Artificial intelligence is the darling of the moment for the technology press and the corporate meme-stream. So much so, it’s forcing some CFOs into tough discussions with executive teams that are overly anxious to gamble with current assets in hopes of gaining future competitive advantages.

Michael Hu

AI certainly holds serious potential, even as research continues to map its exact capabilities.

But CFOs are rightly concerned about overspending and creating enterprise-wide, inflexible AI/IT platforms rather than scalable solutions and capabilities. As we saw with recent overspending on data lakes and clouds, when it comes to technology, imitation for imitation’s sake is dangerous.

Whether executives see AI as a technological magic wand, eliminating all of their current and future woes, or as just the latest “flavor of the month” technology, one thing is clear: there is significant investment risk associated with an incorrect assessment of AI’s value.

Executives who are too optimistic in their assessment of AI’s value risk overspending, failing to achieve projected results, and disrupting their organization. If their view of AI is too pessimistic, they likely will underinvest and potentially fall behind competitors.

When it comes to evaluating potential AI spend, the standard battery of technology risk-assessment questions still need to be asked, including:

• Is the promise of this technology real today?
• How and where has it proven itself in applications similar to those used by your company?
• What mission-critical functions will it expand and/or enable?
• Where does AI fit into your existing portfolio of planned innovation or other business unit spend?
• What are the direct and indirect costs associated with implementation?
• What talent requirements are associated with adding an AI capability?

There’s little question AI will end up being a transformative technology rather than an enabling one. Still, randomly experimenting with ad-hoc AI pilots without an overarching strategic blueprint will inevitably lead to pilot purgatory at worst and the development of limited point solutions at best.

Based on emerging client work, 60-plus interviews on enterprise AI across a number of sectors, and a year-long collaboration with the World Economic Forum on AI’s impact on future production, we have developed a set of guiding principles for building executable, scaled AI strategy and capabilities:

Step One: Identify the Right Use Cases

First, identify actual business problems with tangible impact and develop a list of underlying issues or pain points that AI can address. Start by conducting a systematic scan of current pain points — at the right level of activity specificity — across functions and value chains to identify overlapping activities and needs.

Note that assessing whether bots can automate customer-care calls and interactions is shooting too high. No AI technology exists today, or will in the near future, that can handle such a broad range of tasks and activities.

Drill deeper to surface granular-level problems, such as newly trained customer call-center agents’ lack of experience identifying “best responses” during email inquiry sessions (an issue that requires machine-learning solutions that perform text mining and question-and-answer inference). The solution to these problems will become the foundation for designing and prioritizing AI-enabled solutions and use-case development

When it comes to prioritizing potential use-case opportunities, companies need to ask themselves some basic questions such as:

• Is the underlying AI technology mature enough, or will it be mature enough in 24 to 36 months, to enable the use-case solution?
• Does the organization have access to sufficient data to train and test the AI use-case solution?
• Can these use cases potentially lead to “high-impact” solutions in terms of providing insights, gaining efficiencies, or speeding up decision-making?

Step Two: Manage a Portfolio of Initiatives      

Next, companies should manage deployment and scaling of AI use cases as they would an investment or product-innovation portfolio, optimizing both near-term and longer-term value creation while mitigating risks. They should sequence or group use-case pilots rather than evaluating them individually.

Multiple use cases may need the same training data, creating data-synergy opportunities. Build a data lake by combining shelf point-of-sale data, competitor promotional data, and online sales data. The lake can be leveraged to train AI across multiple use cases — for example, predictive demand forecasting as well as automated inventory allocation and replenishment bots.

Also, the performance of one use case may benefit another. AI that performs “sentiment analysis” can benefit a downstream AI use case for “suggested response to call-center agent based on question.”

Step Three: Engage the Right Ecosystem Partners

Finally, in order to develop rapidly deployable proofs of concepts, companies should leverage internal experts, existing supply-chain partners, and best-of-breed AI technology startups.

While the AI-provider landscape is clearly fragmented, it broadly divides along these dimensions:

• Primary focus: Most AI vendors are specialists in areas such as image recognition, language translation, pattern recognition, or predictive analytics, not generalists across technologies
• Niche players, which focus on developing specific industry and functional use-case solutions and have the benefit of calibrating their algorithms on relevant training data across multiple clients
• Platform players offering more foundational AI algorithm capabilities rather than domain- and function-specific expertise

Companies should leverage platform players to gain access to underlying technologies in order to pursue an in-house path for the development of use-case solutions. This is especially true if a company either has training data that’s highly differentiated from what external niche-solution providers offer, or needs to protect a proprietary process it hopes to integrate into subsequent AI solutions.

An investment trading firm looking at deep learning to improve its algorithmic trading predictions would likely source the underlying technology APIs from Google or IBM but develop the use case in-house. That would be done in order to keep its vast trove of trading data and processes proprietary.

Conclusion

While the jury may be out for some time on AI’s full potential, it is clearly one of the foundational pillars of the technologically enabled future.

The CFO’s mission is to separate fact from the “hype cycle,” identify current and likely use cases, develop a methodology for vendor selection and measurement, and integrate near-term and mid-term AI strategies into their existing technology investment plans.

Michael Hu is a partner in global management consulting firm A.T. Kearney’s operations and performance transformation practice.