AI in the post-watson era
AI was never the endgame. We’re taking stock of the industry and proposing a way forward for enterprise in the wake of IBM Watson’s retreat from drug discovery.
IBM recently announced that it is taking its flagship AI, Watson, off the market for drug discovery applications. There’s been confusion and misinformation in AI since its inception in the 1950s, but following this public admission of retreat, the conversation around AI and deep learning has become even murkier than before. IBM was so successful in marketing Watson as a thinking machine, it set expectations sky high – while setting the stage for their exit today. Believers and non-believers alike must now come to terms with the fact that AI is not what they thought.
Trust plays a less obvious but more impactful role. Sourcing tech innovations from credible giants like IBM was always thought to be a safe and reliable way for companies to get ahead. This has been the “social contract” for decades between tech companies and the enterprises they serve. Now, when the most fundamental assumptions are under question, what should forward-thinking execs do?
In the aftermath, there are two different attitudes when it comes to AI. Some are comfortable with uncertainty and the nebulous, non-deterministic nature of AI. This group is ready to retrain themselves for whatever comes next, as they have always done. The other group has come to a conclusion. AI is never going to fulfill its promise.
Tech evangelists and the media will have you believe that AI is the endgame, but those fighting the uphill battle of business transformation know better. Digitalization is in fact the end goal of enterprise AI. AI is a very powerful and promising tool, but a tool nonetheless. It’s easy to get swept up in the hype around all cutting-edge tech, but if we keep in mind that technology is a tool, we have a way forward despite the uncertainty following Watson’s withdrawal.
The guiding principle for the savvy exec making business decisions around digitalization toolbox is to remember to focus on ROI and total cost of ownership.
Fitting square pegs into round holes
Deep Learning, the AI paradigm powering Watson, is an amazing technology. In an attempt to mimic human thought processes, artificial neural networks analyze hundreds, even thousands of “hidden layers” of information based often on just a few inputs, creating surprising and delightful outputs.
Deep learning promises to identify objects in photos, diagnose heart failure, and compose simple jazz music. But it’s not the only AI paradigm out there. And it’s not the best AI paradigm for all enterprise applications because it’s built for one type of data only.
A Problem Bigger Than Big Data
In the complex world of enterprise, a key but not-often discussed problem with deep learning is its inability to deal with unstructured data. That is, information that’s not organized or easily categorized – think documents in natural language, like emails, memos, and scientific papers, as well as images, videos… Even semi-structured data, which appears structured but can’t be used directly like tables, graphs etc.
Most enterprise data is unstructured. In fact, most data is unstructured, period – it comprises up to 80% of data being created today. Valuable answers and insights locked away in these documents has been largely ignored.
The “big data” problem that no one wants to look in the face is that too much data is locked away in formats which can’t be made use of using popular techniques like simple keyword search or even deep learning.
In the past we’ve written general guidelines for enterprises looking to process their vast data stores depending on format and application. There are nuances, but generally speaking we recommend:
- Use big data analytics for deterministic cases. Many applications don’t need AI to find answers but involve huge data sets – these need only turn to big data analytics.
- Use deep learning if there’s a lot of structured data. Deep learning can unearth surprising insights as long as input data is pristine
- For applications involving natural language, turn to cognitive automation and natural language understanding.
Watson’s troubles began early, when it made the decision to force-fit deep learning, which is meant for structured data, onto applications with volumes of unstructured data.
Even following Watson’s exit, execs haven’t been left high and dry – a cutting-edge branch of AI called Natural Language Understanding focuses specifically on unstructured data.
What is Natural Language Understanding?
Natural Language Understanding is essentially teaching computers to “read between the lines”. Humans have evolved an intuition for abstraction during the many thousands of years we’ve had language, and the rules are fluid at best, completely incomprehensible at worst.
Deep learning has serious trouble drawing conclusions from language, even with huge data sets to guide it, meaning companies who implement deep learning must spend enormous amounts of time and money cleaning, structuring, and annotating all input data. Even the most advanced deep learning applications like Google BERT struggle with context which humans handle so naturally. The mainstream media is talking about BERT like it talked about Watson five years ago – uncertainty is at an all-time high, and the upfront investment is so big and the results so unpredictable given the cost, it’s no surprise some have given up on AI already.
There are other choices. Algorithms based on natural language search such as Calibrated Quantum Mesh are specifically built for the task.
