How to think about potential effects of AI on your workforce
Please note that I don’t believe that the technologies described in this post are not actually ‘AI’, however I acknowledge that this has become the accepted terminology so will use it here.
In almost every edition of my weekly newsletter empiricalHR I cover serious articles are reports on the effect of AI on the workforces. Most are covering it at the aggregate / economy level and whilst this is important for most people it’s not actionable. For decision-making we need to understand the implications at a level we can control, namely our organizations.
Tasks and jobs
It’s worth starting by taking a step back and considering what a job is. Jobs can be seen as collections of tasks handled by a defined person. Some of these tasks might be explicit - defined in a job description or by a process - but others won’t be.
We can often break tasks down to smaller level activities (or tasks). When doing so we get a better view of what is being done, what resource or effort is needed and possibly how it could be automated.
Jobs are often created when an organization creates a bundle of tasks and hires somebody to perform them. However, for many people, there is also a degree of self-crafting a job with slightly different task than if the organization hired a different person. This crafting is important as individual choice will have some effect on how AI can be used as a substitute.
How is the current ‘AI’ best used?
Current applications of ‘AI’ tend to work best when that task is highly specific. It’s much easier to build a model to solve a specific task than it is to build a model to build somethiing much more generic. Therefore, when thinking about how to apply ‘AI’ it’s worth thinking about very specific tasks that need performing.
If we think about what my firm, OrganizationView does - analysing employee survey comments - there are two key components. First, we need to identify which topics are present within the answers and second we need to identify which of those topics are present within each and every statement. Both of these components require different approaches and algorithms and therefore should be treated seperately, including the decision whether to apply ‘AI’ to the task.
In fact, it becomes more specific than that. A system developed for understanding customer feedback will perform less well than one designed for employee feedback. We can increase accuracy further by focusing on employees of one company and even by focussing on the answers to one question.
AI / human interaction
I strongly believe that the best ‘AI’ related approaches involve, at least in some instances, a combination of humans and algorithms. There are some tasks which humans find difficult (maybe requiring memory of very large volumes of examples) which are trivial for algorithms. There are also tasks which are extremely challenging for algorithms which a 4 year old could accomplish easily.
The challenge for the designer of AI-based approaches is to explicitly design the handoff between algorithm and human. Not only does this need to be able to identify when the other party would be more efficient, but the problem needs to be packaged in a way that it is optimised for the other party to perform optimally.
For an algorithm passing to a human we need to consider the user interface required to ensure that decisions are made both efficiently and optimally. Systems can handle probabilistic decisions easily, and if the designer so choses can act with total rationality. Humans are unlikely to do either of these things well.
By defining the activity in small tasks it becomes easier to identify where a task might be best performed by an algorithm or a human. However the ‘cost’ created by this interface needs to be explicitly considered.
Comparing AI and employee costs
Almost every task performed by an ‘AI’ solution could be performed by a human. Whether it makes sense to do so is a different question.
A good way of determining whether something makes sense to automate is to identify several key attributes:
How often the task is completed (the more times, the more likely it will make a good candidate for automation)
How long it would take a human to complete the task
How effective an algorithm is at performing the task
How effective a human is at performing the task
The cost of developing or acquiring the AI solution
The cost of human labour
The complexity of the decision needed. Anything that requires ‘common-sense’ or multiple angles of knowledge is likely to be more effectively made by a human
The cost of making a wrong decision
The legal framework & who is able to see the data.
AI systems have a cost to develop, implement and maintain. Add in the costs of training humans to adapt to the new user-interfaces and we can see that some tasks do not make sense to automate at the current costs, even if an algorithm could perform the task at, or even above, human level performance.
What is worth highlighting is that in certain countries, like here in Switzerland, with high labour-costs it might make sense to automate a task where the same task in a lower-cost market (eg India) might be more cost-effectively done by a human.
AI as an employment substitute
If we consider that AI, under certain circumstances, can more efficiently conduct certain, well-defined tasks then there is obviously the opportunity to remove these tasks from your employees’ jobs and get an automated system to perform them. Jobs, however, are in almost all situations a combination of tasks, some of which might make sense to automate and others not.
As noted above, almost all AI will require some interface between the tasks performed by the AI and those performed by humans.
Every previous wave of automation has resulted in human workers adapting to the machines, not the machines adapting to the humans. I believe this wave will be no different.
I therefore expect AI to redefine how work is done. Whilst the overall volume of tasks might reduce, work will need to adapt to get the most out of the investment in AI-based solutions
Redistribution of resources
There are numerous ways our organizations can compete. For example they can compete on price, on quality, on service, on speed. Even within the same general category organizations can choose different ways to position themselves.
AI can cause a shift in the cost of doing all these things, but that doesn’t mean that organizations will choose to focus just on the costs. Some will see AI as an opportunity to deliver services quicker, some will use AI as a way of increasing the quality of the products or services they deliver.
For most organizations at the moment employees are a scarce resource. We saw in the last recession that many firms changed the ways that they worked to retain workers rather than let them go only to need to rehire when the markets turned.
I expect that we will see many firms will see AI as a way to compete more aggressively with a similar level of employees. Some will use AI to try and drive production levels and gain market share. Some will use AI to improve the level of service they can provide for the same cost.
Changes in demand due to price changes
Some firms will use this significant change in the costs of delivering a service to provide products or services that were probably cost-prohibitive before. This is the route that we have taken. It has always been possible to do the type of qualitative text analysis that we provide however in the past for most firms it was prohibatively expensive. We have used ‘AI’ to automate large parts of the text analysis approach dramatically reducing the cost.
This cost reduction has meant that what was previously not efficient suddenly becomes cost-effective. Most of our clients’ previously had not systematically analysed text answers, be that in surveys or other HR systems. Some would have had approaches to enable comments to be read by those close to the provider (eg the line manager). Some would have not read them at all. Some chose not to ask open questions because they couldn’t understand them at scale.
By using AI to shift the cost of analysing comments our clients have typically shifted not only resources, but also their approaches. Some are relying more on text-based answers because they can deal with the responses. Some are asking different types of questions (eg asking for ideas or identifying issues) than they asked previously. Some are revisiting text data as a way of improving the effectiveness of what they’re doing. In all instances how our clients deliver their services has changed as a result of the new AI-driven service. The algorithms change how work is done and what work is done. I know of none of our clients who’ve just used our service to just reduce cost or reduce employees.
Second-order effects
The last thing to consider and the part that I think is most exciting is what changes might be seen, not as the direct result of AI being used to shift how we do things or what products and services we use but as a secondary result of these things. I think of these as second-order effects.
To see this let’s consider an example. If we were to move to a world where cars were self-driving that implies that they could drive around the block instead of needing a place to park. (They might also go and pick up a different passenger). This change would have the second-order effect of reducing the amount of parking space needed.
If we think of the amount of land-use in our cities devoted to parking what would we use this land for if we remove the need for parking space? It’s easy to see how the AI in self-driving cars could have second-order implications for the property market.
There are these types of connections and secondary effects across our societies. The advancement of automation via AI will create many new markets, or revitalise others. This will have implications on the employment market.
If we think of jobs as collections of tasks we can see that there has been a constant rebundling of tasks as new technologies appear. Thinking of AI as simply reducing the need for workers is simplistic. The truth will be far more nuanced, and opportunities will appear not only to those who create or deploy the AI systems we use, but those who can take advantage of the shifts to the economy and society that these systems will provide.