You signed up for four hours on how HR leads AI adoption. This primer gets everyone into the room with the same vocabulary, so Tracie and I can spend those four hours on your work instead of on definitions. If you can, read it before you do your Leverage Audit. It will make the audit faster.
The vocabulary matters more than usual here, because the market is muddying it. Gartner estimates that only about 130 of the thousands of vendors selling “agentic AI” actually deliver it. The rest are “agent washing,” Gartner’s term for putting a new label on chatbots and robotic process automation that already existed. Gartner analyst Anushree Verma put it plainly: many use cases positioned as agentic today don’t require agentic implementations. You will sit in vendor demos where all three of the technologies below get called “AI.” After this, you’ll know which one you’re actually looking at.
If you’ve used ChatGPT, Claude, Gemini, or Copilot, you’ve used a large language model. A model is trained on an enormous amount of text until it gets very good at one thing: predicting what should come next, given everything that came before. Almost everything else it does is built on that.
That’s why it can write a job description in ten seconds, and it’s also why it can state something false with total confidence. The National Institute of Standards and Technology calls that failure confabulation in its Generative AI Profile. The model produces the most plausible answer, and plausible and correct overlap most of the time. Most of the time is a fine standard for a first draft. It is a bad standard for anything that reaches an employee or a candidate unchecked.
Large language models now power two very different kinds of tool. Add the automation most HR teams already run, and you have three things sold under the same label. The simplest way I’ve found to keep them straight is to ask what you hand each one. You hand automation a rule. You hand generative AI a prompt. You hand agentic AI a goal.
Automation is software executing steps a person wrote down in advance. The logic inside your HRIS that sends a background check request the moment an offer is accepted is automation. So is robotic process automation, or RPA, the bot that copies new-hire records from your applicant tracking system into payroll every night. It produces the same output from the same input every time, and no AI is required.
That predictability is the whole point. Automation is the right tool when the work is high-volume and the correct answer never changes: running payroll on the 15th and the 30th, provisioning a new hire’s laptop and accounts, sending the I-9 reminder on day two.
It breaks the moment reality doesn’t match the rule, whether that’s an offer with a non-standard start date or a remote hire in a state the workflow was never built for. Automation doesn’t improvise and it doesn’t tell you it’s confused. It stops, or worse, it keeps going with the wrong answer.
Generative AI is the chat window you’ve been using. You ask, it drafts, it stops. It takes no action in your systems on its own, so a person decides what happens to every output, and that person owns the result.
It’s best at first drafts and synthesis, such as turning 2,000 open-ended engagement survey comments into themes, aggregating a manager’s performance review notes into a first-draft summary, or drafting an answer to a benefits enrollment question for a specialist to check. In every case the human judgment stays in the loop.
It also has a ceiling. In McKinsey’s 2026 State of AI survey, 80% of respondents say AI has improved their individual productivity, yet only 37% attribute any impact on EBIT to it. A team full of people working faster in a chat window does not add up to a company getting results from AI. Something has to change about how the work itself is designed.
Agentic AI is given an outcome, access to tools, and permission to act. It plans the steps, takes actions in other systems, checks its own results, and decides what to do next. Hand an agent “schedule all 40 final-round interviews for next week” and it reads the hiring panel’s calendars, emails candidates, handles the reschedules, and updates the applicant tracking system without anyone touching each step.
This is where AI stops being a tool someone uses and starts being labor someone manages. It’s strongest on multi-step work that crosses systems and needs light judgment along the way: interview coordination, resolving routine employee service tickets end to end, chasing down missing onboarding documents. Gartner expects at least 15% of day-to-day work decisions to be made autonomously through agentic AI by 2028, up from zero in 2024.
It also fails differently. An error in step two quietly feeds step three, and nobody sees it until a candidate shows up on the wrong day. That is the risk Gartner had in mind when it forecast that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.
|
Traditional automation |
Generative AI |
Agentic AI |
|
|---|---|---|---|
|
You hand it |
A rule |
A prompt |
A goal |
|
It gives back |
The same action, every time |
A draft for a person to judge |
A finished outcome, with actions taken in your systems |
|
Best for |
High-volume work where the right answer never changes |
First drafts and synthesis |
Multi-step work across systems with light judgment between steps |
|
HR example |
New-hire provisioning |
Aggregating performance review comments into themes |
Interview scheduling, end to end |
|
How it fails |
Stops, or keeps going with the wrong answer, when an input breaks the rule |
States something false with full confidence |
An early error compounds through later steps unseen |
|
Who owns the result |
Whoever wrote the rule |
Whoever uses the output |
Nobody, until you decide |
Your pre-work asks you to pick three to five workflows and score them. As you do, write one word next to each: rule, prompt, or goal. New-hire provisioning is probably a rule. Answering policy questions is probably a prompt, with a person checking the answer. Interview scheduling may be a goal.
