When AI Makes Work Worse: Lessons for Agent Builders

Kaiser Permanente nurses are pushing back on the AI tools being rolled out across their workflows, saying the systems add documentation burden, introduce surveillance pressure, and make patient care harder. That's worth sitting with before your next agent design session.

This isn't an anti-AI story. It's a deployment story. And it has specific lessons.

The problem isn't the technology, it's the fit

Most failed AI rollouts in high-stakes environments share a pattern: the tool was designed to optimize something the organization cares about (documentation completeness, throughput, compliance), without modeling what it costs the person doing the work.

Nurses in the Kaiser case reportedly describe AI systems that generate more charting steps, flag alerts that require manual dismissal, and in some cases create a surveillance layer that second-guesses clinical judgment. The AI is doing exactly what it was spec'd to do. That's the problem.

When you build an agent to reduce friction, you have to be precise about whose friction. Reducing friction for the auditor can mean adding friction for the nurse. Reducing friction for the compliance team can mean adding steps for the person at the bedside.

Automation that adds steps is still automation

There's a category error that happens early in agent projects: teams conflate "we added AI" with "we removed work." They're not the same thing.

An LLM that summarizes a patient chart but requires the nurse to verify, correct, and re-sign the output hasn't saved time. It's added a new kind of cognitive task: checking AI output under time pressure, which is often harder than doing the thing yourself because you're now responsible for someone else's draft.

The question to ask before shipping any agent: what is the human now doing that they weren't doing before? If that list is longer than one item, you haven't shipped an automation. You've shipped a co-pilot that needs its own co-pilot.

Surveillance is a design choice, not a side effect

Some of the nurse complaints center on monitoring: AI systems that track how quickly tasks are completed, flag deviations, or create audit trails that feel like performance tracking.

This is a design choice, not an inevitable feature of AI in the workplace. Builders and buyers should name it explicitly during scoping. Is this agent here to help the worker, or to report on the worker? Those two goals can coexist, but only if both are stated upfront and the workers know which one is primary.

When that conversation doesn't happen, you end up with tools that workers correctly identify as adversarial, even if that was never the intent. Trust erodes. Adoption tanks. The ROI case falls apart.

What good deployment looks like in constrained environments

A few patterns that hold up in high-stakes, time-pressured environments:

Start with tasks the worker actively hates. Not tasks leadership thinks are low-value. Ask the nurses what they'd happily hand off. That's where adoption is easiest and resistance is lowest.

Design for zero required interaction when possible. The best agents run in the background and surface a result. The worst ones create a new modal, a new approval step, or a new thing to monitor. If your agent requires the user to do something to get the benefit, you've raised the adoption bar significantly.

Measure the worker's time, not just the process's time. It's easy to show that a step in a workflow got faster. It's harder to show that the person's day got easier. Track both, or you'll ship the Kaiser outcome: a process that tests better in a pilot and lands worse in production.

If you're figuring out where agents would actually help your team (versus where they'd just add overhead), our free AI Opportunity Audit identifies your three highest-impact automations from just your website. Takes a few minutes and gives you a concrete starting point.

The takeaway for agent builders

The Kaiser situation is a useful stress test for any agent you're scoping. Run it through three questions: Does this reduce work for the person doing the task, or just for someone observing it? Does it require new human steps to function? Does it create monitoring that wasn't there before?

If the answers are "no, yes, yes" -- you're building the thing the nurses are complaining about.

Good agent design starts with the worker's actual experience, not the process map on a whiteboard.

Want this built right for your team?

If you're evaluating where AI agents fit your operations and want to avoid the deployment pitfalls above, book a call. We scope and build agents that actually reduce work for the people doing it.