8 Lessons Learnt Implementing Real-World Machine Learning in A Large Corporate

Artificial or Augmented Intelligence (AI) and specifically Machine Learning (ML) are beginning to show real promise at delivering quantum shifts in business performance. The journey has just begun but will accelerate rapidly over the coming years. Senior executives waiting on the sidelines before experimenting will lose a valuable window of opportunity to potentially create significant competitive advantage. To avoid some of the common mistakes and misconceptions associated with ML, I've reflected on my team’s performance over the past year and thought I’d share some of my learnings championing and successfully implementing ML in a large corporate. 

ONE: Imagination and Trust are your biggest constraints. Although a rapidly growing issue, the biggest constraint is not necessarily access to scarce talent but the desire to learn, imagination and conviction to act by senior leaders in the business. If leaders don't understand or trust the basics of ML and how it can be applied, they won't begin to see the myriad of possibilities. In the end, you will spend most of your time justifying and lobbying for support vs. getting on with prototyping new solutions. Similarly, in my experience, trust will also be a big barrier to overcome. Trust in sharing proprietary data with a 3rd party, accepting an ML prediction or having to replace an existing process can all be big issues to overcome. The best way to overcome these is to start small, be specific and use internal structured data to prove the benefits…in the end, monetised insights will do the talking. 

TWO: Beware of omission bias and loss aversion. If senior leaders in the business are not prepared to take some risks and 'play in the sandpit' for fear of failure, then consider going 'underground'. There are countless examples where amazing innovations were birthed in large organisations that were kept 'under the radar'. If your culture allows for it, then consider asking for forgiveness rather than permission because you may still be debating the potential benefits when you could have already worked on proving them.

THREE: Start by developing a robust ‘use case’ strategy. We began by developing a robust use case strategy within a well-defined framework looking for opportunities across our value chain to apply ML. We then developed a ‘road map’ enabling the core business strategy through either radically lower operating costs, better customer experiences or new business models. Finally, we estimated the overall value at stake at a high level and investment story for those chosen use cases. In the end we stuck with prototyping just two use cases given the limited analytical and tech resources available. This also afforded us the time to build trust in the learning capability before scaling up.

As with any new technology hype the temptation is to allow tech vendors, consulting firms and IT department (yes, even those with titles starting with 'Digital'), to convince you that you don't have what it takes or know your own business as well as their experience would suggest. I’m not suggesting ‘going it alone’ in all cases. Knowing when to partner with the right vendor is also crucial in the long run given the pace of technological change. Exploring strategic partnerships for broader capability and integrated solutions is a big enough topic for another day. 

Often, the reality is that you probably already know what 'pain points' your customers, suppliers and staff must deal with every day – or at least you should. Knowing these and knowing how ML can augment to help solve these issues in new ways is half the battle already won. Some are obvious like: demand forecasting; predictive customer churn prompts; or predicting asset maintenance etc. Others can be a little harder to define like: micro customer segmentation; personalising online customer interactions and promotions; or logistics network routing etc.

FOUR: Clearly define the core business problem. Big data and AI isn't the answer, it’s the question. This may sound obvious, but it's easy to get caught up in all the hype and get fixated on what the technology can do and what other structured and unstructured data sets can be incorporated. Whilst this is exciting (my inner nerd), always stay focused on the problem or the question the business needs to solve for. The same set of data can have very different answers, depending on the problem / question you are trying to solve for. Leading any ML initiative from both a strategic and pragmatic operational perspective is important to stay focused and avoid 'analysis paralysis'.

FIVE: Forget traditional business cases in the early stage of development. If you allow linear 'big company' mentality to rule, then you will waste most of your time building fancy business cases for the next annual cycle of budgeting. Rather focus precious resources on developing and testing a prototype in often the same amount of time as creating a business case. In the end you will have real world data to develop a more robust business case far more accurately than if you hadn’t.

SIX: Don't wait to fix the basics. A common reaction by business and IT executives will be that you need to fix the underlying business processes and system issues before pursuing ML. Yes, investing the time upfront to make sure your data architecture, quality and governance processes are in reasonably good shape is time well spent which will save a lot of additional rework down the track. However, there will always be pressing operational performance improvements that need to be made. You may be waiting a long time for your processes and systems to all be ‘optimal’ before getting started. This could end up costing you just as much in time and money designing a new process or redeveloping an existing legacy system which is likely to be redundant in the long run.

For example, we were told the current Sales and Operational planning (S&OP) processes and systems were broken and needed to be fixed from the bottom up before pursuing some ‘shiny new ML toy’. As it turns out, there were multiple teams and parallel processes all working on more accurate and timely demand forecasting. Instead, we ignored the sceptics and created an Advanced Analytics Working Group - essentially a skunkworks made up of passionate experts from across the business and began prototyping a new ML demand forecasting tool. This was operationalized in under six months, significantly improving forecast accuracy and lowering working capital costs – a result of better algorithms and more granular data sets. All this was done in a fraction of the time and complexity it would have taken to optimise existing S&OP processes and systems.

SEVEN: A multidisciplinary and highly collaborative approach will be needed. Few people (if any) possess the business, analytical and technology skills needed to build a fully-fledged advanced analytics capability. So, having ‘business’ leaders who can clearly articulate the problem / opportunity and then work cross-functionally and partner with the relevant SME’s or vendors to translate the business needs into workable solutions is a critical enabler. Note, the emphasis on ‘business’ leader not ‘IT’ leader. Leaving the use case development to technical teams can often result in over-engineered solutions and operating models leading to blown out timelines and costs. 

EIGHT: People, not technology, are your most critical enabler. The most critical enabler or disabler to embedding ML in the organisation will be the people – not the data, technology or processes. In the end, you will still have to persuade people to integrate new insights into daily management and front-line routines. Staying with the example previously mentioned, namely a new demand forecasting tool, convincing the team they could trust the new predictive insights which in turn would ‘free them up’ to do higher value adding activities was a real challenge. Don’t underestimate the time and effort that will be needed here to change mindsets and behaviours. Even if the data clearly demonstrates significant improvements to antiquated or inefficient processes, people are seldom perfectly rational with many biases preventing them from letting go of past ways of working. Another trick to driving adoption is to run the ML process in parallel and let the user see the benefits over time for themselves – a much easier change management approach in my view.

Augmented Intelligence, the term I believe is more relevant in business, is without doubt the next revolution in productivity and growth. I hope these learnings are helpful and admit to at least some of them being a little controversial. I would love to hear more about different opinions and approaches in the comments or post any related topics you'd like me to cover.

All images from https://unsplash.com/