The biggest mistake in HR tech right now is calling everything "AI." A workflow that auto-sends an email on Day 1 of onboarding is automation. A system that predicts which new hires are at risk of early attrition is AI. Both are useful — but they require different investment, different governance, and different change management. Conflating them leads to overspending on simple problems and under-governing complex ones.

Rules-based automation is deterministic logic that a human writes explicitly. There’s no learning, no prediction, no ambiguity. If X happens, do Y. It’s the backbone of every HRIS, every payroll system, and every workflow engine. And it handles an enormous percentage of HR operations perfectly well.

Benefits eligibility: IF employee tenure > 90 days AND status = full-time THEN enable benefits enrollment

PTO enforcement: IF PTO balance < 0 THEN flag for manager review AND block further requests

Compliance alerts: IF I-9 not completed within 3 days of start date THEN escalate to HR manager

Payroll rules: IF state = California AND hours > 8 in a day THEN calculate overtime at 1.5x

RPA is software that automates the manual, repetitive actions a human takes across computer systems. It doesn’t think. It follows a recorded script: click here, copy this, paste there, click submit. Think of it as a very fast, very obedient employee who can work 24/7 — but who will do exactly what you told it, even if the context has changed.

Cross-system data entry: Copying new hire info from ATS to HRIS to payroll to benefits

Report generation: Logging into 4 systems, pulling data, compiling a weekly headcount report

Form processing: Extracting data from structured documents (W-4s, direct deposit forms) into systems

Audit prep: Pulling records from multiple systems and organizing them for review

UI changes: When a vendor updates their interface, the RPA bot breaks because the button it was clicking moved. This is the #1 maintenance headache.

Unstructured data: RPA can’t read a resume and “understand” it. It can only copy text from known locations in known formats.

Exception handling: When something unexpected happens (a field is blank, a format is wrong), RPA stops or does the wrong thing unless you’ve programmed every exception path.

Scale: Each new process requires its own bot. Maintaining dozens of bots becomes its own operational burden.

Machine learning excels when you have large datasets with complex patterns that a human couldn’t manually define. It’s not following rules someone wrote — it’s discovering patterns from historical data and making predictions on new situations.

Attrition prediction: Analyzing dozens of factors (role, tenure, manager changes, comp history, engagement scores) to predict flight risk. No human could write rules for all these interactions.

Resume matching: Scoring candidate fit based on learned patterns from successful hires (with proper bias auditing).

Anomaly detection: Flagging unusual patterns in time-tracking, expense reports, or payroll that might indicate errors or fraud.

Workforce planning: Predicting hiring needs based on historical patterns, growth trajectories, and market data.

Small datasets: If you have 50 employees, ML won’t find meaningful patterns. You need hundreds or thousands of examples.

Simple rules work: If you can write the logic in 5 IF/THEN statements, don’t train a model for it.

No historical data: ML needs examples to learn from. If you’re starting a new process with no history, rules are your starting point.

Low tolerance for errors: ML makes probabilistic predictions — it will sometimes be wrong. For processes that require 100% accuracy (payroll calculations), use deterministic rules.

Generative AI (LLMs like ChatGPT, Claude, Gemini) is a text-in, text-out engine. It doesn’t analyze your data or make predictions — it processes language. It’s exceptionally good at tasks that involve creating, transforming, or summarizing text. This makes it a natural fit for the enormous amount of writing that HR operations requires.

Job descriptions: Draft role-specific JDs in seconds, then edit for accuracy and compliance

Policy drafting: Generate first drafts of policy updates, handbook sections, or FAQ answers from bullet-point input

Survey analysis: Summarize thousands of open-ended survey responses into themes with supporting quotes

FAQ bots: Power employee-facing chatbots that answer benefits, PTO, and policy questions from your knowledge base

Translation: Localize employee communications across languages at scale

Never use as sole source: GenAI hallucinates. Every output touching compliance, legal, or benefits must have human review.

Data privacy: Don’t paste employee PII into public LLM tools. Use enterprise-grade solutions with data processing agreements.

Consistency: GenAI isn’t deterministic. The same prompt can produce different outputs. For standardized communications, use templates with GenAI filling in specific sections, not generating from scratch.

Brand voice: Train your prompts (not the model) to match your organization’s tone and terminology.

AI agents represent a leap from “tool I ask questions” to “system that takes actions on my behalf.” An agent can break a complex goal into steps, use tools and APIs to execute each step, handle exceptions along the way, and report back. It’s the difference between asking a consultant for a recommendation and hiring an employee to do the work.

Onboarding orchestration: Agent receives “new hire starting Monday” and triggers account provisioning, sends welcome emails, schedules orientation, orders equipment, follows up on incomplete paperwork — all without manual intervention.

Benefits enrollment: Agent walks employees through options conversationally, answers questions from the plan documents, submits elections, and confirms coverage.

Offboarding: Agent coordinates across IT (revoke access), payroll (final check), facilities (badge return), and manager (knowledge transfer) to ensure nothing falls through cracks.

Agents sound incredible — and they will be transformative. But they raise operational questions that most organizations aren’t ready for:

Accountability: When an agent makes a wrong decision (sends the wrong benefits enrollment, revokes access too early), who’s responsible?

Audit trail: Can you reconstruct every decision the agent made and why?

Approval chains: Which actions should the agent execute autonomously vs. queue for human approval?

Error recovery: When an agent fails mid-workflow, how do you roll back the steps it already completed?

The mistake: Using AI where simple rules would do the job. A vendor sells you an “AI-powered benefits eligibility engine” when your HRIS already has a rules engine that handles this perfectly.

Real example: An HR team deployed an ML model to predict which employees needed to complete annual compliance training. The answer? All of them. A simple rule (“all active employees, every January”) would have worked at zero additional cost.

The fix: Before evaluating AI solutions, document the current process. If you can describe the decision logic in a flowchart with fewer than 10 decision points, rules will almost certainly work.

The mistake: Manual processes that should have been automated years ago. Someone on your team is still manually copying data between systems, generating reports by hand, or sending reminders one by one.

Real example: An HR coordinator spending 8 hours per week manually entering new hire data from the ATS into the HRIS, payroll, and benefits systems. An RPA bot or native integration could handle this in minutes.

The fix: Audit your team’s time. Any task where someone says “I do this the same way every time” is a candidate for rules or RPA.

The mistake: Using the right technology for the wrong problem. The most common version: using GenAI (an LLM) for tasks that need deterministic accuracy.

Real example: Using ChatGPT to calculate FMLA eligibility dates. LLMs are bad at math and date calculations. This needs a rules engine with the specific federal and state FMLA rules coded in.

Another example: Using RPA to “automate” resume screening by copying resumes into a spreadsheet. The bottleneck isn’t data movement — it’s evaluation. This needs ML (or better processes, or both).