Employing AI in Your Business?
Some questions you should address before you do
By Matthew Rogozinski
Artificial intelligence (AI), and in particular machine learning (ML), is the ‘next big thing’ in business, following right on the heels of digital transformation. However, while its promise is real, it comes with significant challenges and requires smart strategic choices that take full account of not only of the technology’s advantages and limitations, but also its context-specific costs and benefits. Some important choices need to be made early, before major AI planning and deployment decisions are taken.
Why is AI occupying so many executives’ minds? Global tech companies have employed predictive analytics or categorisation applications for some time, using this form of ML to strengthen their competitive advantage. For example, Netflix, Amazon and LinkedIn use it in their recommendation systems; Amazon, Google, Apple and Samsung in their voice-recognition-based virtual assistants; Apple and Samsung in face recognition to unlock hardware; and Facebook and Twitter in the content selected for social media feeds. In addition, ML is used in advertising (e.g., ad placement), telecommunications (e.g., customer churn management), financial services (e.g., credit scoring, payment card fraud detection, insurance pricing, equity investment) and pharmaceuticals (e.g., drug design).
Many other businesses are now looking to adopt ML, taking advantage of recent advances in cloud computing; the adoption of graphics processing units for ML parallel computing; packaged ML services from IBM, Amazon, Microsoft, Google and many specialist players; and an exponential increase in the availability of data to feed data-hungry ML algorithms.
While the pace of AI uptake is increasing, significant differences are observed between industries as well as between businesses in an industry—and the gap between early adopters and others is widening [1,2]. The question for many, then, is how to progress beyond initial discussions, experimentation and pilots.
One important objective is to avoid some mistakes of the recent past, when the need to rewire from analogue to digital became pressing and urgent for a range of businesses. While the need was well known, the steps to achieve it were much less clear. Too often the people accountable for digital transformation talked about a ‘swarm’ of ideas under consideration, and senior executives concerned about progress and direction were accused of getting in the way or ‘just not getting it’. The result was often a fragmentation of effort, with valuable resources dedicated to multiple initiatives that never delivered real impact.
ML amplifies the risks observed in digital transformation because it is not easily scalable across the enterprise: algorithms become unreliable if the environment or context of a particular task changes; an algorithm designed for one task cannot be used for another; and a large amount of data must be made available each time to train and test the algorithms.
This is why we recommend a coordinated, strategic approach to ML across the enterprise, starting with the identification of genuine ML opportunities in light of the business’s competitive position, strategic direction, resources and capabilities, and the off-the-shelf offerings available. Here are the questions we believe need to be considered before a decision on ML is made and certainly before a plan for deployment is developed.
1. How well is a given ML opportunity aligned with business strategy?
You should be able to describe how an application of ML to a specific business activity (task) will contribute to realisation of the business strategy. In doing so, you will likely eliminate a myriad—in a large company possibly thousands—of possible ML opportunities that, even if realised, would have only a marginal impact on competitive success.
We believe differentiation opportunities (rather than, say, efficiency benefits) should be tackled first. Data is already the next frontier for competition and ML offers new ways to leverage data to avoid the commoditisation of products and services (as we argue in our article A Level Playing Field in Digital Banking?). Genuine opportunities to challenge the current business model should also be grasped, given ML’s potential to change underlying economic assumptions.
2. Is ML the best approach to capture target benefits?
ML is seductive because it promises to solve problems we often consign to the ‘too hard basket’. But have we really tried to solve those problems? How? What factors limited the usefulness of our solutions? What does ML bring that will change outcomes for the specific business activity (task) this time?
The academic research reports the successful application of ML to new problems (tasks) but rarely compares results with, for example, traditional statistical methods applied to the same data sets. However, in some circumstances, ML may be less accurate, computationally more expensive and more effort-intensive than human learning combined with other analytical methods [3].
Some guidelines exist to assess whether a task is suitable for ML [3,4], but that’s not the same as determining whether ML is the best approach. Even rudimentary rules of thumb to evaluate the relative suitability of ML versus other approaches are hard to come by. When the outcome of comparing ML with other approaches is uncertain, we suggest rapid (say, 10 weeks) prototyping of ML and non-ML solutions for a chosen task, with the results compared (using test rather than training data) before a decision is made to go ahead with ML.
3. Can we provide sufficient data to make ML effective?
ML is often seen as an analytical technique that matches the explosion of data available from the internet of things, but for many businesses the starting point will be corporate data augmented by selected external data. The data volume needs to be large enough to train the algorithms and continually re-train them to adjust to changes in the environment. A more complex or sophisticated ML algorithm is unlikely to compensate for a smaller training data set [5]. If the application is attractive but the data is insufficient, consider starting to build the data set needed for future implementation.
