Earlier this fall I published an article predicting that a new class of AI-powered applications will reshape the workplace by automating away the individual tasks that make up our jobs. Most of the companies featured in the landscape are early stage startups.

I recently had an interesting conversation with someone who argued against the viability of artificial intelligence startups, given the tremendous budgets and development resources large tech companies like IBM are committing to develop broader platforms — the premise being that these platforms will fill each niche application startups try to carve out (think scheduling meetings or answering common customer services requests).

I’d like to argue against this notion.

AI-Powered Apps Built By Startups Are Designed To Win
AI-powered applications are by design built to close the data gap impeding commercialization.

AI systems need highly relevant data sets with a rich set of semantics on which to train. Startups address this by functioning as add-ons to existing cloud business apps like Gmail, Google Analytics, Salesforce and LinkedIn. For example, X.ai uses email context from Gmail to develop a personalized calendaring assistant, and DigitalGenius uses Salesforce Service Cloud’s API to automate common customer service requests.

Generally speaking, public APIs for cloud business applications provide creators of AI-powered apps:

• Rich semantics to create more intelligent AI systems. With highly structured data sets and lots of metadata, business app APIs make it easy for data scientists to construct the data hierarchies necessary to put the “deep” into deep learning, as The Economistonce described the neural network (an AI construct intended to emulate abductive reasoning). In a neural network, deep-learning algorithms travel along hierarchical data sets much like brain signals transmit along neurons. The more layers of data within a hierarchy, as afforded by metadata around a workflow within a cloud business app, the higher levels of abstraction in reasoning the AI system can reach.
• Narrow data sets to shorten training periods and get to market faster. AI systems thrive off brute statistical analysis on big data. But much like teenagers, they need a little help being pointed in the right direction. The cloud business app is a narrow enough domain for the fledgling AI to flourish. Domain-specific data acts like the high-school basketball coach, helping the untrained AI identify relationships between inputs and desired outcomes, and shortening the training period necessary to run with success. The team at Rainforest QA notes having to put in twice the effort to build impressive AI features because they didn’t have the luxury of a narrow and highly structured data set on which to train machine-learning algorithms.

In contrast, broad AI platforms, used to build proprietary AI systems, aren’t designed to leverage the pre-built semantics and classification of open APIs and libraries to the same extent. This impedes time to market.

IBM Watson, which solves company-specific use cases and trains on proprietary enterprise data sets, requires daunting manual data management and rigorous training that IBM’s global business services division is happy to charge large services fees to undertake.

Individual startups building AI-powered apps will be the dominant force disrupting the workforce over the next five years.

IBM is working to make Watson more scalable, but homegrown AI projects today look alarmingly similar to ERP implementations in the late 1990s — 14 months, 150 consultants and an unfavorable success rate. Some of these projects will yield promising AI systems that reset competitive advantage across industries, but they will take many years to reach critical mass under this paradigm.

What will move the needle for broad AI platforms? Major coalitions of inter-company data exchange and knowledge repositories for common industry practices are a start. Does that sound realistic anytime soon?

Full Steam Ahead: Consolidation In AI Won’t Happen Anytime Soon
As a means of stress-testing my assumptions, one could use the ERP analogy mentioned above to argue that AI-powered apps will consolidate to a standard application suite before we know it. Business processing re-engineering, which I previously compared to AI developments, fizzled out soon after its height in the mid-1990s as ERP software suites, which standardized BPR efforts in the form of out-of-the box software, became table stakes.

But things are different today. For one, AI-powered apps are fickle. No single company, no matter how dynamic, could easily build these things in parallel, because each app is a finely tuned system requiring domain experts and a dedicated team of AI trainers to maintain.

AI-powered apps will eventually consolidate, but by way of acquisition rather than internal development by the enterprise giants.

What this all means is that, although broader platforms like IBM Watson will be critical to solving company-specific problems such as diagnosing patients in hospitals, individual startups building AI-powered apps will be the dominant force disrupting the workforce over the next five years because of their reach and ease of deployment.