A patient was being intubated. He was frightened, beyond words, and began to panic in the middle of the procedure.

In critical situations like these, there are a thousand things to do, every one of them vitally important. But the nurse in the room, faced with her patient’s distress, chose above all else to be human. Rather than consulting documentation or administering medication, she took a moment to sit and hold the patient’s hand.

When the patient recovered, he sought out that same nurse to thank her. “You saw me,” he said.

Stela Lupushor heard this story at the Century Summit hosted by the Stanford Center on Longevity, and keeps sharing it because it illustrates the thing most performance management design fails to do: see the person, regardless of the process.

Stela is a technologist. She believes in data, and in the benefits AI systems can bring to HR. Her question is not whether organizations should adopt these tools, but how. And she is clear that the rush to optimize and automate cannot come at the cost of seeing the people inside the system.

When Mistrust Builds Better Systems

Stela has spent a big part of her career working at the intersection of people analytics and HR technology. At IBM in the early 2010s, she led the team developing some of the first AI-assisted sentiment analysis tools for HR. The technology was brand new and met with its fair share of skepticism, including from the employees it was designed to support.

That mistrust did not sink the project. The way Stela’s team chose to engage with it made the system stronger.

Rather than marginalizing the skeptics, the team brought them in. Vocal critics became an advisory group, and the challenges they raised forced harder questions and better design. The resulting product was more trustworthy and more effective than anything the team could have built without that pressure.

“Having people who are mistrusting you actually in a way helps you build better solutions in the long term... Instead of you trashing us, come join us, help us solve the problems you see.”

Including the voices of those who mistrust you is good engineering, and an indispensable part of ethical, human-focused AI design. The same principle that defined the nurse’s choice applies here: building tools that affect people’s livelihoods means first seeing the people themselves, including the ones who don’t trust you.

Where AI Should Not Go

One of the first use case requests Stela’s team received at IBM came from a health department that wanted to use behavioral analysis to predict employee mental health and stress risk. It was technically feasible. At first glance, there was no reason not to do it.

Not long into the project, IBM’s legal team intervened with one important question: if you find out someone is at risk, what exactly are you going to do about it?

The answer didn’t exist. There was no protocol for intervention, no trained responders on staff. The data could surface a signal, but the organization had no infrastructure to respond humanely. Without a clear, ethical response pathway, surfacing the risk was not a neutral act. It was potentially a harmful one. The use case was dismissed.

This was an early lesson about the limits of data-driven approaches: data collection has to be integrated with a human-centered decision-making system, or it will not be effective. Stela now applies that principle through a checklist she brings to every new use case, informed by the risk-tiering approach in the EU AI Act (which classifies employment-related AI as high-risk by default) and the work of scholars like Ifeoma Ajunwa, whose research on algorithmic management has shaped the academic foundations of this field.

Stela’s Checklist: Before You Use Employee Data

1. What problem are you solving?

2. What decision will you make with the insight?

3. What is the impact on minority or vulnerable groups?

4. Can you verify effectiveness within 6 months?

5. If you were in the dataset, would you consider this fair?

The Real Cost of Efficiency

There is a version of AI adoption that is essentially reinvented outsourcing. Organizations identify tasks that can be automated, hand them to a model, and declare efficiency gains without asking whether the underlying workflow was well-designed in the first place, or what knowledge is lost along with the humans who are removed.

Stela calls this the efficiency trap. The hidden costs are greater than many organizations anticipate. Someone still has to manage the AI. Errors a human would catch slip through. And the kind of understanding that philosopher Michael Polanyi called tacit knowledge, the sort of know-how that can only be acquired by working alongside someone who already has it, begins to erode. That transfer cannot be automated.

This effect is clearest in the entry-level pipeline. PwC has projected a 32% reduction in U.S. entry-level hiring between 2025 and 2028, with global chairman Mohamed Kande publicly acknowledging that AI will likely mean fewer graduate hires going forward. IBM, by contrast, has continued investing in its registered apprenticeship program, a U.S. Department of Labor-recognized initiative that now spans more than 33 roles and is explicitly designed to create entry points for candidates without traditional degrees. These two strategies represent a fundamental disagreement about the source of organizational capability and what is worth investing in.

Stela believes IBM’s strategy is the more sustainable one.

“Companies that are cutting the entry level talent will likely not be able to build a robust leadership pipeline. It is a longer term game than the next quarter efficiency savings.”

A Tale of Two Strategies: Entry-Level Hiring in the AI Era

PwC: Reduce Entry-Level Roles

Bet: AI absorbs junior-level work. Efficiency gains outweigh pipeline investment.

Risk: Hollowed-out capability pipeline; no bench when AI fails or context demands human judgment.

IBM: Double Down on Apprenticeship

Bet: Tacit knowledge and troubleshooting require human apprenticeship. AI augments; it doesn’t replace formation.

