IF INTERESTED #188: Deploying GenAI in HR
Disclaimer: Opinions I voice here and examples I give are not representations of my current organisation.
I am updating this issue as my thinking evolves with real use cases. You might find changes in wordings and flow when visiting a second time.
hi,
By now, I am assuming you have experienced good talent-related GenAI use cases. A number of companies have already gained traction with successful products. I will discuss how to deploy GenAI in the broader areas of HR and possible options to ready your Human Resources organization for the upcoming paradigm shift.
This IF INTERESTED is a longer read, I aim for deep dive chapters for topics that deserve more than a short remark. Here is the flow:
TLDR (Too Long Didn’t Read) Version for HR AI Deployment
Relevant updates in GenAI world
The paradigm shift for HR
Three layers of GenAI for HR
Readying your HR for the first layer
Conclusion
TLDR: HR AI Deployment Tips
You will find a useful evolution of AI use cases below. For the impatient readers among us, the “too long didn’t read” version here:
Don’t rush and buy narrow GenAI platforms like hiring, learning, coaching; unless it is to create awareness and in small scope. Broader models on the horizon will easily handle those narrow cases when they are deployed.
Push your current platforms to have a GenAI solution. Your learning platform should be AI powered by now. Use those for building internal credibility.
Be patient another one or two quarters and wait for more capable models, and build competence in HR. For Human Resources, this is a long distance run.
Build an internal capable AI unit in your company, they will be your best friends and invaluable allies.
Start with knowledge base building, conversational chatbot. Evolve from there.
A few relevant updates in GenAI:
(at the risk of this section being outdated in the next 5 days, here’s a few announcements relevant for this article)
The first autonomous AI software engineer, Devin, launched a few weeks back. We now have human software engineers and AI software engineers. Devin can join meetings as a colleagues, code and test projects and ask questions where he/she feels stuck.
Meanwhile Nvidia unveiled a significantly more powerful GPU (Blackwell B200) continuing its exponential increase in compute power, enabling new and powerful LLM systems.
On the LLM front, GPT-5 is said to be already in trials for enterprise use, and one CEO using it emphasises the easy ability to use company specific data and call AI Agents. Both aspects are critical for any enterprise level GenAI deployments. Elon’s Grok launched and Llama 3 is on its way, making open-source LLMs increasingly relevant for enterprises and providing options for enterprises to build their own applications on top of.
For over a year, GPT-4 was the dominant AI model, clearly much smarter than any of the other LLM systems available. That situation has changed in the last month, there are now three GPT-4 class models, all powering their own chatbots: GPT-4 (accessible through ChatGPT Plus or Microsoft’s CoPilot), Anthropic’s Claude 3 Opus, and Google’s Gemini Advanced. (from Ethan Mollick)
The Paradigm Shift for Human Resources:
Rebuilding Human Resources from first principles:
Let's take a moment to emphasize the monumental shift GenAI is poised to bring to people functions. At the risk of stating the obvious, GenAI's profound capability hints at its transformative potential in HR. GenAI at its core is a large scale intelligence that can become a tutor and given the tools (AI Agents) a transaction executor. It provides us with a digital colleague who knows everything and can execute transactions. Those two aspects challenge us to reconceptualize HR from the ground up. To clarify, HR is traditionally tasked with executing organizational strategy through ensuring capability readiness and talent optimization. However, its operating model is often broken down into specialized areas such as hiring, compensation, benefits, and talent management. This compartmentalization, a construct designed to align with human competencies, might be reconsidered in light of GenAI. I don’t believe the future holds such disjointed structure of HR as it is today.
The current perspective on human resources functions, aiming only to enhance existing structures with GenAI, seems constrained. A deeper exploration of AI’s capabilities could lead to the design of a fundamentally new business enablement function. This approach, based on “first principles”, will unlock the full potential of GenAI, even blurring the need to have departments in organisations at large.
This is the reason why I avoid going deeper into the specific areas like AI in recruitment, or how to use AI in learning. Foundational models together with Autonomous Agents will cater for all those needs once deployed fully.
