One of our primary objectives is enabling our customers to get maximum value from AI, with the most control, and the least friction.

We’ve made a big improvement with our latest feature: fast feedback loops.

One of the most important pain points we’ve encountered while working with customers is enabling feedback loops in their use of AI in business processes.

Feedback loops drive improvements.

Automatic feedback loops drive rapid improvements.

If you can’t think of the feedback loops in your existing workflows, then they are likely implicit, and reactive (or non-existent).

Proactive and deliberate feedback loops lead to better decisions, more deals and efficient scaling.

Making corrections

One of our primary flows is extracting information from raw PDF documents.

We have the ability to provide AI instructions at a workflow level, and we have the ability to define field level nuance.

However, we’ve needed to add information that is specific to certain business situations, at certain points in time.

Two questions we have been facing:

How quickly can users make a correction on an extracted data point? How can we persist and “learn” from the correction?

Without our solution, users are left manually extracting data, and manually correcting fields. This problem plagues almost every financial institution I have ever encountered.

Example: Adjusting EBITDA

Here is a snippet from a customer case study:

A borrower must remain within the Debt Service Coverage Ratio (DSCR) covenant defined in their loan terms. They submit their self-prepared financials on a monthly basis.

The user (a credit analyst) knows that there are specific adjustments to be made based on knowledge they have about that company.

Below are the extracted financials on which the user needs to make a correction:

The analyst provides some further information, in this case adding nuance about a specific business arrangement:

The financials are re-extracted, including a discretionary adjustment field. This is used to run the covenant checks, and the company remains well within their limits.

Critically, this feedback has extracted a new piece of context.

This feedback loop enables future extractions to be run more effectively, and for the system to learn from the corrections made by the credit analyst.

Our system determines how to apply the feedback across the various levels, and with this we can enable highly efficient and accurate document automation.

This solves a problem that plagues all manual extraction processes, all OCR tools and any automation efforts: Applying a nuanced correction for a specific company.

Fast feedback loops

In summary, what we’ve designed is a feedback loop.

You’ve probably seen the following style of interface on modern AI tools. You’ve possibly even clicked on the 👎 and it asked you what was wrong:

This feedback is extremely broad, and designed to improve in aggregate (for you and 500 million other people), rather than to fix exactly what you needed fixing.

We’ve created an extremely tight feedback loop:

• Report an incorrect value
• Explain the correction
• Identify the reason for the correction
• Determine if there is an opportunity to refine the process

So what

Have you experienced business lending software that can learn from a mistake, given only one example?

The past decade of machine learning has enabled feedback loops only when there is enough training data.

Today, we can enable lenders to couple reasoning skills with “one-shot” examples to instantly correct a workflow.

This better enables our company goal: Enabling lenders to increase speed, accuracy, and conviction.

Next steps

We are taking the manual work out of preparing and reviewing deal files.

If you are a Business Lender or Alternative Finance Lender, get in touch.

If you are working with high volumes of business financials, bank statements, invoices, contracts and complex documents, DM me.