Artificial intelligence (AI) and advanced analytics have already approached and sometimes exceed human ability to process data and information. Based on the trajectories of previous transformative technologies, it is virtually impossible to predict what organizations and machine-aided or led decision making will look like 10 years from now, let alone one year from now.
Read another contribution by Alexis Fink and other co-authors on the next chapter of work in the age of AI
Yet for all that seemingly unbounded potential, there are substantial limitations to what AI can help solve – defined by availability of machine-readable critical information.
The term “artificial intelligence” can easily mislead us. Most AI applications are elaborate pattern matching processes entirely reliant on the input data. Organizations have an ever-growing mountain of data. Yet much of it isn’t terribly useful because it captures transactions for financial, legal, operational, and HR purposes. And transactional data lack the contextual information that help illustrate why things happen at the organizational and business process levels.
These limitations put a ceiling on where AI can help in an organization. Attempts to give AI responsibility for addressing organizational effectiveness is a recipe for failure: They simply don’t have access to the data needed to make useful assessments and recommendations. You can’t optimize for what you can’t see.
The information we can’t easily see often hurts us
To illustrate, consider three fundamental truths about organizations:
- We don’t know how to properly measure and evaluate decision making except for imperfect, after-the-fact retrospectives.
- We don’t know how to structure rewards so people focus on the most important objectives consistently and reliably.
- We over-rely on available data while under-emphasizing critical information not captured in ERP and HR systems.
The challenge is one of missing and incomplete information too costly to measure in real time efficiently. The data absence means that AI can’t solve these fundamental problems and they will always be with us.
We’re already seeing AI improve lots of things – but those are mainly around operational efficiency, not organization effectiveness. Building a thriving organization in the AI era requires both – the power of AI and the insight of humans.
Making good decisions on business strategy, organization design, and the operating model
The most perplexing challenges for an organization’s senior leadership happen at the system level, combining elements of strategy, organization design, operating model principles, and byproducts of the people in the system. The human factors include the culture and ways of working; heuristics about managing business processes; norms about how to resolve conflicts; ways of holding people accountable for behaviors and performance; and more.
We’re starting to see vendor solutions aimed at capturing implicit knowledge, but these are primarily aimed at applications like onboarding vs. capturing things like culture, political power within an organization, and norms around decisions rights. It is not economically feasible to record, transcribe, and analyze all interactions within an organization, so the required information is rarely if ever captured in a systematic way. It’s trapped in PowerPoints and contracts scattered across individual laptops, and in the implicit knowledge and habits and beliefs of the people who make up the organization. That reality puts effective and accurate organizational-level diagnosis and problem solving beyond the reach and capability of AI.
Just as important, organizational decisions often are explicitly made to change direction from the past – something that AI systems trained on the past may struggle with.
Structuring rewards
Steve Kerr, the management scholar, not the basketball coach, famously called out in his 1975 article (“On the folly of rewarding A while hoping for B”) the fundamental problem with reward systems: We often measure what is easy and obvious rather than what we really want. This leads us to suboptimal rewards systems. Which in turn requires that we deploy many other levers to entice and reward the desired behaviors.
What results is a rich mix of carrots and sticks that address how people behave over longer time horizons than the yearly performance management cycle. Promotions, access to key assignments, protecting people from downside risks such as layoffs, offering predictability in schedules, investing in effective managers, and more are all part of the holistic view of rewards as more than just compensation and bonuses. Individuals’ motivations and priorities and lives outside of work matter a lot here too.
Though modern performance and reward systems account for an increasing variety of inputs and criteria, a perfect rewards system remains out of reach. AI can effectively and efficiently ensure pay equity, but falls short for perfectly tuning rewards for best alignment with optimal motivation. In fact, such a system might not even be legal: It likely would result in substantially different reward mixes for different individual people doing the same job.
Sourcing data for decision making
The problem with accurately measuring the human behaviors needed for effective reward design are a microcosm of the universal challenge with corporate data repositories: No organizational database contains all the key information needed to diagnose and solve many of the most important challenges that arise from the humans in the system.
The problem is both global and local. Globally, corporate ERP systems (Oracle, SAP, etc.) have incredible amounts of transactional information. Yet accurately diagnosing operational challenges usually requires digging into what is happening with specific business processes, tapping critical information that is missing in the data repositories. Collecting and analyzing that data accurately – to understand the meaning and context – is inherently a human capability that AI cannot replace – though it can augment.
The local challenge is that the data stored in central repositories is only a fraction of what’s available throughout the enterprise. Data exist in a myriad of silos, architected in a way that hampers combining them: Substantial technology investments are needed to create universal databases. Just as important are the investments of human time to make sense of bespoke data that do not fit into the universal data definitions that underpin enterprise systems sold by vendors.
Conclusion
Will AI eventually be able to help guide the analysis and solution design to complex challenges that derive from the structure of the organization, the design of the operating model, and the complexities of all the humans that make the system run? Perhaps. Yet for the foreseeable future, that will be an aspiration, not a near-term expectation, which leaves the imperfect humans firmly in charge of organization effectiveness.