AI for the Enterprise in 2017: Start Small, Scale Decisively

Tales and Best Practices from the Land of Early Adoption

In the Consumer space, Artificial Intelligence and Machine Learning are making a splash. But what about the Enterprise? Only a small percent of business and technology professionals are using AI, mostly because of 3 reasons: the actual use cases are not well understood, they lack the necessary skills, and the technical infrastructure is not yet ready. Where AI is adopted in the enterprise, the results can be stunning, for instance with Cisco's StealthWatch and Tetration. Multiplying this kind of high-impact verticals allows staying ahead of the curve and move into the position to tackle the next market transitions.

The 2017 State of AI in the Enterprise

Siri, Google Now, Echo, Tesla Autopilot, Facebook Feed, Skype Translator,… The number of consumer products that make use of AI has grown dramatically over the last few years. We know because companies tell us, as AI is currently a hot marketing concept. Beyond the promoted ones, consumer companies routinely test new AI algorithms on end users, without publicizing them. This agile approach is indeed important in developing AI products, because it allows companies to maintain the focus near the frontline, emphasizing the benefits of the product and not the technology. In fact, the vast majority of successful AI products in the market today started with human engineered structures and parameters. Truly AI-native products — where AI is a core element from the start — are still very few, and many startups are trying to be “AI-first” companies. What this reveals, of course, is that AI is currently still used as a method to improve existing components and is not yet an enabler of new groundbreaking applications (cars were driving by themselves with comparable results to today but using principles of robotics since the 1990s). Outside the consumer market — notably in health care research — AI is enabling automated methods that perform better than humans. For instance, detecting cancers better than oncologists.

What about the Enterprise space? A study found that only 12% of business and technology professionals are using AI systems (Forrester, Nov 2016). The study further identified 4 reasons why the vast majority of enterprises is not yet into it:

  • There is no defined business case
  • Not clear what AI can be used for
  • Don’t have the required skills
  • Need first to invest in modernizing their data management platform

Given the ability of AI systems to scale with large amounts of data beyond human engineering, and that most enterprises (58%) are researching around AI, the B2B market will change drastically in the next few years. The question that then naturally arises is, how should the enterprise tackle this upcoming market transition?

Successful Verticals

As it happens when new lands are discovered, low-hanging fruits abound. The difficulty is to recognize these opportunities, as they might look very different from what we’re used to.

In general, what are the methods to facilitate the identification of low-hanging fruits in innovative areas? For large corporations, an effective approach for assessing new technologies consists in the following 3-step process:

  1. Understand what the technology actually can do and how
  2. Translate the understanding into the context of each of the company’s business units
  3. Prioritize the opportunities for size and time-to-market

If there are more than 3 business units, it’s most efficient to dedicate an AI Task Force to this activity, to avoid replicating the required technical fluency, maximize the synergies (especially in the critical area of having sufficient relevant data), and create a company-wide strategy that can also be used for marketing purposes.

Note that this approach is valid for in-house development and acquisition strategies. The advantage of building AI teams is that it will lead faster to transforming the troves of internal data into value. The advantage of acquisitions is faster time to market with products and services, while their biggest challenge is, as always, the integration of the new teams into the existing structure. Combining in-house developments and acquisitions through an over-arching strategy makes the most of their complementarity.

As an example of the 3-step process, let’s apply it to AI in the context of IT systems. To be accessible to a sufficiently large audience, this example needs to be rather generic but hopefully sufficient to understand the gist of the method. Each of the steps can be formulated as questions that drive a discussion:

  1. What is AI especially good at?
    Learning millions of parameters to solve a specific task
  2. In an IT business unit, where can large number of parameters be found?
    Back-end management (explicit type of parameters) and meeting discussions (implicit parameters).
    The first is about reducing inefficiencies through automation and optimization, while the latter is about new products to increase productivity, such as automatic generation of meeting minutes.
  3. Which opportunity is larger? Which one is faster to market?
    The size of the opportunities might be comparable, however it is faster to have significant positive results in automating back-end management than in a new product because there is more control and the metrics are better defined right off the bat. Acquiring an NLU+G (Natural Language Understanding+Generation) product already on the market might shift the balance, although an overarching strategy is essential for alignment and solving the challenge of integrating people, systems, and data (which is where the money is).

Although applying AI behind the scenes might not be as sexy as new products, it allows companies to test and learn the fundamentals of working with AI without consequences, and hence build confidence. It is the job of the cross-BU AI Task Force to collect the learnings and bit by bit create best practices that will become part of the AI development DNA of the company.

Once the target vertical application has been identified, the potential of AI needs to be validated, brought into production, and actually used. To achieve this, it is recommended to use an agile approach that takes in consideration the differences of AI from human engineering, the fact that most successful AI solutions rely on large quantities of data, and that there is considerable IP to be created:

  1. Set up small internal teams (2–3) of developers, with or without prior AI knowledge (it is best to internalize efforts because the generated know-how and learnings are the basis for setting up the IP strategy)
  2. Create early PoC solutions using open source frameworks and open datasets. Don’t be afraid to call them “toys”.
  3. Don’t buy from vendors before understanding what’s going on, what you actually need, and before having the right data and questions. Then demand from the suppliers to provide crystal-clear answers and data
  4. Apply your data for progressive performance analysis, and start thinking about IP strategy

This approach allows for effective innovation: Transforming AI ideas into products and services relevant to the business with relatively low people and infrastructure cost. Many of these initiatives often appear spontaneously, but only when the results are shared, the power of using AI as a fundamental approach spreads like a wildfire.