Two cautions. First, if a workflow is a rule, it may not need AI at all, and that’s a good outcome. An agent is the most expensive way to solve a problem a rule could handle. Second, anything that touches hiring decisions, such as resume screening, carries the most risk whichever tool you pick. Don’t overthink the labels. We’ll pressure-test your top pick together in Session 1.
Picking the right technology is the easy part. McKinsey’s same survey found that nearly three-quarters of its AI high performers have fundamentally redesigned workflows because of AI. Among everyone else, it’s one-quarter. Ethan Mollick calls working well alongside AI co-intelligence, and the companies getting there are redesigning the work, not just adding tools to it.
No tool decides which work goes to people and which goes to AI, which work stays human by design, or who answers when an agent gets something wrong. Those are people decisions. That is where the course starts.
If you want to see how mid-market leaders are handling those decisions today, our Co-Intelligent Workforce Report: 2026 Baseline, a survey of 484 Director-and-above leaders across the US and Canada, publishes the week of October 19.
You don’t need to memorize these. Keep this table handy, because we’ll use these terms in the room without always stopping to define them.
|
Term |
What it means |
|---|---|
|
The basics |
|
|
Large language model (LLM) |
AI trained on massive amounts of text to understand and generate human-like language. It is the engine inside tools like ChatGPT, Claude, Gemini and Copilot. |
|
Prompt |
The instructions or questions you give an AI tool. Specific prompts that include real source material, like the actual policy document, get far better results than vague ones. |
|
Generative AI |
AI that creates new content, such as text, images or code, on request. It drafts and then stops, and it takes no action in your systems on its own. |
|
AI assistant (or copilot) |
Generative AI built into a tool you already use, such as email or your HRIS, that suggests and drafts while you stay in control of every step. Many products sold as agents today are really assistants. |
|
Agentic AI (AI agent) |
AI that is given a goal and access to tools, then plans and takes actions across multiple steps, making decisions within limits you set, without a person directing each step. |
|
Risks to understand |
|
|
Hallucination |
When an AI states something false with confidence, such as quoting a policy clause that doesn’t exist. NIST’s formal term for it is confabulation. |
|
Training data |
The information used to teach an AI model. It shapes what the model knows, which is why bias hidden in historical hiring or pay data tends to show up in the output. |
|
Zero data retention |
A vendor commitment that your data is processed and then deleted, and is never stored or used to train the vendor’s models. Get it in writing before any employee data goes in. |
|
Putting AI to work |
|
|
Grounding (or RAG) |
Connecting an AI tool to your own documents so it answers from your actual policies rather than general internet knowledge. RAG, short for retrieval-augmented generation, is the most common method. It reduces hallucination but does not eliminate it. |
|
MCP (Model Context Protocol) |
An open standard that lets AI tools connect securely to other systems, such as an HRIS or ATS, to read or write data. We cover it in Session 2. |
|
Harness |
The software, instructions, connections and human checkpoints that surround a model and turn it from a chatbot into something you can rely on inside a real workflow. You will build one in the course. |
|
Human in the loop |
A named person who reviews or approves AI output before it reaches an employee, a candidate or a system of record. |
Bring your top workflow and your questions. There are no dumb ones. See you in Session 1.
Additional AI Fundamentals and AI Literacy Learning
The AI Fluency for HR Masterclass is designed to equip HR leaders with the knowledge and skills to lead AI transformation initiatives at scale. That is what we mean by AI fluency. It is different from AI literacy, which is understanding how AI works and the science and technology that power it.
The course assumes a working level of AI literacy, roughly what you would have after using ChatGPT or a similar tool for a few weeks and reading this primer. If you want a stronger foundation before Session 1, each major AI company offers an introductory course. If you only have an hour, start with OpenAI's.
Free Trainings Offered by Frontier LLM Companies
Claude Academy by Anthropic: Anthropic's free learning platform offers courses, tutorials, and real-world use cases on how AI works and how to use it safely and effectively. Recommended courses:
OpenAI Academy: OpenAI's free learning hub offers courses, guides, and live events on using ChatGPT at work. Recommended courses:
Google Skills: Google's learning platform brings its AI and technology training together in one place, with courses ranging from beginner to advanced. Recommended courses:
Let’s Orchestrate!
Sources
Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” press release, June 25, 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1), July 2024. https://doi.org/10.6028/NIST.AI.600-1
McKinsey & Company, “The state of AI in 2026: On the road to ROI,” August 25, 2026. n=1,719 respondents in 97 countries. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
Ethan Mollick, Co-Intelligence: Living and Working with AI (Portfolio, 2024). https://www.penguinrandomhouse.com/books/741805/co-intelligence-by-ethan-mollick/