Kaiser nurses say AI, surveillance are making their jobs and patient care worse

TL;DR Highlight

Kaiser nurses say call-time pressure and AI empathy scoring are hurting patient care

Who Should Read

Backend or product developers building AI monitoring or performance-scoring features for call centers or healthcare services

Core Mechanics

Calls over 15 minutes drew criticism
- Seven nurses told CalMatters that calls longer than 15 minutes bring management criticism or evaluation meetings.
- They said call time feeds into monthly performance scores.
AI graded empathy and tone of voice
- Nurses said Kaiser began testing an AI tool in summer 2024 that assesses empathy and tone in nurse and patient voices.
- The tests ended in November 2024, but the union was told managers may bring it back.
Kaiser denies using Average Handle Time
- A spokesperson said Kaiser does not use Average Handle Time to assess performance or enforce call time metrics.
- Kaiser declined to answer questions about the AI tool.
Breaks between calls reportedly shrank
- Nurses said they once got about 10 minutes for charting but now get 30 seconds or less when lines are busy.
- They said this pace can cause missed cues about a patient's wellbeing.

Community Reactions

The problem is metrics, not AI, some argue
- One commenter said most complaints in the article concern call center metrics and pressure to ration care.
- They noted the AI empathy tool was a discontinued 2024 pilot.
Medical AI tools help, per personal experience
- A commenter whose wife works at Kaiser said live translation and note summaries save her time.
- Another said their doctor reported much lower stress because AI handles note-taking.
Metrics degrade care, per experience
- A commenter said nurses at their rural hospital are tracked on campus and time in each room is reviewed.
- They noted nurses now zip in and out although staffing went up, and blamed the metric.
Goodhart's law and vendor snooping concerns
- One commenter pointed to Goodhart's law as a well-understood effect of efficiency pushes.
- Another noted a big contact-center vendor offers AI sentiment analysis of calls in progress.

How to Apply

Add exception rules to your metrics first
- Open your call or support quality metrics today and find any with a length threshold.
- Exclude cases tagged as crisis, interpreter, or serious diagnosis from length metrics.
Let humans overturn AI scores
- Show the evidence segment and an appeal button on the score screen.
- Ask for a reason instead of penalizing when staff override an AI recommendation.
Create a disclosure checklist before AI scoring
- Before shipping emotion or tone analysis, define the notice text and opt-out path for staff and customers.
- Log which model scored what.

Try It

copy & run

# Prompt to paste into Claude or ChatGPT (simplified metric review based on the article's case)

Our team is designing agent performance metrics.
Metric list: {{metric list}}
Support types: {{support types}}

Review them as follows.
1. For each support type, find 3 situations where a single metric like call length could harm care or quality.
2. Predict how people might game the metric (Goodhart effect).
3. Propose exception rules and a human review process in a table.

Next step: apply the proposed exception rules to your real dashboard queries.

Terminology

Average Handle TimeThe average time taken to finish one support interaction. Like a taxi meter, it counts time but not quality.

Algorithmic managementManagement where algorithms assign, monitor, and evaluate work. It is like an automatic grader watching over your shoulder.

Goodhart's lawWhen a measure becomes a target, it stops being a good measure. People start chasing the score itself.

TriageSorting patients by urgency based on their condition. Similar to the sorting desk at an emergency room.