Why Isn't AI Helping With HR Compliance Work?

Should AI be making leave or compliance determinations at all?

AI should narrow and prepare compliance determinations, and a named person should make them, because the risk of a fluent but wrong answer is highest where judgment and accountability matter most.

There is good evidence for that caution. In a preregistered experiment with 758 knowledge workers, run with Boston Consulting Group on realistic consulting tasks and published as a Harvard Business School working paper in September 2023, people using GPT-4 on tasks inside AI's capabilities completed more of them, faster and at higher quality. On a complex task chosen to sit outside those capabilities, they were 19% less likely to reach a correct solution than colleagues working without AI. The researchers called that uneven boundary a "jagged technological frontier." Compliance decisions, full of edge cases and institutional precedent, sit on the harder side of it.

Weak AI output in this context rarely looks weak. It reads fluently, cites a policy section and sounds confident. That is what makes it dangerous where someone has to own the outcome if a determination is challenged.

So the useful question for leaders is a set of questions, asked before AI goes near a case: what problem the work is solving, what kind of judgment it requires, where a person makes the biggest difference, where trust matters most, and who owns the outcome when something goes wrong.

Where the edge sits

The edge of AI's reliability in HR compliance moves. It depends on how well the policy is documented, how closely the case resembles earlier ones, and how much discretion the role has. Finding that edge in your own organization is groundwork, and it has to happen before AI is trusted with anything.

Why do HR teams still handle case-by-case decisions manually after rolling out AI?

Managers were never given the guardrails, training or permission they needed to trust AI's output, so they default to the slower method they know works. The license got distributed and the decision authority stayed where it was.

Managers need a credible story about why the organization adopted AI for this work and what a good outcome looks like. They need clear guardrails on where AI use is encouraged or restricted, and explicit permission to manage the transition rather than enforce it from above. Without those, a manager facing a borderline leave case will fall back on the manual process every time, because that is the one where they know who is accountable if it goes wrong.

Training on the software is not enough either. People need to understand how AI fits into the compliance workflow: where to use it, how to judge its output, which guardrails apply, and when a second person has to look.

A specialist who keeps doing the manual check is telling you something true about where trust has not been earned yet. Reading that signal is more useful than mandating faster adoption around it.

What does it take to redesign an HR process so AI helps with it?

Start by mapping the existing workflow honestly: which steps are repetitive, which require judgment, where the bottlenecks sit, and which parts nobody has examined in years. AI gets placed into that map afterwards.

The book makes the point directly: it is very difficult to redesign work that nobody has properly examined. If the organization has not worked out which parts of the compliance process are broken, where workarounds have crept in, and which tasks are repetitive and which need a person to weigh competing facts, AI creates more confusion than it resolves. Clean up the workflow first.

From there, the redesign has a clear shape. AI drafts the first-pass summary of a case and flags the applicable policies. A specialist reviews the summary against the parts of the case that need judgment. The final determination carries a named owner whether or not AI touched it. The specialist's job changes, with less repetitive drafting and more of the decision-making their expertise is for.

Leaders also need evidence that this is working, in terms people recognize. An early win here is specific: case turnaround drops, the backlog shrinks, or specialists spend less time on repetitive steps and more on the judgment calls. "We launched it" tells people nothing.

How do you know if the redesign is working or just adding another layer of process?

It is working when specialists choose to route more cases through the new workflow, and when the time saved shows up in turnaround and backlog figures. Usage numbers alone will not tell you that.

Watch the distance between what leadership reports and what specialists do when nobody is checking. If people complete the mandatory steps in the new process and then redo the judgment call by hand because they do not trust the AI-assisted version, the redesign has added a layer and changed nothing.

Another honest signal is whether managers can say, without hesitating, where AI is supposed to help in this workflow and where a person has to own the call. If that answer is fuzzy, the guardrails were never clear enough to build trust, and more training will not fix it.

What's the real cost of leaving this unresolved?

The license is the smaller cost. The larger one is the trust lost between HR leadership and the specialists who were told AI would help them and found it added a step.

That loss compounds. The next technology initiative lands on a team that has learned to treat new systems as extra work. Rebuilding that trust takes longer than building it would have the first time, and it starts the same way: mapping the workflow, naming where judgment has to stay with a person, and giving managers the backing to make the transition real. The Elephant in the Algorithm walks through this in more depth, including why the jagged frontier makes compliance work such a hard case for blanket automation.

If you already know this is where your adoption stalled and want a structured way to rebuild the process around the people accountable for the outcome, see how the AI Profit Partnership works. It ends in a plan for the path from case intake to determination that gives AI a bounded role and leaves specialists the judgment work that needs them.