Skip to content

Verification#

Automatic Verification#

While we provide state-of-the-art extraction results, we cannot guarantee that our AI is perfectly accurate. Therefore, we encourage you to implement additional cross-checks of our extractions against your business data as outlined below:

Cross-Checks#

The following three cross-checking approaches may be applicable for your business:

  • Master Database: If you have a database with core data, you can match our extractions against it. One could for example:
    • Do an exact match between the tax ID extracted by the natif API and the tax IDs stored in a database containing all of your vendors
    • Do a fuzzy match between the vendor name extracted by the natif API and the vendor names stored in a database containing all of your vendors
    • Check if the items we extracted from your invoice can be found in your ordering system
  • Confidence: We do not just provide the values we extracted, but always pass along a confidence value. The closer this value is to 1, the more certain we are that we extracted the correct value. Thus, you can decide that you trust our extractions if they have reached a minimum confidence threshold.
  • Logical Checks: You can implement simple rules that you expect to be true based on your use case. E.g., you can check that a certain identifier matches a known pattern. You could also decide that you do not want to fully automatically process documents if the risk of doing so would be too high (let's say invoices with a gross amount of more than 10,000 Euros).

Combining Several Verifications#

Ideally, you combine several different verifications across multiple extraction fields.

As a positive example, where verification helps to increase the automation rate, consider an invoice where the vendor's address, IBAN, and VAT ID were extracted with high confidence and are contained in your master database. In that case, having a low confidence for the vendor name probably does not matter, as you already have a strong indication of who sent you that document.

As a negative example, illustrating how combining several verification methods could reduce the automation rate and save you from profit loss, consider a high-risk document (according to the due payable amount) where the extracted IBAN has high but not perfect confidence and the VAT ID is not contained in your database. In that case, you probably want a human to manually check the document.

Strictness of Verification#

Naturally it depends on your use case how strictly you want to check our output. If you want to grant a loan of a million Euros, you probably always want a human to verify the information we extracted. Note, however, that this verification can be done much quicker than fully manual processing, as all fields can be pre-filled by our AI.

If the risk is low (let's say an invoice of 2 Euros), you might want to process it fully automatically even if you do not fully trust it, since the effort of manually checking exceeds the cost of manually checking the document.

Outcome of Verification#

Overall, your checks can have two outcomes:

  1. The document can be processed fully automatically
  2. The document should be verified manually by a human

Human Verification#

If you decide that a human should manually inspect the outcome of our AI, you can either use your own interfaces or embed our interfaces into yours.

Using Your Own Existing Interfaces#

Using your own existing interfaces obviously has the advantage that your employees can work the way they are used to. Simply pre-fill all input boxes with the values our AI extracted for you and ask your employees to verify them. To make their lives easier, you can of course highlight fields were the automatic verification failed in red and fields where the automatic verification succeeded in green, such that less manual work is needed.

Using the natif.ai Stand-Alone Interface#

Since many of our customers do not have the resources to build their own interfaces, we made it easy to use our interfaces and integrate them into your workflow.

This works as follows: When the outcome of your automatic verification is that a human should verify the information

  1. Create a share token by passing the document_id (uuid) and an expiry_at (date) to the route /sharing/documents (see the documentation). Thus, you basically use your existing authentication (bearer token or API key, see the docs) to create a token with which only a single document can be accessed (until a certain date). This makes the document sharing more secure, as the token does not grant access to anything but this single document. The response of this POST request contains a token to be used in the next step.
  2. Using this token and the known document_id, build a URL with the following pattern: https://platform.natif.ai/workflows/extraction/stand-alone/<document_id>?token=<token>
  3. Open this url in any browser, edit the document, and save the results (with the top-right button). The natif.ai stand-alone extractions interface
  4. After saving, fetch the results again from our API, and automatically process the document with verified information.

Probably you do not want to break the user experience for your employees, i.e., you want to prevent them from switching between your system and a browser all day long. Therefore, we recommend that you open the above URL within an <iframe> in your application.

To make the integration even more simple, our application listens to events. Thus, you can place a button outside of the <iframe> that does the following:

  1. Save within our application by sending a save-document event
  2. Re-fetch the updated results from our API and trigger automatic processing

A minimal working example looks as follows:

<div>
    <h1>
        My website
    </h1>
    <iframe
    id="iframe"
    style="width:100%!important; height:80vh!important;"
    src="https://platform.natif.ai/workflows/extraction/stand-alone/<document_id>?token=<token>"
    ></iframe>
    <br />
    <button onclick="myCustomSaveFunction()">My Save Button</button>
</div>
<script>
    function myCustomSaveFunction() {
        const frame = document.getElementById('iframe');
        frame.contentWindow.postMessage("save-document", 'https://platform.natif.ai')
        // trigger fetching of data and automatic processing
    }
</script>

If you do not want to show the cookie banner, you can accept cookies by passing the URL parameter "allowCookies=essential" or "allowCookies=all", resulting in an overall URL 

https://platform.natif.ai/workflows/extraction/stand-alone/<document_id>?token=<token>&allowCookies=<your_choice>