Democratizing Data and AI for the Enterprise

“How do we create a democratization strategy for data and artificial intelligence,” a C-suite client recently asked me during one of my product research calls.

It’s clear that data and AI democratization is top of mind among knowledge worker industries and sectors. As organizations explore ways to successfully democratize data and AI, here are five tips that are particularly relevant to enterprise and sector-specific computing:

1. Create an internal data and algorithm marketplace.

It all begins with clean data, a solid data structure and well-labeled data. Organizations might ask themselves, “Who is going to be the data janitor?”

To make this happen, I often recommend that enterprises consider creating a private marketplace to internally crowdsource the creation of well-labeled data products, data structures and pre-trained algorithms.

For example, the recent AI boom can be traced to a seminal moment: the creation of a manually annotated and labeled database -- ImageNet -- created by a research team at Stanford. The ImageNet dataset was built and annotated using crowdsourced data marketplaces.

Large enterprises have yet to create their own “ImageNet moment.” For example, there is no famous dataset in financial services that multiple enterprises are using to bootstrap their internal AI algorithms. However, there are now many off-the-shelf products that can help enterprises bootstrap the creation of well governed, internal, and private data cataloging, data labeling and data annotation “marketplaces.”

Crowdsourcing the non-glamorous, data janitorial work is the first step to more transformational use cases such as industry-specific intelligent recommender systems.

2. Adopt a designer’s mindset.

I can hear you asking: How on earth does design or app experience matter in the world of AI engineering and data science?

Quoting one of the great technology designers, Dieter Rams, “Good design is as little design as possible.”

We cannot expect wide adoption of enterprise data/AI products without designing to the needs of the end user. For example, if it’s a home consumer, they might want an AI appliance that talks or has voice, but if it’s a contact center, they might be more comfortable with a text-based chat-app.

For enterprise developers, the experience might even take the form of a well-designed AI application programming interface (API) that embeds contextual intelligence within a larger enterprise application.

In fact, an API-based approach to AI might be more relevant than anything else within a Fortune 100 organization. Embedding industry or domain-specific intelligence into enterprise applications by using AI APIs is better than building an enterprise application from scratch. APIs aid democratization among developers, as many algorithms and models actually languish on isolated laptops scattered across Fortune 100 giants.

One of the best examples of this approach is the Open AI API built on top of the GPT-3 model -- a language model that uses deep learning to produce human-like text. The explosion of creative use-cases built on top of this easy-to-use API is remarkable. My favorite was an AI-enabled general-purpose Excel function that looks up and automatically fills in state populations.

Imagine intelligent APIs that augment apps used by financial advisors and knowledge workers within every industry and sector.

3. Embrace the toolkit “culture” of your community.

An organization’s data scientists and data engineers express their technical culture through the popular tools and data assets that they use to communicate across silos. Not too long ago, many CIOs were still in denial of the open-source revolution. With open ecosystems such as code repositories and code messaging boards proliferating, it would appear that the democratization of data and AI must involve listening to toolkit standards with an enterprise’s data scientist and developer community.

Any unilateral attempt to police data science toolkit standards will likely fail. For example, which Python notebook is most popular within your organization? Did someone create a rogue wiki to catalog all the data elements? Is there a common data model that analysts share and reuse to help wrangle common flat-files within a typical sector? Are there pre-trained R-algorithms that your analysts are using within your domain?

Why not listen to the community when shaping standards? Democratization is ultimately about shining a light on the data science innovation happening within your organization’s trenches.

4. Continuously audit your algorithms.

Who audits an algorithm that decides whether a segment gets a mortgage? What about an algorithm that decides whether you should get a job interview or keep your job based on past performance?

There have been news stories recently about large companies that have experimented with using algorithms to prioritize job interview candidates or even evaluate job performances of factory workers. The “black box” nature of certain types of algorithms makes the adoption process problematic when defending against alleged or accidental bias inside algorithms.

This is an example of how important it is today to “verify trust” in data and algorithms. This is not a point-in-time exercise. I advise teams to adopt a culture of continuously testing and auditing data and algorithmic products within an enterprise. There is emerging research using new algorithm audit methodologies such as contrastive explanations to bring transparency to “black box” models.

5. Champion a community leader program.

Democratization is ultimately about people and bottom-up leadership. Find star performers who can serve as community leaders, and then make them in charge of a new center of excellence (CoE). If you don’t know who your stars already are or are in recruiting mode -- look for contributors to public code repositories and technical blogs.

You might have to incentivize or gamify contributions to data science-specific knowledge management systems internally -- especially if they have poor historical adoption in your organization.

Such “leaderboard” scores will not only help you identify future champions for a data-driven transformation but will also help monitor the data and AI “democratization” score for your organization. Monitoring this score regularly and creating the right incentives could even help create a virtuous cycle of internal data and algorithmic products.

While these steps only serve as a guide, it’s worth reinforcing how critical it is to democratize data and AI enterprise wide. Democratizing these departments empowers employees to find their own solutions, saving them time, and allowing AI and data teams to focus on strategic initiatives, rather than ad-hoc support.

The views reflected in this article are the views of the author and do not necessarily reflect the views of the global EY organization or its member firms.