It’s been long six months since my last newsletter. Some things have changed but my passion remains the same: people analytics and digital transformation of HR.
Today, I’d like to continue sharing my view on the generative AI and its impact on HR. Let’s go!
In April, I shared some thoughts about generative AI revolution. At that moment, my intention was to reflect on how the newly introduced ChatGPT 3.0, that was released in November 2022, will affect the work and HR. Since that time, we learned a lot about hybrid collaboration between human and AI. Although, it still remains a domain of the early adopters, and one might say that the risks of wide introduction of large language models (LLMs) into the processes and products is heavier that prospective benefits, the market penetration of the new technology is increasing. We have already seen that the big players bet on the new generation of AI as the emerging cornerstone of their business with such examples as Google Bard, Microsoft Co-pilot, Microsoft Bing, or Adobe Photoshop 2024, just to name a few.
Knowing what we know about generative AI, and considering all uncertainties, are we able to predict the future of work? Let’s try.
What This Generative AI Phenomenon Is Really About?
Generative AI and ChatGPT in particular are hot topic, but how it really works?
ChatGPT is akin to other LLMs. This is a new generation of technology. Unlike the traditional transactional systems, it isn’t basing on the architecture of the database and data the user enters to the system, but on the parameters used to train the system. Josh Bersin published an excellent report Understanding AI in HR that I highly recommend you read if you like to learn more about the technology. Below are two definitions that comes from this report:
A language model is a probability distribution over a sequence of words. To put this another way, when given a sequence of words, it assigns a probability to the sequence. Language models have been in existence for a long time and predate any of the developments with deep learning. They are typically used to complete a sequence, meaning they predict the word that comes next when given a piece of text.
A large language model is a language model that uses the latest technologies of deep neural networks with a very large number of parameters trained over a very large corpus of text.
In short, ChatGPT works this way that it extends the text based on the preceding context. It uses probability to match meaning of the next words and predicts them. Additionally, there is an element of randomness, because it turns out that during prediction some lower-ranked words result in more engaging texts. Additionally, this mechanism ensures that the output is varying for identical prompts, enhancing creativity in the text generation.
The whole revolution we observe since November 2022, when OpenAI introduced ChatGPT 3.0 to the wide market, has one foundation: the amount of data used by OpenAI to train their model. OpenAI increased the number of parameters used to train the model to 175B, compared to 10B that was the previous record. This allowed them to reach completely new quality in LLMs.
Don’t Let Chat GPT Confuse You
As simple as it is in the assumptions, the results of LLMs production are very useful, often in good quality, and among other - robustly fast! Thus, you might want to know how to incorporate the tool to your current application landscape and processes?
First of all, we have to understand and acknowledge that the amount of data we produce in our companies, especially in HR, won’t be sufficient to train LLMs like ChatGPT. Most likely, we will have to partner with vendors who are offering tailored solutions and then leverage the data we have in our core systems to achieve the results we want. I think there is not that many practical implementations of generative AI in HR yet. We observe a lot of uncertainty in the field where the experts are trying to understand the value, calculate ROI, remove potential bias, and adjust processes accordingly.
But don’t let the ChatGPT confuse you. The generative AI is not about the next generation of chatbots. This is the new type of software that is capable to create output based on the data it was trained with.
There is one subset of that class of software that is briefly called “co-pilot”, but there is also another one which is utilizing deep learning to significantly speed up our business (HR) processes and introduce utterly new quality to it.
Deep learning neural networks greatly increase the number of layers and interconnections. This was made possible because of the increase in GPU (graphics processing unit) capacity and a drop in price of compute. GPUs allow for multiple parallel computations, which makes deep neural networks possible because they rely on matrix multiplication.
Machine learning refers to the ability of a system to “learn from past performance to improve future performance.” A system that predicts retention and gets better over time would qualify as a machine learning system.
According to Josh Bersin and already mentioned Understanding AI in HR, vendors should specialize in narrow domain and train their systems accordingly. Then, we choose the right one and implement it to get the best results and maximize the benefits.
Suppose you buy a recruiting tool that has been deeply trained to find and select great engineers and software professionals. Would it accurately predict the best people to work in a coffee shop or manufacturing facility? Maybe not, but if the system is well trained with data in those job domains the answer may be “yes.” However, if you buy it out of the box and you are the first customer with that use case, the system may not work as well as you’d like.
We believe these systems will quickly become specialized by domain. Talk with your vendors about the industries they serve and where they’ve been the most successful. Chatbots that work for sales and lead generation, for example, may not understand or be trained for the questions and issues you face in sourcing and recruiting. That’s why vendors like Paradox.ai, who have spent years training its systems for recruiting, remain focused on this problem, because its system is getting smarter and smarter about recruiting every day.”