The Three Layers of GenAI in HR:
I have identified three distinct tiers of application of GenAI or LLM Systems. Initially, as the first tier, I cover the basics: conversational chatbots and AI agents capable of performing tasks independently.
The second tier is the LLM System evolving into a role akin to personal coaches and analysts, offering tailored advice and insights as a trusted co-worker would.
The last tier of this evolution is the company's digital avatar, a digital entity. The entity becomes a (digitally) living and breathing avatar of the company. That entity eventually shapes as the single governance body.
The notion of a digital entity steering enterprise transformation might seem far fetched, it's a horizon that I believe will materialize within the next three to five years. Meanwhile, the foundational and intermediary tiers are rapidly advancing, powered by sophisticated Large Language Models (LLMs) and AI Agents. In the following sections, we'll delve deeper into these tiers.
1. Assistant:
The Assistant is the essence of knowledge-based support, whether through an engaging conversational chatbot, a dynamic AI-powered search tool navigating the vast expanse of the company's knowledge base, or an efficient digital worker handling the tedious, system-based tasks that clutter our workdays.
Conversational Chatbots:
This is what almost all modern companies are working on, primarily utilised for the customer support or enhancing the effectiveness of internal sales force. Yet, when we pivot to HR applications, the term "conversational" demands authenticity. Today's chatbots often stumble over complex queries and fail maintaining fluid dialogues.
Key considerations:
An LLM model capable enough to have a conversation (GPT4 Turbo and similar)
Meticulously fine-tuned data sets with no inaccuracies
A memory component: Ability to remember what the previous conversations were, and who the person is (this is something very difficult to do with LLMs, not to be confused with the apparent short memory that comes from the prompt context window or session)
AI Powered Search:
For those familiar with Perplexity or Google Gemini, the concept won’t be foreign. Imagine a search engine tailored for your HR queries, capable not only of summarising responses but also guiding you directly to the detailed documents, policies, or records you need—like pinpointing the specifics of your company’s sabbatical policy amidst a sea of data. This is what we refer to as a Retrieval Augmented Generation chatbot, or RAGbot.
Below is a screenshot from Perplexity.
Task Completion: Autonomous AI Agents
An AI agent, in a simple way, is a program designed to autonomously take actions to achieve specific goals.
AI agents are at the forefront of artificial intelligence, revolutionizing how we interact with technology. For a tangible glimpse of this innovation, the “MultiOn” demo on YouTube showcases an AI Agent autonomously purchasing groceries with a simple prompt like “can you buy me bananas and water”.
The basic Human Resources version will be “can you please open a software engineer position and then post the job ad on LinkedIn”. AI Agents very soon will be able to do those tasks, significantly increasing Employee Experience and cutting down bureaucratic steps, creating efficiencies.
If you merge this with the chatbot use case above: “Am I eligible for a company car allowance instead of company car? If yes, can you register me for allowance?”
There's a compelling case for the aggressive deployment of AI Agents within Human Resources. This arena with disjointed procedures for initiating, approving, and completing transactional operations, will gain immensely from the efficiency and cohesion of AI Agents.
That is why the initial GPT-5 news about the focus on AI Agents represent a tremendous opportunity in the second half of 2024.
Conclusion
The three main use cases above give a frame to the first layer I call “Assistant”, as you now know, we are talking about a very capable assistant who can answer every question in detail and execute multi step tasks on its own.
The Assistant reveals a dynamic entity capable of informed responses and managing complex, multi-layered tasks independently. The conversational chatbot and search features are stepping stones towards centralizing the company’s entire knowledge database. It is easier to start with Human Resources topics and policies, and then scale it to all policies and practices in the enterprise. Or starting with product information and customer support allows for iterative improvements in LLMs, setting the stage for a company-wide knowledge base, with HR positioned to benefit significantly in the subsequent phase. Both roads can lead to success.
2. Trusted Co-Worker:
This layer is about AI taking the roles of both analyst and coach, acting as a mentor, a trusted co-worker, and a business-savvy expert all in one. This is co-pilot on steroids.
