Operationalising AI: What’s your strategy?

Many Australian enterprises have spent years trying to justify their investments in data analytics models. On average, only half of the analytic models built by organisations will ever make it to production.

Clearly, organisations that operationalise and monetise their artificial intelligence (AI) and analytics capabilities are more likely to succeed with their customer engagements.

Tech execs gathered at a virtual roundtable recently to discuss the challenges they face when moving their AI and data analytics programs from an experiment inside their business to one that is a key part of their core operations. The conversation was sponsored by SAS.

Ray Greenwood, domain lead, advanced analytics at SAS, says that ModelOps – or how analytical models are cycled from the data science team to the IT production team in a regular cadence of deployment and updates – helps organisations get business value from AI.

“ModelOps can bring the same productivity gains and support for innovation to advanced analytics that DevOps brought to software development.”

“By establishing standardised practices and smoother system integration, previously time-consuming phases are condensed or automated. This leads to more models being built which are implemented sooner and stay at peak performance for longer.” 

“Each model therefore contributes more to the organisation’s goals compared to its cost of development. ModelOps bring rigour to the process of getting models into production,” he says.

ModelOps is the next big step to help surface the benefits of data and analytics, says Ian Oppermann, chief data scientist at the Department of Customer Service.

The philosophy and discipline of DevOps being applied to modelling will make usable what is currently not transparent and/or readily explainable.

“However, care must be taken to allow the deeper contemplation and evaluation required for modelling to not fall into the production line thinking of ‘ops.’ Standups and Kanban are great for progress and transparency but do not sit well with non-productive activities such as deeper thinking and experimentation,” he says.

Designing an efficient and scalable model of operations is paramount to the success of ModelOps, as well as a shift in workplace culture.

According to Anna Liebel, chief digital and information officer at UniSuper, “Data scientists and IT operations teams work in different ways and have different goals. As we design our new way of working at UniSuper for analytical models, I contemplate how these teams will collaborate effectively and the appropriate governance of models at scale.”

Key steps to operationalising your data analytics efforts

While an organisation’s selection of tools and techniques for AI projects will continue to evolve, putting standard processes and practices in place that cover key technical and non-technical aspects of an analytics project will increase the chance of success.

This includes ensuring executive support is committed beyond any initial experimental phase into production and beyond.

Committing to best practices around governance will also drive long-term benefit in terms of repeatability and scalability as the demand for AI initiatives increases.

“Operationalising models requires collaboration from IT, product owners, analysts, audit functions and more. Ensure that everyone is engaged and aligned early,” Greenwood says.

Glenn McPhee, chief operating officer at ManpowerGroup Australia and New Zealand, adds that advanced analytics doesn’t change the need to get the basics right – in fact, it makes it even more critical.

“It all starts and ends with data so you need a comprehensive information management framework,” he says.

McPhee says that there are three key steps his organisation sees as important to operationalising analytics efforts. Firstly, historical data needs to be cleaned, sorted and filtered and new data should be captured and stored correctly.

Secondly, data capture and reporting needs to be built into organisational processes not just tacked on at the end of an existing process. Thirdly, operational managers need to have a continuous training program put in place.

“They need to understand the power of analytics and AI and what it is capable of so that they can ask the right questions that will allow them to make informed business decisions in the future,” he says.

UniSuper’s Anna Leibel adds that the company operationalises its data analytics efforts by using data champions to create curiosity about the power of data. She says that these staff dive deep into business data to understand how it can influence the member experience and improve decision making. 

“As the business demand increases, IT will continue to mature our capability in AI, predictive analytics and data visualisation,” Leibel says.

Meanwhile, the Department of Customer Service is ‘recalibrating’ its expectations around data analytics, says Oppermann.

“The COVID-19 response in NSW showed just how much better we can understand, prioritise or explore ‘what if’ when different groups share models and assumptions,” he says.

The agency’s challenge, he says, is to keep as much of these gains as possible as the crisis recedes. Baking in the gains through willingness to share, quantification of data quality issues and willingness to use insights requires the department to capture, in governance frameworks, the “high tide” of data sharing that it has experienced during its COVID-19 response, he says.

Securing data science talent

For many organisations, finding talented data scientists to help make the most of their data can be difficult. There’s an ongoing global skills crisis and data scientists can often demand large salaries.

SAS’ Greenwood says that data scientists are often more motivated by the opportunity to make a difference in their organisations than by money alone.

“Frustration with seeing the results of their labour languish in experimental mode and, all too often, not deployed at full scale is what leads them to seek out alternatives where the chance for innovative work and the promise of their work reaching the front lines draws them away,” Greenwood says.

He advises that organisations help data scientists see the value their work adds in production as often as possible and reduce the frustrations related to manual model deployment and maintenance.

UniSuper’s Leibel adds that a data strategy which clearly defines how information will help achieve company goals will attract data scientists.

“Retaining data scientists is more complex and challenging,” she says.

“The primary objective of a data scientist is to extract business value. Our role as leaders is to design an operating model which prioritises this objective, nurture a culture of innovation and resolve critical issues,” she says.

Finally, data scientists – amongst other skills – are not only niche but critical and undersupplied in the market, adds ManpowerGroup’s McPhee.

“It’s a candidate-driven market. This means it can be expensive for companies to continue to hire to fill vacancies that could be avoided.

“This is important to understand as the time investment in retaining and the learning and development costs of developing these types of workers will be seldom wasted and will have a high payoff,” McPhee says.

He adds that retaining talent comes down to understanding the wants and needs of the individual and ensuring the ‘employer value proposition’ is broadly aligned.

“It’s about getting to know your employees and what drives and motivates them. For some, it will be money but for others, it will be about the challenge of the work they are doing or belief in the purpose of your organisation.

Ultimately, organisations that take the time to invest in their artificial intelligence and data analytics strategies as well as their people give themselves the best chance of success in the longer term. Companies that make the best use of their data and build a culture of lifelong learning for their staff will be most prepared to withstand digital disruptions across their market sectors for many years to come.