Blog: An illustrated guide to AI for the handsome corporate executive
A lot of companies are making headways with machine learning (ML) these days. The leaps in image recognition and natural language processing enabled by recent advancements in deep learning and massive increases in available data have put AI* on the corporate radar almost everywhere. ML has outperformed humans and enabled companies to process service requests at a scale quite simply not possible with a 100% human service desk. It has seen successful applications in domains ranging from fraud detection to medical image analysis, from machine translation to personalised recommendations. Since quite a number of the headlines in ML are made by the GAFA lot, it is only natural that you as an executive or service manager are looking towards them when you start out exploring applications of artificial intelligence for your business. Besides the obvious issues in following this approach—your company is probably still a largely offline endeavour—there are some less-discussed issues I want to address in this post.
*AI = applied machine learning
First is that the gains advertised in big tech headlines obfuscate the massive undertaking and years of development, people operations, investments and failed projects at the base of their success. Of course, this should not be a barrier for you to start exploring machine learning applications, but it can be useful to keep this in mind when building your first machine learning team or sponsoring your first proof of concept — ML is far from mature, and some projects will require large upfront investments for them to be applicable in an industry setting. To give an example, if I were to translate the diagram above to an LSE environment with no proper data infrastructure in place, it would amount to roughly the equivalent of 10 FTE working two years to get anywhere near a basic scalable machine learning platform. And all that time and money will have only bought you some infrastructure (models not included). Once your teams reach that point, expect them to be working full time on data pipelines, change management, operations, integrations, and model development. That’s a significant investment for an LSE — anywhere between EUR 1M and EUR 10 M depending on your ambitions, size, and geographical location. Including the required prior investments in data infrastructure (your data lake/moat/swamp/drain), the projected costs of such an endeavour can quickly run up into the tens of millions of euros.
Strange enough, the most common approach towards machine learning adoption in the enterprise I’ve seen is flying blind (a notable exception was Ronny Fehling’s vision for industrialising AI at Airbus; he’s since moved to BCG). It seems as though the sheer amount of hype surrounding artificial intelligence has excused your typical board member from the responsibility for sensible project management and long-term strategic thinking. This is a dangerous situation to be in, since a lot of these ‘AI everywhere’ investments are flowing directly from the fact that everyone is investing in AI, rather than from a viable business case and release strategy. And we all know that FOMO, like financial markets, is unpredictable at best. With data having been compared to oil (spoiler — it’s not, different externalities), AI to electricity (also incorrect, lots of routine stuff doesn’t require cognitive capabilities), and ML to cooking (for practitioners), I’d like to propose a new analogy, that will hopefully allow you to better understand the trajectory that the adoption of AI will take in the global economy in the next fifty years: that of flight.
With all the current excitement for instant drone-based delivery (let me drone my bagel), it is easy to forget that has taken more than 100 years to get to the point where flight will be a true commodity. Even if we reach that point in the near future, the majority of the expected 855 billion USD in airborne revenues generated by commercial airlines in 2019, will come from transporting people — a business model as old as the first commercial airline. Similarly, when we look at machine learning, the primary value added contribution to our 21st century world as measured by operating income and stock prices is still by and large paid advertisements. This begs the question of whether we can expect industrial applications of machine learning to quickly evolve beyond personalisation-based applications in the years to come. A lot of people are betting on Internet-of-Things (IoT) applications. While the IoT is indeed generating a lot of data, without economies of scale and proper infrastructure the value of ML applications in this domain will not outweigh their costs. And industrial-scale ML systems control — different from a POC — requires quality control, continuous maintenance, and ongoing improvement — a hefty price tag if the improvements offered by ML are only marginal. That is to say, I don’t expect that running the ML analogue of an air traffic control tower for smart refrigerators will be economically viable. So far predictive maintenance is the one application outside of the consumer domain in which ML has found widespread adoption. On the other hand, ML applications in the consumer domain have proven themselves time and time again. Human-machine interactions have benefitted enormously from ML. We are witnessing AI become the primary consumer interface with machines for the vast majority of humans on this planet, through search engines, smart assistants, smart cameras, and autonomous driving among other things.
This is a dangerous situation to be in, since a lot of these ‘AI everywhere’ investments are flowing directly from the fact that everyone is investing in AI, rather than from a viable business case and release strategy.
To return to the analogy of flight, besides the onboard crew that operates the aircraft an airplane requires ground crews and sufficient infrastructure to make sure the people and cargo on the routes the airplanes fly link in with the rest of the economy. In other words, airlines are sustained by a bigger value chain and ecosystem outside of the airline industry. So is ML — it doesn’t reside in a vacuum. The team operationally responsible for developing and maintaining the models — data scientists, data engineers, machine learning managers, business analysts etc — is part of an enterprise value chain and wider IT ecosystem consisting of enterprise applications, business processes, data infrastructure, and compute resources. While obvious to the people that design and build airports, these externalities are often blatantly overlooked in machine learning projects. Even though a well-designed machine learning algorithm will certainly not harm your chances of success, without proper integration of the model predictions into internal or customer-facing business processes—machine learning UI and UX — and a wider value chain and ecosystem, your project is almost certainly guaranteed to fail. In fact, in the absence of empirical evidence (of which I am aware) I’d like to bet that the success or failure of a lot of ML projects is influenced more by the corporate sponsor and stakeholder than by whatever the ML team itself manages to produce during the project.
