Orchestrating a Successful Enterprise AI Transformation
By Deb Goswami, Data Scientist Lead, Traveloka
Deb Goswami, Data Scientist Lead, Traveloka
The artificial intelligence (AI ) revolution is well and truly underway. An article by the Financial Times reveals that the number of AI-related patent applications globally increased from 18,995 in 2013 to 55,660 in 2017 (as per analysis by Geneva-based Wipo). This means that over 50 percent of the entirety of AI patents have been generated in the last 5 years!
Fuelled by an increasingly global digital economy, massive data and a proliferation of consumer compute (mobile), this surge in AI has led to a significant shift in how organisations are approaching innovation enablement. Personalization, chatbots, recommendation engines and augmented human-computer interfaces are just some of the ways in which the world has changed around us.
Surveys conducted by Deloitte indicate that up to 37 percent of C-executives for large firms polled are considering or have already set up a Centre of Excellence (CoE)/Innovation Hub to enable AI transformation within their organisations. 88 percent of those polled are considering increasing their funding efforts in the coming year, with 82 percent of the respondents already claiming a positive financial return on the AI investment.
So you want to AI, but how?
With all the hype around AI, it is only natural for the surge in enterprise AI transformation. However, numerous initiatives to set up Innovation Hubs or CoE’s for AI fizzle out with little to no measurable impact. Such failures have the added cost of setting back AI engagement by years, which significantly hinders the competitive edge in the current innovation-centric landscape. In fact, the above Deloitte Survey indicates that the greatest challenges faced by early adopters of AI are implementation, integration, cost and effectively measuring/proving value. The canny reader will note that this boils down to...well pretty much everything that matters!
It is apparent that there is an issue where organisations are struggling to find a right implementation pathway for their AI transformations.
Disambiguating the value proposition
It is essential for any organization looking to enable this AI transformation to ascertain exactly how committed they are to this process, and what they are willing to spend to get there. Many such initiatives fall prey to the trap of doing AI for AI's sake. This certainly backfires in the long term, by misdirecting the initial enthusiasm and engagement. In the absence of a clear strategy to ascertain impact, there will always be the temptation to fall back to the narrative - we are simply not ready as an organisation to do AI.
The most critical step when crafting an AI transformation roadmap is to correctly identify what success looks like for your organization. I propose three key areas to drive this assessment – Impact, Time and Complexity.
Impact
One of the primary reasons that AI transformations fail to integrate with core business units is due to an inability to concretely demonstrate value-add or ROI.
The most critical step when crafting an AI transformation roadmap is to correctly identify what success looks like for your organization
At the same time, not all organisations are looking to frontline their AI teams on business critical problems. Instead, their core focus may be geared at branding or Proof of Concepts (PoC’s) to drive greater AI awareness within the company. Regardless of priorities, it is important to acknowledge the aims and set up the teams accordingly. An innovation center that is structured as a cost center cannot innovate.
Time
In addition to solving the right problems, commit to a time-frame in which results are expected. Crucially, ensure clarity on the time-frame within which the above impact needs to be delivered. In order to ensure that these timelines are tethered to reality - who with real-world experience in implementing AI for a related domain has vetted this timeline?
Complexity
Finally, it is also important to ascertain the technical feasibility and complexity required to build out for.
Taking an example use-case to build a colloquial speech analysis product – does sufficient labelled data exist to train a machine-learning model? What is the cost and time of annotating this? Will this be a constant effort? If these models are being deployed as internal APIs, what is the effort associated with implementing a monitoring framework? Has the impact of leveraging platform interactions to label data for further supervised learning problems been evaluated? What does Quality Assurance in this domain look like?
This is often a good point to start hiring the first (senior) data scientists, who should be able to formulate an accurate estimate for these.
Measuring and Enabling Success
Subsequent to crystallizing the value-proposition, it is time to formulate an appropriate evaluation paradigm and structure the CoE correctly.
When it comes to evaluation - Have metrics been identified that reflect a successful implementation? Linking back to the impact section, AI output metrics that are synergized with business KPIs are often an easy place to start. Avoid the trap of using machine learning metrics for individual models (like AuC, precision or recall) to evaluate success. Instead, validate what a particular AI product has enabled (e.g. KYC time-reduction, cost-per-acquisition)
When it comes to structuring the CoE, it is often the case that an internal champion is appointed to spearhead the unit. In such setups, it is of paramount importance to ensure a separation of strategy vs execution. The unit head need not be an AI expert. However, their mandate ought to be strategic in nature. The execution of the solution should be the responsibility of a technical lead, who absolutely needs to have a proven track record of building AI products.
Cultural Pivots are the Secret Ingredient
The few truly successful AI transformations I have seen have been a result of close collaboration between the business units and the AI teams to identify and solve critical pain points. This can often be a difficult transition to navigate, as individuals within the business may feel like they’re ceding control to a technology or process that they do not fully understand. Thus, a virtual embedding of data scientists within the business is strongly recommended, so that the process of building the AI solution is as transparent and democratized as possible.
To conclude, I’d like to draw the reader’s attention to Andrew Ng (ex-Head of AI at Baidu) who had this to say on the topic of enabling enterprise AI- “A shopping mall with a website isn’t an e-commerce company.” In the same vein, an organisation which has an AI team is not an AI-driven organisation.