It is “THE” trusted co-worker who knows company’s products, finances, and talent insights, or is an expert on human and organizational behavior, guiding you through difficult situations. This is a textbook AI augmented employee journey, making the employees better at everything they do.
The Analyst Mode transforms the digital colleague into a beacon for decision-makers, guiding them through a fact-based decision-making process. It comprehensively evaluates variables, discerns their correlations, and conducts scenario analyses. This epitome of "augmented decision-making" empowers employees to converse directly with data, soliciting tailored recommendations from the ANALYST, leveraging advancements in GenAI technology.
The Coach Mode emphasizes behavioral insights. With access to an employee’s interactions—meetings, emails, chats—(always with consent), the COACH gains an unparalleled understanding of the individual. This depth of insight enables the COACH to guide employees towards behaviors and strategies that suit the context best. Again transforming the individual to their best version.
The COACH and ANALYST represent two sides of the same trusted co-worker. The specific needs or inquiries of an employee dictate the adoption of a particular “personality” or a blend of the two, showing the adaptability of the AI model.
Even though there are examples of AI Coaching in the market, they have nowhere near the capabilities mentioned above. However, they represent the first generation.
3. Digital Entity: The Company’s Digital Avatar
The final layer, which often sparks debate among my friends, envisions the GenAI model evolving into the Company’s Digital Entity. This model has comprehensive knowledge about the company, including its stakeholders, market dynamics, and competitors. Imagine an omnipresent, flawless CEO, available at any moment to support, approve and guide—this is the vision for the ultimate digital entity.
Knowing the strategy and frames of the company, this model can approve anything that the employers can put through. It can keep track of the business transformation efforts, monitoring every decision. It ensures obsolete products are phased out as planned, guards against misallocation of resources, and can even correct or enhance customer proposals in real-time. Beyond operational decisions, it offers guidance on behavior, grasps the nuanced cultures within the organization, and actively promotes collaboration and alignment with the company’s values among employees and units.
At this point, traditional human management begins to appear redundant. And I shall not write further in this section : ).
Deploying GenAI in HR:
The principles here to deploy GenAI are not necessarily limited to HR. I will focus on the first layer, that is The Assistant, and leave the deployment of the other layers for future IF INTERESTED issues.
Deployment of AI needs good AI knowledge in the company, a good partner like Google / MS and a significant data preparation from HR.
Here are some pragmatic headings that will help you during the preparation and deployment:
Side Benefit: Radical Simplification before AI Takes Over
For LLM systems to effectively assist employees, they must learn the nuances of a company's internal policies and practices. That requires fine-tuned data with the company specific policies and practices, and a good information retriever (RAG). Enterprise policies, may they be Human Resources Policies, Procurement Policies or Customer Entertainment Policies, they are often reside across varied and disjointed platforms, from PowerPoint presentations to SharePoint sites, compounded by the prevalence of outdated content.
The need for the LLM system to go through rigorous fine-tuned data to minimise hallucinations, presents a prime chance for those enterprise functions (HR, Procurement etc) to streamline and condense those policies and simplify them before integration. This step ensures that LLM system ( or the vector databases, RAG’s etc) operates on clear, concise, and up-to-date information.
While there are AI tools that can streamline the processes, a better approach involves engaging a UX designer from outside HR alongside a process optimisation expert to radically simplify these policies and processes. While they are at it, Human Resources practitioners can embrace a holistic view of the employee experience and not look at the policies from functional boundaries like “compensation and benefits policies”.
Seizing this unique moment can eliminate redundant procedures, streamline and add a human touch to policies, laying the groundwork for an improved employee experience that resonates across the organisation.
Key to Success: The Right Team and the Right Technology
Selecting an adept internal AI team paired with a proficient LLM system vendor is non-negotiable. The gap between satisfactory and exceptional AI guidance is vast, impacting critical aspects such as architecture and APIs, which are far from mere details. These elements dictate whether your conversational AI feels like a human or falls flat as a clunky bot. I can’t overstate the importance of the synergy between expert human insight and the optimal LLM system.
If you are not sure about your internal AI competence, wait for 8-12 months so that the large-scale AI for HR solutions become productised, and you buy it off the shelf.
