You need to manage algorithms
It’s generally believed we are entering the algorithmic age, if we aren’t already well and truly there. So how do we ensure we optimise the benefits and reduce the risks inherent in their use.
In our daily lives, at work and at home, we are impacted by data and algorithms in ways that we would never have dreamed of just a few years ago. You don’t have to look far to see examples where ‘upstart’ businesses have entered an industry and/or market and completely revolutionised the way we, as customers, interact and benefit. Just look at Garmin, Uber, Netflix, Amazon, AirBnB to name just a few.
Managing in an algorithmic age is going to be much more challenging than in the analogue era that most of us have operated in. In an analogue world we have had the comfort of traditional approaches to business – hierarchy, manager and subordinate, business units and silos. Today, and in the future, our value and success as leaders will be determined by our flexibility, collaboration and above all the management of dynamically interconnected relationships.
As executives and managers our job, in part, is to make informed predictions about a wide range of subjects including how we interpret the future, who to recruit, what to prioritise and which ‘go to market’ solutions we adopt. These are just the tip of the metaphorical management iceberg. Where before we would gather some data and do our own analysis, or in some cases rely on our gut, more and more businesses are turning to data science, including algorithms. Algorithms can perform many analytical iterations at incredible speed and scale. Add to this the 3 Vs of Big Data – volume, velocity, variety – and their accuracy further improves. However, there are some inherent risks in using algorithms that require understanding and managing.
It’s important to recognise three key limiting factors, which if we do not manage could lead to problems. Just ask Oscar Munoz, United’s CEO regarding their customer service failure on flight 3411, or YouTube and Facebook, who in 2017 ran into difficulties with their machine learning algorithms placing several ads in some unsavoury places.
Firstly, algorithms get it wrong. Yes, on average they are better at predicting than us, but they still err; after all the algorithms have been created by a human who has potential biases and limitations. Secondly, they are black boxes and don’t necessarily help us understand the true cause, or why something happened. Lastly, they are literal and will do exactly what you are asking them to do, no more no less (however, read our upcoming insight paper on deep learning). Taking a step back, it’s not that the algorithm is at fault, it’s working exactly as programmed. What gives rise to these issues is how we interact and manage them.
Here are some key lessons that can help us manage algorithms and ensure we get the best and not the worst out of them.
is important in making the most of algorithms. It’s no good releasing an
algorithm and expecting the organisation to behave differently. A data culture
needs to be established and not just in one department it needs to be
any military leader you need to spend quality time in fully defining the
problem, obstacle or opportunity you are trying address and solve. As Steve
Jobs said, “if you define the problem correctly, you almost have the solution”.
better your intelligence gathering, the better you will be able to frame the problem
about the alternative ways of moving forward and don’t over-engineer. The situation
may not require complex machine learning.
the adage garbage in garbage out. Algorithms will not be useful built on
explicit about all your goals, both the short- and long-run. Failure could
bring early success but long-term issues. For instance, short term sales
increases could come at the expense of quality and therefore long-term
profitability. You need to consider all the interacting goals when defining your
aims. We want x but we also want Y and Z.
the ‘law of unintended consequences. Involve a wide and diverse audience in the
problem identification and design. Spend quality time in considering all the
potential risks and consequences. Then, design out these unintended
possible, build in soft goals and trade-offs, however sometimes this can be too
difficult, in this case make sure you have built in a human review process that
makes decisions and acts on the information.
as many classes of data as possible, both about the subject and other related streams. It’s better to have a wider range of
different data, diversity really does matter and adds to the richness of the
prediction. Remember that one of the key benefits of machine learning is that
it can spot the subtle patterns and interconnectivity in many strands of data.
and test before ‘going live’ – iron out issues back at base and sense-check the
different algorithms and test them against each other.
careful of data that has in-built bias and that could skew the prediction.
Think through the data you are using – look for potential potholes in the data.
abrogate responsibility once the prediction has been made. As managers acting on the predictions,
consider whether it is in line with all the long term aims of the organisation.
Build in control points. All of which provides an additional quality step.
- Critically assess what you are being told – ask yourself the counter-factual questions and remember that there is such a thing as ‘guilt by association’. Just because you see positive affirmation doesn’t mean it is the cause. So, test the hypothesis and its underlying assumptions.
Algorithms and machine learning can be incredibly powerful, but not understanding and acting on their limitations can be debilitating and cost the organisation in terms of time, money, brand reputation and its future. Managing algorithms is a key skill that cannot be underestimated and needs to be treated as a key success factor in any forward-thinking strategy, as does critical thinking and analysis. As managers we need to make sure that the culture is right, it’s flexible, empowering and open. Lastly our organisations need to be data enterprises.