AI Adoption: Challenges and the Future
Artificial Intelligence (AI) is one of the most hyped topics today. AI has beaten humans in Star Craft 2, Dota 2, Chess, Go, it personalizes searches on Google, drives self-driving cars and recommends movies on Netflix. Still the industry is fairly small in terms of revenues: in 2017 AI software market was estimated at less then USD 1 billion, which is peanuts comparing to a USD 30 billion CRM market.
There are several reasons for this:
- The absolute number of deployments is still very small: less than 100,000 enterprises are expected to adopt AI in 2019, for comparison, only in US there are close to 30 million small-medium businesses
- Half of all the deployments are pilot projects and PoC’s which are far from commercial deployments in terms of scale and functionality[1].
AI Adoption Challenges
To better understand what are the challenges for AI adoption we need to separate two main types of AI:
- Embedded AI. AI is a component of a business application, one of the examples is Salesforce with Einstein capabilities embedded into their CRM.
- Standalone AI. AI capabilities are a separate implementation independent from other business applications. Most of the most famous AI implementations are standalone implementations (Google, Netflix, Amazon etc.). These companies develop AI capabilities in-house themselves.
Standalone AI
The implementation of Standalone AI is a challenging task due to high cost and high-risks of failing to deliver expected business results.
Despite all advancements in data science it is still considered to be an art. This results in extremely high skill acquisition costs. A single team to deliver AI project can easily run into 7 figure payroll expenses:
- Data Scientists: average base pay $139,840/yr (Source: Glassdor, US)
- Data Engineers: average base pay $151,307/yr (Source: Glassdor, US)
- Software Developers: average base pay $91,887/yr (Source: Glassdor, US)
Like many complex IT projects, it is a high-risk venture: even after spending millions of dollars there is no guarantee that you will get the expected benefits:
- There might not be an answer to the business problem with the data available
- The prediction accuracy might not be high enough to justify the business case (i.e. the business case breaks even at accuracy of churn prediction being 99.5% while with the available data the maximum accuracy is 95.7%)
These factors act as barriers to entry for small and medium business which efficiently limits the implementation of Standalone AI to large enterprises.
Embedded AI
With the Embedded AI, the costs and risks of the AI implementation are a burden of software vendors rather than the enterprises using the software.
Still there are several challenges for the implementation of Embedded AI:
- The data used for AI purposes will most likely be limited to what resides in the business application, as data integration is a separate complex task which will require extensive effort
- Even when using the data residing within specific business application, there is no guarantee that what worked for one customer will work for another, since everyone has different data, especially in different markets
This leads to the fact that software vendors need to invest significant resources into developing functionality which might not “work” for some customers who will not pay for it (obviously).
Future of AI adoption
There is enough evidence that AI can generate significant value for the business, so it is rather “when” than “if” AI will be widely adopted by businesses worldwide:
- Automated Machine Learning (Auto ML) tools are expected to automated some of the data scientist work to reduce the cos of skills acquisition improving adoption levels of Standalone AI. Examples: Google Auto ML, TPot.
- Over time software vendors will be able to evolve Embedded AI as they get more and more customers with real data using the AI capabilities.
- Number of AI start-ups has been growing expnentially since 2008 bringin new innovation into the industry
Whether the hockey-stick type of industry revneue forecasts we see today will live up to their expectations is still a question, but the growth of the AI is adoption inevitable.
