Automating HR: Where AI Adds Value and Where It Creates Liability
Two years ago, a client asked us to build an AI resume screening system. We declined and explained why. Six months later, a different client asked us to build a system that automated onboarding document collection and reminded managers when probation review dates were approaching. We built it in three weeks and it has been running without incident since. The difference between those two requests is the difference between AI that takes decisions about people and AI that removes friction from administrative processes. The line matters more in HR than almost anywhere else, because crossing it creates legal exposure that no efficiency gain justifies.
Where AI in HR clearly works: administrative automation
The HR processes that are genuinely good candidates for automation are the ones that are currently manual, time-consuming, and do not involve subjective judgment about people. Three categories where we have built systems that work:
Onboarding workflows. Document collection, system access requests, equipment provisioning tracking, training completion follow-ups. A new hire's first week involves 30–50 administrative touchpoints across IT, facilities, payroll, and HR. Automating the orchestration of those touchpoints — sending the right form at the right time, tracking completion, escalating to the hiring manager when something is overdue — reduces the administrative load on HR by hours per hire without making any judgment calls about the employee.
Leave management and scheduling support. Calculating leave balances, checking coverage requirements before approving leave requests, notifying managers when team leave overlaps with project deadlines. These are rule-based processes with defined criteria. Automating them reduces processing time and removes human error in balance calculations.
HR data maintenance and compliance tracking. Contract expiry reminders, work permit renewal tracking, mandatory training completion monitoring, compensation review cycle management. All of these are calendar-driven and criteria-based. Automating them means nothing falls through the cracks when the HR manager is on leave.
Where AI in HR is genuinely useful but requires careful design
Some HR applications are more complex — they involve analysis or pattern recognition that is legitimately useful, but the risk of bias or misuse is high enough that the design must be careful. Two categories:
Employee sentiment analysis from survey data. Analyzing anonymized employee survey responses to identify themes, track sentiment over time, and flag departments with significant changes in engagement. This is legitimate analytics on aggregated data. The risks: de-anonymization if the analysis is too granular, acting on AI-identified patterns without human verification, and the organizational politics of having an AI flag specific team leaders for low sentiment scores.
Our approach: sentiment analysis only on data where the response pool is large enough (minimum 10 respondents per cohort to prevent de-anonymization), outputs reviewed by HR before being shared with management, trend analysis shared in aggregate rather than as flags against specific people or teams.
Job description analysis and pay equity auditing. Using AI to analyze job descriptions for gendered language, or to audit pay data for unexplained compensation gaps across demographics. Genuinely useful. Requires HR and legal review of the methodology before acting on the outputs. The AI is doing statistical analysis — a human with appropriate authority needs to make the decisions that follow from it.
The useful category of HR AI is analysis that surfaces patterns for human review, not decisions that replace human judgment. A pay equity audit that shows an unexplained 12% gap for a specific demographic group is valuable. A system that automatically adjusts salaries based on that analysis without a human making the decision is not — it is a discrimination lawsuit waiting to happen.
Where we will not build: hiring and performance decisions
We decline work in two HR AI categories. Not because the technology does not work, but because the risk of harm is high and the regulatory and legal environment is increasingly hostile to AI-driven decisions in these areas.
AI resume screening and candidate ranking. Every trained AI system for resume screening learns patterns from historical hiring decisions. Historical hiring decisions contain human biases — conscious and unconscious. An AI trained on ten years of hiring data for a company that has historically underrepresented certain demographics will learn to underrepresent those demographics. The system replicates and scales the bias. It also creates legal exposure under employment discrimination law in most jurisdictions.
The usual counterargument is that humans are biased too. This is true. But a biased human is more defensible in court than a biased algorithm, the bias is slower to scale, and the bias can be corrected through training and accountability. Algorithmic bias is faster, scales perfectly, and is harder to attribute and correct.
AI performance monitoring and termination risk scoring. Systems that continuously monitor employee productivity, flag underperformers, or score termination risk based on behavioral patterns. These create legal risk under labor law, create toxic work environments when employees know they are being scored, and generate outputs that are used in disciplinary decisions without the employee being able to understand or contest the basis for the score.
Our take
The design principle that guides our HR work
The question we ask before any HR automation project: does this system make decisions about people, or does it remove friction from processes that people then manage?
Automating the collection of onboarding documents removes friction. Automating the escalation when a probation review is late removes friction. Automating a leave balance calculation removes friction. None of these make decisions about the employee as a person.
Screening resumes makes a decision about the candidate. Scoring performance makes a decision about the employee. These cross the line. The practical test: if an employee asked "on what basis was this decision made about me?" and the honest answer is "an AI system decided," we should not have built it. If the honest answer is "the HR manager reviewed and decided, and an automated system made that review more efficient," we are on the right side of the line.