InsightsAI and workforce8 min read
AI in HR without the data risk
Where AI earns its place in a people function, where it does not, and the four lines you do not cross with employee data to get there.
By Finesse Blumenthal, Founder and Managing Partner.
Most conversations about AI in HR start in the wrong place. They start with the tool, then look for a problem. The result is a chatbot that answers policy questions nobody was asking, a survey summarizer that replaces the one conversation that mattered, and a quiet decision to paste employee data into something nobody in the company controls.
Start instead with the processes that eat the most time and produce the least trusted numbers. Those are the ones where a machine doing the repetitive part, under rules a person set, gives the people function back its hours and gives the board numbers it can rely on. Then draw the lines that make it safe, and do not cross them.
Where it earns its place
Repetitive, rule-bound, and read by a person
The test is simple. If a process is done the same way every month, depends on data the company already holds, and ends in a document a person reads before acting, it is a candidate. Four that qualify in almost every growing company:
Workforce planning that refreshes itself
The headcount and cost plan rebuilt from the HRIS and finance actuals every month, with spans, ratios and cost per head against plan. The work is arithmetic on data you already hold. A person reads the result; nobody re-keys the inputs.
A critical-role register that stays current
Which roles matter most, who holds them, how long they have been there, whether a successor is named. Kept current from the roster instead of from memory, so the board question has an answer before it is asked.
Onboarding that runs without chasing
Every step of a new hire's first month, owned, dated, and nudged. The manager is reminded; the new hire is not left wondering. Nothing about this needs judgment. All of it needs persistence, which is what machines are for.
Integration tracking from signing to day 100
The decision calendar, the people risk register, the retention watch list, kept in one place and reported weekly without someone building the slide by hand.
Where it does not
Three places to keep people in the loop
Deciding who is promoted, paid or let go
Decisions about individual people are made by named people. A model can prepare the information. It does not make the call, and it is never the reason given.
Reading engagement surveys for sentiment
The value of a survey is the conversation it starts. Summaries that replace the conversation remove the value and keep the cost.
Anything the owner cannot explain
If the person accountable for a process cannot describe what the automation does and why it produced what it produced, it is not ready to run.
The principle
Automate the preparation. Never automate the decision. The model gets the numbers ready; the person accountable makes the call and can say why.
The four lines
What you do not do with employee data
The data risk in AI is not exotic. It is the ordinary risk of handing personal information to a system you do not control, and it is entirely avoidable with four rules that cost nothing to keep.
- 01Employee identities never reach the model. Names and identifiers are replaced on the way in and rejoined only inside your own systems.
- 02The AI runs under your own account, in tools you already hold. Your data is not used for training and does not leave.
- 03A person signs off before anything reaches an employee or the board. Every output has a named owner.
- 04Every run leaves a record of what it did and why, in plain language a diligence team could read.
How to check a vendor, or yourself
Ask which of the four they can show you, in writing, today. A vendor that cannot answer the first one is not a vendor for employee data. A process of your own that cannot answer the third is not ready to run.
The point is the hours
An HR manager in a company of a hundred people spends a large share of the month building the same reports by hand, chasing the same onboarding steps, and answering the same board question from memory. Give those hours back and the same person can do the work that needs judgment: the manager who is struggling, the pay exception that needs a decision, the leader who is about to leave. That is where AI in HR pays. Not in replacing the people function, but in letting it do its job.