This type of AI software will help us make smarter decisions and use data as information more efficiently. AI will do the analysis and will increase the amount of information have available when decisions are made. The final judgement will be still on the human side.
The Age of Co-pilots
Apparently, the AI used to improve, speed up, and make more efficient the decision making is one application of the technology. Another one are co-pilots.
This is where the addition of Copilot, which relies on the Microsoft Graph for collaboration and behavioural data, likely changes the game. By adding Copilot to the mix, Glint customers should be able to ask questions, right within Glint, that span all of these different data sources (the Microsoft Graph, the other Viva applications, and Glint).
Co-pilots are intended to speed up your daily work through natural language processing and task automation. Would you like to create a chart in Excel based on your data? Ask your co-pilot a question. Other potential elements of application are presentation drafts done through prompting the context or based on a document, workflow automation created by prompting what and how should be done, document synthesis, idea generation, email draft, response to inquiries, etc.
Generative AI can help in time consuming activities like writing job descriptions for job postings, or by supporting onboarding of the candidate thanks to dedicated chatbot which was taught about the work domain, company, and market context in terms to help the newcomer to get on quickly. And by the same time release manager’s or buddy’s time. The latter is especially important in the hybrid workplaces where the managers responsibility for onboarding rose heavily.
In HR we can think about much more applications of co-pilots.
Managers would appreciate help regarding employee well-being. Who is at risk of burnout or overworking? Who haven’t take the annual leave yet? What we have talked about during the last performance touchpoint and what are the proposals for follow up for this week? All of these could be easier thanks to generative AI working alongside with managers and specialists.
I know it often sounds naïve. I have exactly this feeling – above mentioned examples aren’t something that will occur in nearby future, because of the time needed to roll-out, resources shortage, expected downturn, and other important priorities for the companies. But you know what? I have two arguments that the change is closer that you may think.
First and foremost, we invest of we lag behind. Work automation and job transformation is happening whether we want it or not. It won’t wait for us. MIT research showed significant increase in white collar productivity from ChatGPT. According to the researchers, the group of people using ChatGPT was in average 37% time faster at completing tasks than similar group which was doing similar tasks without help of the generative AI. Moreover, the quality of work between both groups were similar. When both groups repeated their work the speed of the ChatGPT group went up even more. ChatGPT made the work faster without sacrificing quality and then made it easier to improve the work even more using the tool.
Co-pilots will make the work faster, will reduce the dull tasks even more compared to the previous wave of automation, and will enable employees to focus on more value adding activities. Doesn’t it sound like an employee value proposition element to you?
Second argument is what Adobe has done with Photoshop. They are segment leaders in the software for photo, video, and picture editing. Their tools like Photoshop are unbeaten market leaders. What Adobe did when generative AI blew? They invested in product development and innovation. As a result, late summer 2023 we got Photoshop 2024 that received generative AI elements such as enabling you to add, extend, or remove content from your photos and images non-destructively, using simple text prompts to achieve realistic results. It works and I’m sure it saved hundreds of hours already for all professionals who have to clean, cut, or generate elements in the photos and images. Amazing, but it’s already here!
Writing good job posting, or job description cannot be more complicated than generating a jaguar drinking from pond in the middle of a carpet in a library.
What’s In It For HR?
HR, unlike many other company functions, doesn’t collect many information from sensors, business processes, or customers. It means we’re not a gold mine for big data. Although, we have another strength which are textual data sets.
Resumes, performance nine-box grids, potential, promotability, succession, learning, skills, job description, verbatim questions and comments in your engagement surveys; all of these are elements of large data sets that are difficult to analyse by human and traditional statistical modelling. But what if we use deep learning algorithms, like generative AI, to get our insights? Again, Josh Bersin:
Consider the two most common parts of HR: a job requisition (job posting) and a job description. Both these artifacts are “thrown together” by hiring managers or HR professionals, often based on what people think a job is like, a set of company standards, and some “technical skills” we know this person will use. As we all know, these artifacts don’t really predict who will succeed, because so much of “success” is based on ambition, learning agility, culture fit, and alignment with purpose.
In other words, this is one of the most complicated and fascinating “mixed data” problems in the world, and making decisions a few percent better can drive billions of dollars of business value.
AI can offer us Talent Intelligence for Recruiting, Mobility, Development, Pay Equity. LLMs based systems help to identify hundreds of parameters, for example skills, in the workforce, enabling companies to diligently source candidates, move people to the new opportunities, identify pay inequities, and implement strategic workforce planning. One of the key benefits of talent intelligence could be avoidance of over hiring that is harmful to people, inefficient, and could lead to damaged employer branding because of following layoffs.