Don’t work with HR companies in this area, work with AI companies, you dont need HR expertise, you need AI.
Broad Beginnings: Scope Your AI Assistant Wisely
Contrary to the minimalist approach of starting small, aim for a wide-reaching scope from the get-go. This strategy ensures your AI assistant immediately offers comprehensive support across various HR policies, instead of being limited in functionality and frustrate users. This broad foundation encourages stronger adoption and a more impactful employee experience. Nevertheless, such a scope requires meticulous planning and an understanding of your organisation's most pressing needs to execute successfully.
Scope:
Consider all HR policies and topics as the scope and to be simplified and fed to LLM system as fine-tuned training data.
Focusing on the inquiries coming to your HR support teams. Aim to cover 80% of the questions for tier 1 and 2 supports with the AI conversational bot.
The areas you are getting the most questions and tickets are the most unclear, simplify and clean those policies first. I assure you these areas are not “succession planning” or “leadership development”; they are more about payroll, taking vacation or sick leave.
Start experimenting with AI Agents later in 2024 if you have a capable AI internal unit, if not wait for 2025 or later.
Foundational Steps: Ready Your Infrastructure for AI
The groundwork for a successful AI implementation in HR lies in data consolidation. Create a unified data lake to centralize talent information, ensuring all data, from promotions to performance evaluations, is easily accessible. Simplify your tech stack by reducing the number of HR applications, aiming for a cohesive ecosystem.
Consolidate your databases, build a common data lake for your enterprise where all talent data flows to one secure place. Build your data architecture from the start to accommodate the incoming data.
Consolidate your HR platforms. Don’t rush after fragmented AI solutions, and buy new products from the market like Skills AI or internal Marketplace. The last thing you want is to move to AI era with 8 HR applications integrated to an HCM (Human Capital Management) like Workday. Consolidate everything under one or two places.
Design your processes and your data architecture to feed every data to your data lake. Example promotion process: Every promotion, the selection process, interview notes, assessment reports, salary increase amount should flow to your data lake. Nothing stays in emails or in people’s minds.
Record all your HR operations inquiries, use your ticketing systems wisely and record the solutions as well.
The above points are easier said than done, I am aware. Get help, work with your IT and invest in your HR tech and HR analytics teams. They will help you.
Start with augmentation (Trusted Co-Worker) in 2025:
There are already few companies that can provide a simple version of leadership coaching. If you want to build (or already have) an experimenting and early adoption culture you can start to work with those companies at small scopes. However, first impressions are important, and the data you feed into those systems are crucial.
Be mindful of EU limitations
The recent legislation about AI will limit many small players in the field. Their access to your data will be only possible with significant security certificates, especially with the LLM systems. Big players will likely get those umbrella certificates early, and handle most of the bureaucracy for you. (as Bill Gurley says, regulation favours the incumbent).
Get your security people and IT people early in the game, and don’t underestimate the paperwork it will take to get a small player into your systems.
Don’t worry about Change Management
Aim to build an amazing product with the Assistant, underpromise and overdeliver: but deliver a product that doesn’t need significant change management. Remember: nobody held workshops for you to use AirPods, they looked stupid when they first came out, but they were so functional people just used them.
Focus on the product.
Summary and Conclusion:
Enlightenment is: absolute cooperation with the inevitable.
Anthony De Mello
The exponential evolution of GenAI’s is difficult to keep pace with. I am firmly optimistic about AI. I envision a future where AI catalyses significant improvements in the human condition. In the immediate term, it’s critical for organisations to pivot their perspective on GenAI—from viewing it merely as a tool for efficiency headcount reduction to recognising its potential to elevate the quality of work and create new, innovative business opportunities. And Human Resources leaders should be the beacon of this message.
To my friends and colleagues in HR operations and subject matter expertise, consider this a friendly yet urgent nudge: the landscape is shifting rapidly with LLM Systems. Many traditional roles in our field are exposed to transformation, if not outright automation. This is not a signal to despair but a call to seize the moment—invest in upskilling and reskilling.
If interested
burak