Multiply and Thrive

Each successful AI vertical has the same characteristics:

  • Millions of samples have been prepared for training
  • Use cases tied to measurable objectives have been defined following heuristics (i.e. rule of thumb)
  • Millions of parameters have been iteratively optimized (a small but crucial part of which by human architects)
  • Data is processed very fast

Even without the core element of AI — namely that it is the computer, not the human, that formulates and tests the relationship between cause and effect — these characteristics are a hallmark of how any project should be run. It follows a process that is structured following heuristics, is data-driven, and has quick feedback loops. Additionally, the technical infrastructure is composed of efficient data pipelines, powerful computing capabilities, and visualizations designed for the use case. If all projects of any given company would have these elements, their executive leadership teams would be challenged to manage growth.

The hidden benefit of adopting AI is precisely this: It is an agent of change towards a high-performance mindset across the board. Exactly what most enterprises need to tackle accelerating rates of change.

Who says “change”, says “resistance”, but profound innovations are inevitable. Despite the technical depth of AI, currently it’s just another tool, it isn’t yet what drives a sentient machine. As a tool, anyone can use it. In the 1950s, mechanical calculators were all the rage, and human operators would spend their days pressing buttons, copying results to paper, and send them in envelopes to their colleagues for further operations. Then when computers arrived, hundreds of calculators were replaced by a single spreadsheet. Days became seconds. And mechanical calculator companies went out of business. Would anyone today have trouble learning how to use a spreadsheet? Not so long ago it was inconceivable for non-specialized staff, let alone that computers would be so common. With AI, we’re at the same early stage of the innovation curve.

Concept expanded from Andreessen Horowitz

As innovation curves go, the adoption happens in phases. And different phases have different needs. At first, it’s early adopters — the innovators — finding success in individual projects. Then, convinced by the results, the early majority tries to adopt the novelty. If successful and after inertia is overcome, the innovation becomes a standard with the late majority. Finally, everybody is on board because that’s just the way things are done.

The speed at which an innovation moves through the phases depends on several market and organisation factors, but the longest part is overcoming the inertia between early and late majority. Think about your sister-in-law who never wanted a cell phone, and now finally replaced her Nokia 3310 with a smartphone. Obviously now you can’t get her off the 312 WhatsApp groups she’s in (yes, in 2017, that’s a thing). The choices of your sister-in-law might be difficult to understand, but the reasons of most enterprise teams are clear: They are either paid to (1) achieve specific metrics or to (2) attain some fuzzy objectives. In the first case, they’re too busy to care. Think about teams so busy pushing a company cart with square wheels that they don’t see how they could find time and resources to try round wheels. In the second case, the freedom given by the lack of clear objectives can be convenient as well as risky. Usually the net result is a tendency to maintain the status quo. Both cases can cause longer hours, worse environments, and worse output quality than what technology actually would allow.

It is beyond doubt that working better is healthier for people and for companies. The key to moving closer to that sweet spot is to provide the right incentive structure, that is, paying people for working better. Properly innovative enterprise technologies allow for working better, it is thus the role of the strategic leadership is to provide the conditions for adoption and allow for diffusion. If the early majority cannot reproduce the success of the innovators or there is no reward for the calculated risk of trying, the new technology gets stuck and a bad reputation. If the late majority has no incentive to change, the innovation stays in isolated projects and is killed at the next budget optimization. These are small and very human actions, that compounded lead to missing the train and slowly sink into oblivion.

Now, what can be done in the context of AI? Adopting a few AI projects can be simple enough, as explained above. But how to provide the conditions for systematizing the AI approaches throughout the company? For starters, the vision must be bold and crystal clear:

We work better by focusing on framing the challenges and enabling computers to solve them. This leaves us more time to personally take care of our relations with the customers.

Such an AI-first vision serves the early adopters to contextualize and remind them of greater rewards down the line. The vision serves as a beacon for late adopters on their journey of transformation. On this path, the following elements are developed:

  • Knowledgeable and skilled people
  • Common and easily accessible best practices, standards, and learning structures
  • Efficient data pipelines, powerful computing capabilities, and quickly customizable visualization systems

When AI grows in the company, the above elements appear in separate functional entities, such as HR, IT, and Sales. At a certain point, when a number of successful projects exist, the executive leaders must take a decision and follow-through by creating the appropriate structures and incentives. The number of successful projects needed is directly correlated with the boldness of the executives. Late decision-makers will need to face talent bleeding and solve inefficiencies. Early decision-makers will be rewarded by being recognized as the driving force in capturing market opportunities and transiting the company into the future.

Because the number 1 requirement of AI is data, the appropriate structure for scaling is a single functional entity that manages the proprietary data, and all other aspects around it. This will avoid inefficiencies (e.g., heterogeneous labeling in data sets) and bring scale effects. This “Corporate AI” has also to be responsible for appropriate sales and hiring processes, training programs, and computing structures. At first, it is best to create a separate entity to avoid integration and legacy issues, as well as take advantage of marketing opportunities. Eventually, Corporate AI becomes part of Corporate IT. Mechanical calculators were replaced by computers — bicycles for the mind will in turn be replaced by self-driving cars for the mind.

~ Nicola Rohrseitz