Founded 2014 | HQ Boston, MA | 70 employees | >$5M annual revenue

Indico is among the new breed of disruptive AI software providers who have figured out how to make deep learning workable for intelligent process automation (IPA). The company has an impressive strategy to democratize neural networks and bring the benefits into mainstream document workflows.


The Company

Indico offers a suite of cognitive capture software products powered by deep learning and natural language processing (NLP). Founded by a group of engineers from Olin College of Engineering in Massachusetts, the company’s tagline is IPA for Document Intake and Understanding (IPA stands for intelligent process automation).

The Indico team trained its machine learning algorithms on more than 500 million publicly available documents and images: a quintessential deep learning project. The software acquired enough knowledge that Indico customers can feed the algorithms as few as 200 of their own documents and create enterprise-class document classification and extraction models ready for deployment. This approach is known as “transfer” learning; it enables users to train their own machine learning models with the benefits of deep learning neural networks but requires far less data and computing power than traditional deep learning and machine learning approaches.

Indico has targeted the robotic process automation (RPA) market, where current customers are looking to extend and expand their RPA investments into business processes that include unstructured content such as contracts, accounting documents, records, correspondence, and handwritten documents. The company told us that invoices in particular are still a huge pain point, and every prospect asks for help, but its software can process far more than invoices. Although most RPA vendors have their own capture offerings, they tend to be basic OCR and rules-based solutions effective only on structured documents with minimal variability, such as forms or W-2s. Automation Anywhere (AA), for example, offers IQBot for document capture. However, when it comes to semi-structured and unstructured documents, cognitive capture products with advanced AI are required to achieve the straight-through processing that makes RPA most effective. Indico created partnerships with RPA leaders AA, Blue Prism, UiPath, and Microsoft Power Automate, among others. The company presented at AA’s recent online conference and showed a compelling value proposition to extend IQBot for unstructured content. It plans to expand soon into other document-intensive workflow markets, with a Salesforce integration on the horizon, for example.

Indico sells mainly in the US and targets customers in the Fortune 1000. The company is in a strong financial position to grow rapidly. In December 2020, a Series B round of $22 million brought its total capital raised to $36 million.

The Technology

Indico IPA software provides a point-and-click environment to create and manage unstructured content classification, extraction, and workflows. The user interface focuses on helping the user teach the computer how to perform human tasks by creating AI models.

Indico IPA consists of several modules. Starting with Indico Teach, users can quickly build custom machine learning models tailored precisely to their document challenges (see Figure 1). The first step is for subject matter experts (SMEs) to label approximately 200 samples for each document type. The software learns from examples, and learning improvements are most noticeable where the document variability is the highest.

Indico demonstrated a loan processing use case for unbundling and separating large loan files. We know from experience this is one of the more difficult tasks to automate. Using NLP along with both visual and text algorithms, the software quickly created workable classifiers from a few dozen samples, all without any need for specialized data science or AI development training. This is a major improvement when compared to deep learning systems such as Microsoft’s that require tens of thousands of sample files.

After being fed a dozen documents or so, the prediction function – based on real-time learning – activates and the machine begins to assist the SME in completing the task. To wring out unintentional bias, Indico Teach can assign labeling tasks to multiple SMEs and track the activities.

Next, the Workflow Orchestration module sequences the AI models into a workflow to perform document classification, data extraction, and data validation. As documents are being processed, the Indico Review module provides an interface for a human in the loop (HITL) to review and handle any exceptions. Corrections are fed back into the learning system. The Workflow Analytics module is a dashboard where a project leader can monitor statistics such as document processing volumes and total time on task to produce ROI reports.

Another barrier to enterprise adoption of deep learning has been the lack of transparency. Deep learning is sometimes called the “black box” of AI, a closed and mysterious system of neural networks that issue decisions without accountability or auditability. This cloak of secrecy and unknowability quite frankly scares a lot of people.

Indico has addressed this issue head-on with Indico Explain, the module used to explore AI models. The software displays several metrics about how the data label decisions were made for each model. The SMEs and the Data Science team can do a deep dive into the modeling, enabling them to fully understand the decisions and performance, and can use insights to update the production models.

With Explain, Indico has effectively demystified AI and has the potential to become quite popular with the data governance and risk management teams who are hesitant to endorse neural-network-based processes. The same people will appreciate another benefit of the Indico approach: there is no data leakage; that is, none of the customer’s own data ever flows back into Indico’s master data set.

Another use case for Indico is for complex financial documents. A global financial advisory service firm specializing in the debt and derivatives markets needed help to automate workflows such as trade order confirmations, cap rate confirmations, and contract reconciliation. Using Indico IPA, the firm created custom deep learning models to classify and extract data from scanned, semi-structured PDFs.

Lastly, Indico utilizes a modern container-based architecture, which enables the solution to be deployed in the Indico cloud, within a customer’s private cloud, or on-premises, and provides built-in scalability able to handle huge document volumes.

Figure 1
Indico Transfer Learning

Our Opinion

Indico is among the new breed of disruptive AI software providers who have figured out how to make deep learning workable for intelligent process automation. The company has an impressive strategy to democratize neural networks and bring the benefits into mainstream document workflows. It has made a good start at dismantling the barriers to deep learning adoption by bringing to business the upsides of neural networks (a vast knowledge base and multiple decision layers) without any of the downsides (huge data sets, high computing costs, and opaque decision-making).

Indico has also created what we’re tempted to call “Deep Learning for Dummies.” The software offers step-by-step on-screen instructions that walk even a novice user through the process of building an enterprise-grade AI model to, for example, sort the scanned mail, or something equally prosaic and useful. And by solving the AI explainability problem with its transparency model, Indico demonstrated that it understands the enterprise view of AI technology.

Advice to Buyers

Any organization with an investment in RPA should consider Indico as the cognitive capture component when expanding RPA into document-intensive workflows. With its recent capital raise, the company is financially sound and should be a long-term player.

We also think Indico would be a good acquisition candidate for an RPA vendor or a software company that specializes in a vertical market such as healthcare claims or mortgage origination.


SOAR Analysis

Strengths

  • Master training data set
  • Transfer learning technology

Aspirations

  • Drastically reduce the number of customer training samples
  • Be a thought leader in the practical application of deep learning and neural networks

Opportunities

  • Become the cognitive capture leader for RPA
  • Acquisition by or tuck-in for a large software company

Results

  • Blue chip Fortune 1000 customer base
  • Strong partnerships with RPA vendors