CQM is Coseer’s unique solution to the big data problem, developed over seven years to handle natural language in an enterprise environment. A few key design elements allow CQM to handle the subtlety of language by emphasizing that all words have multiple meanings, and that everything is dependent on context.Because of CQM’s ability to handle language, it is truly built to add value where other AI paradigms cannot – by processing unstructured data. It also allows for very fast training and deployment; as little as 4-12 weeks, and reaches 95-98% accuracy.
Vaporware, Snake Oil and AI
Even before any doubt of Watson’s capabilities, vaporware and high-tech snake oil cluttered the AI marketplace. A recent survey by London venture capital firm MMC found that 40% of Europe’s so-called AI startups don’t use AI in any way meaningful to their business. It’s easy to see why companies would want to jump on the bandwagon, as “AI” startups command 15-50% more funding.
Many of these VC-backed startups have no real business model or proven use cases, perhaps because the reality of enterprise AI isn’t as exciting as the hype. Most use cases are chatbot deployments or fraud detection – hardly the futuristic stuff of sci-fi.
Execs navigating the waters post-Watson have to wade through all this vaporware to get anything grounded in reality. An unpleasant task to be sure; however there is a reliable metric which offers a liferaft: customer references. We’ll get into details below.
It’s imperative to work with people who have been successful in solving enterprise-grade issues. There are plenty of vendors out there who resist the temptation to over-hype themselves or over-promise their capabilities using a single AI technique. With a clearer understanding of the options at hand, it’s time to talk about ROI in context.
What’s different about AI ROI?
In order to walk the line between pie-in-the-sky thinking and complete pessimism in the post-Watson era, the guiding light for enterprise must be ROI. AI, being a tool for digitalization, is not much different than other projects once you’re tracking the right variables.
In our six years experience debating cost curves and ROI, we’ve learned that there are three laws to keep in mind when choosing the right AI vendor and during deployment.
- Avoid data preparation wherever possible
- In the whirlwind of KPIs to choose from, make sure you’re tracking accuracy
- Data scientists are important, but they’re not usually the bottleneck for ROI
If you’re shelling out enormous amounts of cash for data prep before any indication that your AI solution will work for your application, you could be in trouble. Data preparation usually means SME’s must go through scores of data to determine what’s relevant and tag/annotate accordingly. This can take months, and their time doesn’t come cheap. Combine that sunk cost with no guarantees about the you’ll get actionable insights once the data is input into the engine, and you’re likely already way behind on ROI.
AI will save your business time and money if 1) it can help solve the problem you need it to, and 2) the problem is diagnosed correctly by the AI. This is how we define accuracy. Accuracy is the key driver for lasting profitability as the system gets better with use. Take a look at how ROI improves as accuracy goes from 60% to 90%:
Accurate systems reward the user, driving more traffic to the AI system, which trains it better, in turn making the system more accurate. This virtuous cycle is important to ensure costs are covered, but it pays off big in the future.
Which brings us to data scientists. It’s common sense that talented data scientists will accelerate the development process, but we’ve seen that management processes are a bigger bottleneck for AI deployment than lack of data science talent. A brilliant data science team won’t be able to drive ROI unless the organizational ecosystem can support them. By bringing talented data scientists into your AI equation whether internally or through a vendor, you can certainly reach ROI quicker – but don’t lose sight of the bigger picture.
Key questions to ask when mapping the AI landscape
Choosing the right AI paradigm to suit your problem and properly tracking ROI will set you up for success in your AI deployments, but we understand that it’s rarely so cut and dry in the real world, especially when choosing an AI partner. Even as the sun sets on Watson’s highest ambitions, hype is high.
The best indicator of a solid partner is always customer reference. Total cost of ownership includes so much more than just the fee you’ll pay your vendor; it’s cost of hardware, SME consultation, data annotation, opportunity risk… the list goes on. But it’s absolutely critical to your digitalization strategy that you have a solid grasp on the costs and risks involved before you can move ahead confidently.
The five most important questions we recommend asking of any potential AI partners are as follows:
- Has the vendor solved similar problems at other enterprises? Are they willing to talk about their experiences?
- Is the technology that the vendor uses the right fit for the use case? If they use open source technology, how are they different from a services organization?
- What is the likely total cost? Be sure to include software cost, hardware costs, cost of data acquisition, cost of data preparation, direct personnel cost/ opportunity cost of personnel.
- What happens when AI makes a wrong prediction? What kind of controls are necessary to manage the wrong predictions? How much do they cost?
- How does this project lead to more projects based on similar AI?
As always, if you’re ready to explore next steps in your AI journey, we’re here to help. Setup a call with us to learn more about NLS, ROI for AI projects, and how to reach your digitalization goals.