Which brings me to the second point. Creating value with machine learning is more often than not an organisational rather than a technical problem. As such, it is not one that machine learning practitioners can solve independently of the rest of the organisation. A lot of machine learning projects that aim to create business value fail because they require collaboration between business people and technologists. This means that companies that are venturing to move into this brave new 21st century digital world not only need to adopt a new and unfamiliar technology (AI), but they also need move from organisational compartmentalisation (and iceberg risk management) to more fluid organisational lines drawn by value creation rather than business function (see this McKinsey write-up for an introduction into platform-based organisations). Once you’ve juggled these transitions successfully there can be incredible rewards in terms of improvements in customer interactions, cost efficiencies, and general awesomeness to reap from your bet on (investment in) machine learning. Since the digital tools companies have at their disposal today remove the need for most direct human-to-corporate-human interactions, potentially any cross-functional team can interact directly with their customer through digital interfaces, obviating the need for the sales, HR, or service personnel that was once the face of the company. This logic holds both for teams where the primary customer is internal (e.g. HR or finance), and for teams with external customers (sales, service delivery, etc.).
To reiterate, AI needs to be embedded in business processes to create value. Consequently, this means touching, changing and enhancing deterministic business logic implemented in the company’s IT landscape. Which, to tell you upfront, will result in a lot of work, and then more unexpected work on top of that lot of work. And on top of this rework of your existing business processes, operational AI-driven products also require a machine learning platform that facilitates the rapid development, deployment and monitoring of a multitude of predictive processes embedded in the firm’s products, businesses and operations. That means end-to-end data traceability, low-latency data streams for near real-time processes, and data access of cross-functional teams to all data streams running through an organisation. In this sense, machine learning applications are quite different from previous waves of automation. The introduction of automated cognitive and probabilistic capabilities brings with it an increased need for quality control of data-driven processes. And that — unfortunately for you — requires increased and improved collaboration between business stakeholders and technologists.
The strange thing is that often companies will leave the make or break part of the ML project to the data scientist. Whether this data scientist is an external contractor or internal employee, he or she is typically ill-equipped to deal with company politics, and often lacks the corporate mandate to make changes to existing business processes. Which leads to those dreaded MLPOCs that fail to deliver any business value. Poor graduate and post-graduate level educated data scientists apply the workflow they’ve learned in academia to solve a business problem. Somebody finds a server somewhere and the model is scheduled without monitoring while the data scientist moves on to the next MLPOC. Everybody’s happy (we’ve done AI!), and nothing has really changed. This is a very black-and-white representation, but it serves to highlight one of the main challenges in industrialising AI. Without proper design and project execution, machine learning projects run the risk of getting caught in a no-value-added ML proof of concept (MLPOC) in which executives author ML projects out of FOMO, data scientists apply what they’ve learned at graduate school, and no-one is willing to take the responsibility for making sure the project generates real ROMLIs (return on machine learning investments).
In this sense, machine learning applications are quite different from previous waves of automation. The introduction of automated cognitive and probabilistic capabilities brings with it an increased need for quality control of data-driven processes.
Put succinctly, the haphazard nature of the machine learning efforts being funded today are a recipe for disaster. On top of this, not everyone in your organisation is one of those hopeless optimist that thinks ML has the potential to improve working conditions worldwide by automating routine cognitive tasks, thus freeing us humans up for more interesting work. There are a lot of folks out there that scared that automation will take over their job, scared of losing the power to make decisions, or just scared of change, which lead to the protracted political shadow and trench wars that are slowing the machines from taking over the world (bold sad-faced emoji). The long-term risk for both AI practitioners and consumers is that after a slew of overgenerous promises and underwhelming results businesses will continue to operate as usual for decades to come. And that’s a shame, because it means your customers will be forced to continue living in the technological equivalent of the 1980s. To summarise, there are several preconditions that need to be present for any machine learning project to work:
For the ML team,
- Data infrastructure for the machine learning models to tap into.
- Compute infrastructure for the machine learning models to train on.
- A path to production integrating the model predictions into applications.
- Monitoring and AB-testing infrastructure to measure the impact of the model predictions on business and analytical KPIs.
On the business side,
- The identification of business processes where automated predictions can be applied with enough ROMLI to warrant the effort: high volume, low-latency, high-cost, routine cognitive, hazardous, etc. business processes.
- Aligning cross-functional stakeholders around business processes that can be enhanced with or facilitated through machine learning.
- Developing the hard constraints — the epistemic bounds — within with the machine learning models can operate.
- Monitoring business KPIs and providing (preferably statistically grounded) feedback to the ML team.
As you can see from the bullet points above, the ML team has the easy job in all this. The hard part is aligning and changing your organisation to allows you to take full advantage of the opportunities offered by the technological advancements in machine learning. This is where most ML projects continue to strand today. Since the ML team or its direct sponsor often doesn’t have the mandate to work on cross-departmental business processes, if you want to get serious with ML, you’ll need to supervise and guide these projects personally, or appoint someone with a wide enough mandate to spot and facilitate ML-enabled changes to existing business processes or develop new processes based on ML capabilities.
Disclaimer: I am a freelance machine learning engineer, so it is in my best interest that machine learning lands successfully in organisations and provides my clients with maximum ROMLI (return on machine learning investment). Have a cookie.
The images are movie stills from Alfred Hitchcock’s 1959 North by Northwest.