Risk: Higher near-term cost; requires organizational patience and structured learning infrastructure.

Performance Needs Context

Most performance management systems measure success in one way: did the employee produce the output? The question they often answer poorly, or fail to answer, is the more important one: why did they perform the way they did, and what support do they need going forward?

It’s a much more meaningful question to ask: did they perform so well because they had access to the resources they needed, because they had psychological safety on the team and the support from the manager, and there was encouragement of risk-taking and celebration of failures as an opportunity to learn.

On Stela’s sentiment analysis team at IBM, one of the researchers routinely solved the project’s most stubborn technical problems overnight, and as a consequence often slept well into the morning. Measured by output, she was exceptional. Measured by timeliness and attendance, she fell well short of expectations.

Measurement systems are rarely set up to tell the difference between the “what” and the “how,” and they tend to penalize the most valuable contributors who don’t fit the mold. What you measure signals what you value. There is nothing wrong with using attendance as a proxy for participation, but organizations that over-index on proxies risk missing the more important performance factors underneath.

In the case of that researcher, typical proxy measures would have obscured the contribution she was actually making. The manager needed to see her with adequate context, and provide the flexibility she needed to keep doing the work.

Signals of Context-Aware Performance

What most systems miss, and what context-aware design surfaces:

• Psychological safety and quality of manager relationship

• Equitable workload distribution across the team

• Access to resources, tools, and sponsorship

• Team transitions: new hires, key departures, reorganizations

• Whether risk-taking and productive failure are celebrated or penalized

The Review That Doesn’t See You

Annual performance appraisals are among the most studied and most persistently criticized practices in organizational psychology. Gartner research has found that fewer than one in five HR leaders believe their performance management process actually achieves its primary objective, and 81% of organizations are still tinkering with the same basic model. Elaine Pulakos, one of the field’s most cited researchers, has spent years documenting why traditional reviews fail. Stela has plenty of criticisms of her own:

“Annual evaluation cycle is painful, it is not effective, influenced by recency bias... It evaluates the confidence of the manager more than the individual who performs. It is backwards looking.”

How can managers buck the trend and provide unbiased, forward-focused reviews? Increasing the amount and rate of feedback might seem like an easy answer, but that often creates problems of its own: feedback fatigue, shallow assessments, and the loss of the big-picture view.

Stela suggests a different approach. Input that is timely relative to the event, with specific and actionable suggestions, at a cadence that is frequent without becoming burdensome, and built around the work rather than the calendar.

“‘Continuous’ doesn’t mean every day, every moment of the day... It’s the time distance from the event and specificity that makes feedback impactful on performance. Is it recent enough so people still remember the occurrence and is it specific enough so people can work out how to do it differently next time... It has to be frequent enough but not overburdening.”

Stela believes younger and more progressive organizations will lead this transition, not because large incumbents don’t see the problem, but because the infrastructure built around annual reviews is slow to change. Companies still building their calibration cycles, legal documentation, and compensation linkage are more nimble. No matter the size or age, organizations willing to experiment with new systems of review now will have a significant head start.

Managers Can’t Be Everything

The modern manager role has accumulated expectations that no single person can meet reliably. Managers are asked to be strategic contributors, project coordinators, performance evaluators, career coaches, and much more. Stela believes it is unreasonable to expect managers to perform at this level, and that doing so actively hurts workplace culture.

“In a marriage some expect the partner to be everything for them, but that’s nearly impossible and likely unhealthy... Same for managers too. A better model could be to have a ‘career manager’ who is helping you with your growth and development, and a distinct ‘project manager’ who is leading the tasks and activities involved with specific projects or teams you’re working on.”

The person who assigns your work is not necessarily the best person to guide your development. At PwC, separating those responsibilities was standard practice, with career coaching and project management distributed across different team members. Coaching-as-service models, like Mento, take this idea further, allowing employees to seek development support independently from outside their organization.

Spreading the load is not a sign that management has failed. It reflects how important the function actually is. Distributing these responsibilities produces better outcomes than asking one person to do everything.

Design to See

The nurse in that hospital room didn’t optimize her workflow. She saw a frightened human being in a moment of vulnerability, and she responded with empathy. That simple act of being seen was what the patient remembered above everything else.

The question for anyone designing performance systems right now should be: what would it mean to build something that sees people the way that nurse saw her patient? Outputs are part of the picture. So are the conditions that produced them, the team transitions, the access to resources, and whether the support existed for someone to succeed.

AI, used well, can help answer some of these questions, providing context, prompting timely feedback, and reducing the cognitive load on managers so they have more capacity to be present. Technology is not the obstacle. The design intent is.

The future of performance management isn’t about doing more with less or faster.

It’s about seeing the human and their unique needs, so we can support each other better.

-Stela Lupushor