Employee experience is another field of application of AI. Intelligent employee chatbot could bring together documents, support materials, and transactional systems into and easy to use single point of truth.
There are many ways to do this. OpenAI has a feature called “function calling” that lets a develop take any input (“I want to log my vacation.”) and turn it into a simple call to an API like the vacation page in Workday or SAP.
That type of chatbot can be a replacement for many employee portals. It will significantly increase potential of service management tools.
Learning and compliance are perfect use cases for generative AI.
We’ve seen tools that generate training from documents, automatically create quizzes, and take existing content and turn it into a “teaching assistant.” Just yesterday I talked with a client who has a new leadership development program they just built with a vendor. We discussed taking that content and putting it into our Copilot to make it available “on-demand” with a conversational interface for managers. That is not a difficult project once you have the AI platform in place.
Much of the Learning & Development jobs are around content creation. Let’s just come back to the context of creator using Photoshop 2024 generative AI elements. It can be easily extrapolated further to the other multimedia like text, images, video, scenario generators being widely used by the L&D teams. Imagine augmenting it by co-pilot functions to automatically cut, turn into chapters, and promote material, or new things to learn.
Another use of AI in HR is people and career growth. These systems will look at your skills, experience, career path, and show you in an accessible way the options you have to grow. The input will be based on the experience of millions other professionals.
These pathways are all exposed and explained by AI, and these new systems show you precisely what you need to learn, what certifications you must acquire, and even who you can talk to about this path.
These career counsellor systems are going to transform lives of many professionals.
The AI systems should not write performance reviews, but they can help a lot. Currently, to solve underperformance issues with the team, project, or an individual, managers often ask HR consultants or business partners for help. What if, instead of asking HR department, managers could get tailored made consultancy regarding performance design and organizational design?
That’s how it works today: each manager has to guess, figure out, and decide “what to do” about a low performing individual, team, or project. Why not let the AI do some of this for us? We have seen apps, for example, that show you the integrated “view” of performance in a company. This is, in many ways, a data problem.
What if we find, for example, that the project teams that are over a certain size simply don’t get things done? What if we look at the skills composition of a team and see that an important one is missing? Maybe tenure is the problem (it often is, by the way). Maybe diversity is holding teams back.
The final big area of AI application is studying, analysing, and improving employee retention, well-being, and engagement. With well trained, robust AI systems we can simply ask questions like “What are the top factors contributing to turnover in the sales department?” or “What factors are hindering our efforts to increase engagement of the R&D team?” and get answers what is the root cause of the problem.
Today, we do surveys, we use other listening methods, but even with growing people analytics as an organizational practice it is difficult to look into scattered data and provide insights. With AI on board we will be capable to do that. Fast.
A Promise or a Hoax?
Are the algorithms we build biased? Yes, they are. Because we are full of bias.
Pay equity has become a massive issue. Google paid over $118 million to employees who proved that the company practiced gender discrimination in hiring and promotion, and Goldman just paid close to $200 million for a similar lawsuit.
The AI is as smart as the data allows it to be. Therefore, we will remain vulnerable to the human factor for a very long time. The data quality seems to be key.
In a still largely exploratory work, Microsoft researchers showed that when small language models (SLMs) are trained with very specialized and curated datasets, they can rival models which are 50x larger. They also find that these models’ neurons are more interpretable.
Another question is whether the development of AI systems, their quality and broad application won’t be hindered by running out human-generated data.
Assuming current data consumption and production rates will hold, research from Epoch AI predicts that “we will have exhausted the stock of low-quality language data by 2030 to 2050, high-quality language data before 2026, and vision data by 2030 to 2060.”
There is a lot of uncertainty regarding the future of AI. No one really knows how we will be using it and how it eventually change our lives. I like to always remind that YouTube was invented in 2008, so only 15 years ago, and since that time we went through many other digital revolutions. One of them we’re witnessing right now and it is related to the generative AI. These digital revolutions aren’t happening overnight and the progress sometimes is slow so we don’t notice it. But if we look back 6 or 12 months then we see the difference.
Embrace AI and your new reality! Partner with it and watch how your HR processes are becoming more efficient, smarter, and more human-oriented than ever before. AI is a tool. You can use it or not, the choice is yours, but if you want to follow the leaders, then actually you have no choice. Start using it.
"No man will make a great leader who wants to do it all himself, or to get all the credit for doing it." - Andrew Carnegie
Want to be a Sponsor?
Would you like to be a sponsor of my newsletter? Contact me: firstname.lastname@example.org.
Thank you for reading my newsletter! I’ll be delighted to get your feedback about the newsletter and this issue.
Do you know someone who might be interested in this newsletter? Share it with them.
You can also: