From messy forms to clean data with AI-powered extraction → Unstract.com

From messy forms to clean data with AI-powered extraction

Conducted on 13th August 2025

AI + Human-in-the-loop = 99% Accurate Data Extraction (Here’s How)

Transcript

[00:00:00]
Hi everybody. Hope you’re all doing well today. Thank you so much for joining this session on how AI and Human Intelligence together can give you 99% accurate document data extraction. I’m Mahashree, Product Marketing Specialist at Unstract and also your speaker for this webinar today. So we’re all aware of how rapidly AI is growing today, paving the way for seamless automation, even in complex business workflows.

But in many cases, especially in document extraction use cases, human intervention in these workflows still remains an anchor to the overall success of the process. So in this session we’ll cover an overview of human in the Loop deployments, especially in document extraction pipelines, how untract supports human review as a feature, and also what is the impact that we see out of it.

Now, before getting started, I wanted to quickly go over a few session essentials or layout, few ground rules for the session. So [00:01:00] all attendees will automatically be on mute throughout the course of this webinar. In case you have any questions, please do drop them in the q and a tab at any time during this session and we’ll be able to get back to you with the answers via text. Now you can also interact with fellow attendees using the chat tab.

This is also where you’ll let us know in case you run into any technical difficulties during this webinar. And as a final point, when you exit the session, you’ll be redirected to a feedback form where I request you to leave a review so that we can continue to improve our webinars going forward.

Agenda

So that said, here’s the agenda for today to give you a quick bird’s eye view of what you can expect. So we’ll start off with the human in the loop overview, where we’ll cover what human in the loop is, why we need it, and where it can be deployed in the entire document extraction pipeline. This will be followed by the core chunk of the session that is the live demo of HITL, or human in the loop in Unstract.

So we’ll see how we can set it up, how to define access permissions, and also how to go about the entire review process. Finally, before we conclude, we’ll also go over a few HITL best practices, and in case we have any questions remaining, we’ll also be venturing into a q and A session where one of our experts will be on air to answer your questions.

So that said, let’s kickstart this session.

Understanding Human in the Loop (HITL)

[00:02:30]
Human in the loop or HITL at its core is the practice of bringing human judgment in AI driven document processing workflows. Now that is the simple definition of it. What actually makes this truly powerful is that it enables better data quality, stronger compliance with regulations, and also greater trust in your automation.

Now, AI today can automate a lot, and it is only improving at a rapid pace. But again, AI is not infallible, and that is why human review becomes essential. There are still edge cases, low confidence extractions, and high-stakes scenarios in document extraction, which are just way too important for it to go wrong, and that is where human review becomes essential.

Document Processing Pipeline

So where do we exactly place this in the entire document extraction pipeline? [00:03:00]
So what you see over here is a typical document processing pipeline. Again, there are various use cases, but I’ve just taken a generic pipeline over here, so you get that document influx from your source. You might wanna classify your document and based on the document type, you would route it to the different document workflows that you have, and after that, you would extract the relevant data from the documents, prepare it, and finally store it in the destination. So this entire process can be automated by AI. So where do we insert the manual review without disrupting this process?

Now, there isn’t any hard and fast rule about this. You can go about it depending on what your business requirements are, but what we commonly see is that one of the main areas where human review becomes important or relevant is right after document extraction. You might want to add a human over. You might want to have a human to oversee how the extraction has been done to verify where the data has come from and whether it is correct. And you might also want this person to change certain values in case they’re wrong before it goes into the destination.

And/or you can also have it right at the beginning where you’re looking to classify your document and you might want human oversight to see where your documents are landing up, which workflows they are directed to, and all that.

Importance of HITL

So now that we’ve seen what human in the loop is and where it can be inserted in the entire document processing, let’s just go over why HITL is really required today. [00:05:00]
So we already touched upon this, but to reiterate, we see two major implications where HITL becomes important: accuracy and compliance. In elements today when the advanced models can sometimes misinterpret difficult documents, it could be ambiguous layouts, misclassified documents, or even poor scanned quality. When this happens, human review becomes the safety net to catch these misses before they end up having an impact on the downstream consequences.

Secondly, having a human in the loop is often a compliance requirement in highly regulated industries like finance, insurance, or healthcare. Regulations such as GDPR, HIPAA, and financial reporting standards don’t just demand accurate data; they demand traceable and also auditable processes. Including human checkpoints in your automation pipeline would reduce this risk and also create a defensible audit trail to prove that diligence was followed while extracting data from these documents.

