, or logs in to a portal that has no API and enters or downloads data on a limited account.
- Scanners and photos
- Reads scanned delivery notes, handwritten order sheets and photos sent by staff or customers.
Connects to your tools. You stay in control.
- Works with what you have
- Gmail, Outlook, shared folders, scanned documents, web forms, Odoo, SAP, Microsoft Dynamics, QuickBooks, Xero, Salesforce, Excel, Google Sheets, websites without an API.
- A person approves what matters
- Payments, orders, discounts or anything you choose wait for approval. Every action is logged.
- Can run on your own servers
- The agent, its data, its logs and even the language model (open models such as Llama, Qwen or Mistral) can run on your servers, so documents and customer data stay inside your network. How that works
What it does not do
- It does not guess a value it cannot read. A smudged quantity or an unknown item code goes to a person.
- It does not enter a record that fails your checks, even when the rest of the document looks fine.
- It does not change or delete existing records unless you have allowed that specific action. By default it only adds.
- It does not approve payments or confirm orders on its own. Entering a document is not the same as releasing it.
- It does not fix mistakes that are in the source document. It flags the mismatch and asks.
Questions
How accurate is it?
It depends on your documents, so we measure it on a sample of your own files before launch. The checks matter as much as the reading: a value that fails a rule is held back instead of entered, and the agent asks when a field is unclear.
Our system has no API. Can it still enter data?
Yes. If the system has an import file, we use that. If it only has screens, the agent works in a browser on its own limited login and fills them the way a person does, stopping on any page it does not recognise.
Can it move data from PDF to Excel, or from Excel into a website?
Yes. It reads tables and fields from PDFs into spreadsheet columns, and it takes rows from a workbook and submits them to a web form or API, logging which rows went in. If the website has an API, we use it, because form filling breaks whenever the page changes.
Can AI do data entry for us at all, or is it too unreliable?
It can, if the language model only reads and rules do the checking. The model finds the values in a messy document. Code then confirms that totals add up, the customer exists and the document is not a duplicate. Whatever fails goes to a person.
What happens when a sender changes their invoice or order layout?
The agent usually still reads it, because it reads the content rather than fixed positions. If the values start failing your checks, the correction rate for that sender rises and we adjust the reading instructions. Nothing wrong is written in the meantime.
Is it safe to send scanned documents containing personal data to an AI model?
It depends on the setup. We can run the agent on your own servers, mask fields it does not need, and use a model provider whose data terms you have approved. Access is limited to the folders and systems the job needs, and every read and write is logged.
Do we need to change our software?
No. The agent enters data into the ERP, accounting software or spreadsheet you already use. If you keep one spreadsheet today, it can fill that spreadsheet.
How much does an agent like this cost?
It depends on the task and the systems it connects to. After a first call we give you a clear estimate of cost and time.
Want this agent built around your work?
Tell us in a few lines what the job looks like today and we will set up an online intro call.
Request sent
We will contact you to set up an intro call.