While A.I. and machine learning has made rapid progress in consumer facing applications such as Amazon Alexa, playing Go, and making art, its deployment across the enterprise is still emerging and uneven.
Opinions differ on the scope and pace of A.I. to automate and enhance activities across the organization. On one end, a recent Oxford study suggested that >50% of all jobs could be automated by A.I. in the next two decades. On the other spectrum, many executives still remain doubtful to the practical impact of AI. In addition, the current hype cycle and misinformation in the media further cloud perspectives.
Based on hands-on client engagements and interviews across academia, startups, and enterprise users, we highlight 4 key facts and strategic implications on AI enterprise automation:
Fact 1: The current A.I. boom is sustainable and executives should not ignore it
For the first time in history, machine learning algorithms are beating human performance across tasks such as image recognition, voice to text translation and complex games such as Go. The current machine learning boom is fueled by a convergence of three underlying drivers – (1) continued breakthrough in Deep Learning algorithm sophistication, (2) rapid increase in big data (or structured data) to train such algorithms and (3) an exponential speedup in machine learning hardware such as the GPU chipset that cuts down training time from months to days and hours.
All three underlying drivers of today’s machine learning advances are expected to accelerate into the near-term future (see Figure below). By 2020, almost 70% of all enterprise data will be stored and processed via cloud data centers – providing unprecedented big data infrastructure as input for machine learning algorithm training. Similarly, hardware chips for speeding up machine learning algorithm training and processing is making rapid inroads. Google, NVidia, Intel, and others recently announced next-generation GPU chip hardware that will further accelerate training speed by 10-100X fold. Lastly, advances in the underlying machine learning algorithms continue to accelerate, as evidenced by an increase in patent filings and applications.
The above trends offer strong evidence that advances in A.I. capabilities and performance will continue to improve in the near-term future. Ignoring A.I. is no longer an option for business leaders.
Fact 2: A.I. use-cases are pervasive across the organization – but limited in scope
Given that AI advances will continue to accelerate, what can A.I. do in the next 5 – 7 year near horizon? The consensus expert view is that most of the enterprise use cases will be driven by narrow A.I. (e.g. supervised learning) and that achieving broad, human like general intelligence is decades away.
In short, AI algorithms will have the capability to learn to automate a task from training data. But once it learns a task, the solution will be narrow in scope and – in most cases – cannot be generalized to perform other tasks. See figure below for a sample of use-cases that are possible today or emerging in the next 5 years:
Consider the following computer vision use case. A company might want to train an AI algorithm to automate scanning of PDF and hand-written invoices, validating field formatting and triggering automatic accounts payable processes. More advanced algorithms (i.e. unsupervised learning) have emerged where the solution can be trained even on unlabeled data – i.e. a collection of past invoices without having its fields specifically labeled as valid or invalid.
However, the resulting AI solution is fundamentally limited to automating the task of text field recognition and formatting. If the company wanted to use the invoice processing bot to say detect fraud, they would need to design and train a completely new solution by focusing on other underlying features and patterns.
AI applications in the foreseeable future will be along the narrow “supervised learning using training data” paradigm. This has two strategic implications:
– Acquiring labeled data for training becomes a strategic capability and source of differentiation
– AI solutions require deep functional and domain specific human co-creation and process redesign
Fact 3: Focus on prioritizing A->B activities
The Japanese insurer Fukoko recently announced replacing human agents with AI for claims processing. Goldman Sachs between 2000 and 2016 transformed its 600 trader unit into a much leaner 200 person team augmented by machine learning.
However, not all organizational activities are suitable for AI automation under today’s narrow learning using data paradigm.
A helpful way to characterize machine learnable tasks is what Andrew Ng calls A->B activities; activities that take a set of clear, unambiguous inputs A and produces a response B.
For example, retail demand forecasting can be considered an A->B activity. By taking a diverse array of input signals such as seasonality, regional sales history, point of sales trends, social media signals, and pricing sensitivity data, the algorithm can likely achieve human level prediction of future demand.
Financial trading is also an A->B activity. A trading algorithm would take a set of inputs such as historic prices, macro trend drivers, arbitrage rules complied from past traders, etc. and produce an output of either Buy or Sell. Making the right trading call (B) based on inputs (A) may be a hard problem given the underlying unpredictability of the market, but an AI solution is still attractive if it can outperform humans over large number of trades.
See figure below for a representative sample of other A->B tasks suitable for AI as well as activities that are not suitable for AI automation.
Mapping enterprise processes and activities into a spectrum of A->B vs. non A->B categories can help managers ensure a systematic opportunity scan for AI automation and augmentation strategies.
Fact 4: AI Adoption requires more than technical feasibility
Our client experience suggest that some AI applications will achieve faster adoption rate than others, even though the underlying technical requirements are comparable. Companies need to consider broader deployment adoption drivers to ensure that its portfolio of A.I. initiatives balance unlocking near impact value with longer term aspirations. Key adoption drivers include:
One-time cost: the initial capital outlay for developing the AI solution such as algorithm development and training data acquisition. The availability of open source access to AI algorithms, and pay as you go, “AI as service” platforms can help reduce fix cost hurdles. Access to training data often becomes an expensive bottleneck is a key source for differentiation.
Switching cost: the associated costs and hurdles in displacing the current solution with a new AI solution. This includes both technical hurdles such as the ability to open the AI algorithm black-box to trace and explain decisions and also human hurdles such as political, cultural and change resistance barriers.
Eco-system needs: the need for complementary technologies as part of an integrated solution. For example, an AI solution that requires integration with innovative IoT sensors and emerging robotics will experience higher adoption complexity.
System externality hurdles: the extent to which the AI solution exhibit network externalities, where the value of adoption increases with more adoption.
See the figure below for three illustrative A.I. use cases with different adoption challenges and potential timeline to broad organizational impact:
Consider the use case for automating sentiment analysis based on consumer voice or chat to enhance call agent interactions. The solution has significantly higher switching cost in terms of cultural and risk barriers. The client is much more sensitive to starting small to minimize negative consumer impact. Even if the solution works, the client also needs to redesign their end to end training process to allow the AI engine to push suggested agent responses. Lastly the solution has relatively high network externalities, where higher adoption will generate more training data to further improve performance but collecting the initial critical mass of training data will take time and leadership “leap of faith”. Given such complexities, an AI sentiment analysis bot will likely exhibit a longer 7-10 year rather than a rapid 2-3 year adoption horizon.
A.I. automation use cases are rapidly becoming realities across organizations and value chains. Business leaders should start today to adopt a disciplined and portfolio based approach to develop machine learning capabilities, data and partnerships to ensure relevancy.
Based on our whitepaper: “Unleashing the Power of AI for Enterprise Automation” | Read Chinese version here
The author would like to thank Eric Gervet, Florian Dickgreber, Sean Monahan, Denis Bassler, Joseph Edwar, and Petr Materna for their feedback.