And beyond just accuracy and compliance, there’s often a third overlooked benefit that is trust. So when stakeholders know that there’s a layer of human oversight, they’re more likely to adopt and trust the automation system that you have in place. So to sum it up, HITL helps bring, helps bridge the gap between rapid automation and responsible AI.

Untract Overview

Now that we’ve taken an overview of HITL, let’s see how this can actually be deployed in practice using Unstract. [00:06:30]
Untract is an LLM powered unstructured data ETL platform. If I had to briefly bucket the capabilities of the platform, I’d have two main categories: that is the development phase and the deployment phase. In the development phase, you would upload your documents for extraction—this is done in a prompt engineering environment called Prompt Studio, where you would upload your documents, define prompts to extract data from these documents.

The first step that is done in Prompt Studio is to basically extract the raw text from your original document using a text extraction tool. One of the text extraction tools that you can connect with on the platform is LLMWhisperer, which is again Unstract’s in-house text extractor, known to produce really good LLM-ready documents.

After extracting the text, you can then define prompts over your documents which would define what data you want to extract as well as the schema you want this to be extracted in. [00:08:00]
Once you’re happy with the data that is being extracted and the format in which it is being extracted in, you can then export your prompts to your project as a tool and then deploy it in any of these four deployment options that we have:

API Deployment: where you process incoming documents from an application and send the output data back to another application.
ETL Pipeline: when you get your data from a file system like Google Drive, process it, and send the output data to a database or a data warehouse.
Task Pipeline: similar to ETL but sends documents back to another file system.

In any of these layers, we can also include human in the loop as an extra layer. [00:09:00]
What happens when you include human in the loop in the ETL pipeline is basically you get the document, you get the data extracted, and then there is a layer of human oversight to see whether the extracted data is accurate before passing it on to the database or whatever the relevant destination is.

Setting Up the Untract Demo

So that is the core demo segment that we’ll be looking at in this webinar—how to set it up and how to operate this entire workflow. Just to give you an idea, we have over 5.5k stars on GitHub, a 950 plus member Slack community, and today we are processing over 8 million pages per month by paid users alone.

So that said, let’s move on into the Unstract demo. [00:09:30]
I’ll give you a quick overview of the platform and how it works so that you’d have better context when we move into the HITL demo after that.

Alright, so what you see over here is the Unstract interface, and if I were logging in or signing up for the first time, I’d have to set up certain prerequisite connectors. You’d want to connect with some of the popular LLMs out there, and you have the option to do that over here.

So once you set up these connectors, you can then move on to creating your Prompt Studio project, where you’d be uploading your documents and also defining the prompts that you want to run against them. For instance, in this project, we have uploaded a loan application, and we also have a bunch of prompts to extract relevant data from this particular document.

Document Layout Preservation

[00:11:00]
The system preserves the layout of the original document during the text extraction, which is crucial for LLMs to get the complete context. This text extraction is one of the powerful capabilities of LLMWhisperer, which can also provide a confidence score on each text extracted, as well as the corresponding bounding box coordinates.

Confidence Scores and Highlighting

The confidence scores indicate how confident the system is about the text extracted, and it can determine which documents should go for human review based on this score. Moreover, it allows for source document highlighting, which is vital during human review. You can interactively view where the data was obtained from and assess its accuracy.

Review Interface and Permissions

Next, we delve into how to set access permissions, because dealing with sensitive documents requires careful monitoring. Under platform settings, you can define different user roles such as Unstract Admin, Unstract User, Unstract Reviewer, and Unstract Supervisor, each having different access levels to perform human review processes. [00:25:00]
The review interface allows for data to be highlighted based on confidence scores, making it easier to identify which data needs verifying. Once the review is done, you can send the document for approval, and once approved, the altered data gets sent to the database.

Conclusion and Best Practices

To summarize, we explored HITL's deployment through an ETL pipeline and API, and discussed best practices:

Set clear review triggers based on case importance.
Use confidence scores to filter reviews.
Maintain access controls for sensitive data.
Implement a two-step review process if needed.
Automate review triggers using cron schedules or API calls.
Continuously refine review logic to remain effective.

If you want to explore further, you can schedule a free call for a personalized demo. Thank you for joining this session!