Analytics

Why analytics tools are prime vehicles for “democratizing” AI/ML

By Peter Bailis - December 21, 2020

In the rush to “democratize AI and ML,” the $100B question remains: what does this actually look like from a product perspective? We’ve seen rapid advances in visual and text processing productized via APIs and verticalized in offerings for call centers, manufacturing, and retail. However, for the everyday enterprise user, AI/ML largely remains the realm of flashy demos and specialized data scientists.

At Sisu, we believe there’s a better way to deliver machine-assisted capabilities to end-users and aren’t shy about this open secret: analytics tools are the most effective vehicle to deliver AI and ML capabilities in the enterprise. Although these platforms were originally designed for descriptive analysis — dashboarding and reporting — there are three major reasons why analytics use cases are a particularly promising vehicle for bringing advanced AI/ML to more users:

    1. Broad organizational footprint. While the adoption of AI/ML tools remains scarce even in the largest enterprises, analytics tools – primarily in the form of dashboards and reports – are ubiquitous. I regularly encounter companies that have thousands of employees using analytics tools on a weekly basis. If we can develop interfaces to embed AI/ML capabilities into these analytics tools, there is a large, pre-existing, multi-departmental user base to empower. Moreover, as the de facto translators between organizational data and business users, analysts and users of analytics tools are primed to take model outputs and turn them into actionable recommendations for the business.
    2. Low-friction access to first-party data. Advanced analytics depends on large amounts of data to provide statistically sound predictions and recommendations. Especially as cloud data warehouses grow in volume and complexity, these databases are a target-rich source of proprietary, well-labeled training data about a given enterprises’ operations, products, and customers. Analytics tools have been designed from the ground up to connect to these warehouses – providing a low-friction, well-trodden path to data, and ultimately value. Moreover, by capturing how users interact with their existing analytics tools, analytics provide easy access to weakly-supervised training data for personalized ranking and relevance.
    3. Human-in-the-loop interaction. The early adoption of AutoML tools makes it clear that building high-quality, robust models requires substantial human involvement in the form of data labeling, quality assurance, and end-user feedback. Many successful AI/ML-enabled products still depend on a “human in the loop.” This interaction model can be challenging to productize (“what do you mean I have to rate the quality of my chat interaction?”). However, analytics platforms have always been designed for interactive, human-in-the-loop interactions, in the form of interactive data exploration, slicing and dicing, and drill-downs. Rather than reinventing the wheel for human-in-the-loop interactions, building AI/ML into analytics tools provides a convenient and familiar interface for users.

Given this broad reach, access to large amounts of data, and familiar interaction models, analytics tools are a promising means of truly “democratizing” AI/ML. The key challenge in realizing these benefits comes in the co-design of AI/ML capabilities and product interfaces. This is a departure from conventional product development of analytics tools, which tend to be visualization- or data-centric.

We’ve seen the first steps towards this kind of algorithm-interface co-design in early features like Tableau’s Explain Data and Power BI’s “Automatic Insights,” which layer AI/ML-powered results on top of existing visualization engines. However, based on our experiences in both research and with customers, our thesis at Sisu is that fully realizing this potential requires a complete architectural rewrite, with scalable and powerful AI/ML at the core of the product experience and visualization layered above. We believe the market is much like search — the results (provided by the search engine) will matter most, not the visualization (provided by the browser).


Read more

Humans, not machines, are the main bottleneck in modern analytics

With advances in data storage and compute, the fundamental bottleneck in analytics has shifted from the infrastructure to people. In this post, Peter breaks down this shift and explains how Augmented Analytics will re-balance this human bottleneck.

Read more

Three Inconvenient Truths about the State of Enterprise ML

Last week at TieCON 2019, Sisu CEO Peter Bailis challenged conventional wisdom around enterprise ML with three principles for making ML accessible to all.

Read more