Some vendors might offer off-the-shelf ML solutions trained on their proprietary data, but it’s important to determine whether the result will be a ‘me-too’ solution rather than a source of competitive advantage. Remember that investment in ML should offer a differentiation opportunity that’s likely unavailable through other means.
In addition, the output of ML algorithms can only be as good as the underlying data and this can create new problems. For example, third-party training data may be biased (e.g., customers with some characteristics may be underrepresented). These biases will be embedded in their ML algorithms. The result might be a missed business opportunity or compliance and even ethical issues.
4. What risks will the ML application introduce?
A small improvement in predictive or categorisation performance may offer a significant competitive, economic, or public benefit but the consequences of an error must also be considered. ML is, after all, a probabilistic discipline. Ideally, an ML application should have significant benefit-risk asymmetry for the business. The Amazon purchase recommendation system is a good example—in that case, the business benefits (in both revenue and brand terms) when a customer purchases a recommended item but suffers no brand erosion if a recommendation is ignored.
Further, a business is sometimes required to explain how a decision was made. This can be difficult when people acting on known rules are responsible, let alone when an ML algorithm produces a ‘black box’ answer with no auditable explanation. Telling a customer ‘we acted based on tracking your online activity’, for example, could be seen as creepily invasive rather than helpful.
We mentioned the risk of embedding a bias in an ML algorithm through the training data. Bias can also be created through the optimisation method used to tune the ML algorithm. Variance, in the form of different outcomes from different algorithms presented with the same data, is also possible [5].
ML risks aren’t by definition more problematic than those involved in manual processing and human decision-making, but they should be explicitly considered in the context of each application and the relevant regulation (e.g., the EU’s General Data Protection Regulation [6]). It’s important to remember, too, that the community and regulators are increasingly demanding clear, direct executive accountability.
5. What new human capabilities will we need and how will we access them?
The building blocks of ML include specialised computational platforms (massively parallel processing), which need to be fed the corporate data, with tight latency requirements for (near) real-time applications. In addition, ML algorithms need to be selected, customised, trained, maintained and updated. Where and how to access the scarce data science skills required for these tasks are critical factors to consider. Do they already exist in the business? If not, is it well positioned to acquire and retain them, in competition with the established market leaders? Can it access them all through third parties and still be confident ML will differentiate its products and services? Even if third parties supply the required skills, does the business need some in-house capability to inform decision-making?
As well as data science skills, ML deployment requires complementary capabilities—for example, to manage data risks, including those arising when valuable corporate data is made available to vendors, and to ensure all regulatory and compliance requirements (some of which are or will become ML-specific) are met. It’s sensible to scan the business for any gaps in these and other support requirements so what needs to be done in this area can be clearly articulated in the planning round.
* * *
Launching ML in an organisation requires expertise to be brought together from the strategy, business, operations, digital, technology, data governance, data science, HR, risk, legal and compliance areas. Perhaps ironically, AI creates an imperative for increased human interaction via close collaboration across boundaries. This collaboration should begin by bringing multidisciplinary teams together to answer the kinds of questions we’ve outlined above. It should continue when a decision to proceed is made and a detailed strategic plan for deployment is developed. We believe the workshop methodology we describe in Cut-through to Break Through is well suited to laying the groundwork to identify ML opportunities, then finalising the planning process (an article on that process is in preparation).
[1] Ransbotham, S, Gerbert, P, Reeves, M, Kiron, D & Spira, M 2018, ‘Artificial Intelligence in business gets real’, MIT Sloan Management Review and The Boston Consulting Group, September
[2] McKinsey & Company 2018, ‘The promise and challenge of the age of artificial intelligence, Briefing note prepared for the Tallinn Digital Summit, October
[3] Makridakis, S, Spiliotis, E & Assimakopoulos, V 2018, ‘Statistical and machine learning forecasting methods: Concerns and ways forward’, PLoS ONE, vol. 13, no. 3
[4] Brynjolfsson, E & Mitchell, T 2017, ‘What can machine learning do? Workforce implications’, Science (New York, N.Y.), vol. 358, no. 6370, pp. 1530–1534
[5] Domingos, P 2012, ‘A few useful things to know about machine learning’, Communications of the ACM, vol. 55, no. 10, pp. 78–87
[6] General Data Protection Regulation (GDPR), Recital 71