Background vs Foreground AI
It’s the quiet stuff that matters.
Background
Why are we not swooning over the spell-check function in our writing programmes? Or when Apple Map accurately estimates our time of arrival while both traffic condition and our driving speed is changing dynamically? Or when our email filters out spam despite it not obviously looking like spam. These are all examples of what I would label as “background AI” — the non-sexy AI algorithms that get the job done quietly and without fuss. These AI solutions have been around for decades, and yet the world wasn’t a fraction as excited as the day Midjourney or ChatGPT unveiled their generative prowess. These new generativeAI solutions require the user to actively engage with it; we use words like “co-create” to project a semblance of control. I call these “foreground AI”. In the scheme of things, it’s the background AI that will ultimately provide more utility than the foreground AI. And so I dedicate my 49th weekly article to the topic of why this perspective can prove useful when working towards AI success for an organisation.
(I write a weekly series of articles where I call out bad thinking and bad practices in data analytics / data science which you can find here.)
More Than Semantics
Background AI can be described as AI-powered solutions that do not require iterative interactions from the user. The battery management software on your iPhone is an AI software. It makes its own decisions to accelerate or slow down the charging of your iPhone without any human inputs. Solutions such as spell-check simply need the user to accept or disregard the recommended changes. The same goes for the AI-powered recommendations on Netflix. Because we don’t “see” the AI busy at work, we take it for granted and simply expect it to work.
Foreground AI can be described as AI-powered solutions that require iterative interactions with the user. An AI-enabled chatbot would be a classic example of a foreground AI. Microsoft co-pilot functionality in its Office360 product suite would be another example. One can now argue that certain AI-powered recommendation engines can become foreground AI — e.g. if the recommendation output requires successive rounds of decision-tree style inputs from the user to arrive at the final converged recommendation. We are seeing examples of such in the customer relationship management (CRM) domain.
Why should we care about this definition of background vs foreground AI? It’s just semantics, isn’t it? Not really. It’s really about the user experience (UX).
AI UX
One of my clients defines user experience along 4 dimensions — utility, usability, trust, and appeal. I honestly like it. It’s succinct and practical. (I’ve written here and here about the challenges of measuring UX, including referencing these 4 said dimensions.)
Now, background AI typically has a lower cognitive load on the user population, and can be scaled easily given its lower compute cost as well. And because of its narrow use case implementation, it can be made very accurate in achieving its desired output. It should be obvious at this stage that given the attributes of background AI, it confines itself largely within the utility and usability aspects of UX, and much more weighted towards the former.
Foreground AI, on the other hand, typically have higher cognitive load on the user population, and is harder to scale due to its significantly non-zero marginal cost (e.g. ChatGPT’s marginal cost is huge!) And because of the exposure to overt human iterative interactions, its output accuracy tends to be relatively lower. It should also be obvious that given these attributes of foreground AI, it plays in all the 4 UX dimensions, including trust and appeal. Usability and trust are paramount in its design.
Value Extraction
I argue that the value extraction of background AI surpasses forground AI. Background AI reduces process friction and is not in direct competition with their human counterparts. They are much more scalable. Their implementation use cases are much more diverse, and in many ways, extremely creative from an engineering perspective. Consider how the AI software in your iPhone battery management programme gives you up to 20% more battery life; a non-obvious use case.
Organisations should logically focus on background AI to lift their business performance and capabilities before getting into the foreground AI space. This is because background AI will raise productivity with a lower cost of investment (both in terms of capital expenditure and employee re-skilling).
Foreground AI requires much more thoughtfulness and scenario planning, due to the unpredictability of human interactions. So much more things can go wrong. Those who want to get into the foreground AI domain must therefore make a strong case that the problem cannot be resolved through the implementation of background AI. That shiny new Gen AI toy is sometimes just that — a shiny toy. So don’t get sucked in without doing some serious homework. Many of the sexy foreground AI use cases being touted are in fact a combination of background and foreground AI working in tandem; again reinforcing the value that background AI brings to the table. Consider the case of Apple’s Siri or Amazon’s Alexa. The voice-activation and voice-recognition is “sexy”, but the actual work is being done by background AI.
Conclusion
Background AI is non-threatening, easier to develop and scale, and achieves the prerequisite accuracy thresholds. In contrast, foreground AI is noisy, expensive, and still very much an unproven technology (from an application and value extraction perspective). As the hype in generative AI sends the stock market into overdrive and senior executives try to outdo each other in getting press mileage for their (sometimes pretentious) Gen AI adoption, the real economics will start to hit home fast. Already, there are signs of the AI bubble stretching to bursting point. In writing this article, I wanted to present a simplified approach to thinking about AI adoption, using a background vs a foreground perspective to elucidate value extraction. The economics thus far would suggest that humans will not be replaced by AI any time soon.
