Can AI Do Performance Reviews? Rippling Says Yes.

This week Rippling, a fast-growing mid-market HR Tech company, introduced a fascinating product called Talent Signal. This product reads work output data from Github (for engineers), Salesforce and Gong (for sales people), and Zendesk (for service reps) and uses AI to analyze their performance.

How does it work? The system uses several inputs. With only information about the employee’s job level (no data on age, tenure, gender), it does the following:

First it measures performance (lines of code, peer reviews, number of closed deals, closed cases, etc.), all of which is available in these enterprise systems. Using a broad rubric developed by Rippling’s Talent Signal client data, it evaluates the performance as “on-track,” “high-performing,” or “needs attention.”

Second, it looks at behaviors (language, code quality, sentiment) in the high performers and tries to apply these “rules” to others. So if all the top sales people are highly inquisitive, it evaluates others on this standard.

And third, it imports your rubric to evaluate performance based on company values. So you can “teach it” if your company values collaboration, or innovation, or customer intimacy, or perhaps quality.

The output has three forms.

First, it provides a rating (on-track, high-performing, or needs attention). Second, it ranks employees in a team within a function and job level.

Third, and probably the most valuable, it offers developmental feedback. The manager sees narrative about why an employee was rated a certain way and can dig into specific recommendations. On manager discretion, this information can be shared with employees.

Today, to limit costs and get early feedback, the system only operates on “the last 90 days” of performance.

The Implications Of AI-Driven Performance Assessment

Well now that the cat is out of the bag, let’s discuss what it means. Should we let generative AI evaluate employee performance? Do we trust it to compare us to peers? What about the “soft skills” implicit in a job?

Well before you get worried, note that this is already happening. Both Meta and Atlassian “grade” engineers based on code quality. Platforms like Eightfold, HiredScore (Workday), and most ATS vendors score job candidates against job descriptions. And advertising systems “score” us against product preferences hundreds of times a day.

Here’s my take.

This is coming fast and we just have to get ready for it. Workday launched AI-driven development plans almost a year ago, most L&D companies are building AI tools to assess skills. And “digital twins” are right around the corner (digital representations of your work, which tell the company what you’re good at).

I was on stage with Matt MacInnis, Rippling’s COO, when he launched it in Vegas. The audience was quiet, but once people saw how it worked they started to ask questions. There were no major objections, only fascination.

I pointed out that consulting firms (Korn Ferry, Heidrick & Struggles, SHL) have done “high-performance” analysis on teams for years. This form of performance consulting is essentially what Talent Signal does by algorithm. So the methodology is not new, only the AI implementation.

On the topic of bias, I’d argue this system is less biased than managers. Managers have proximity bias and whole variety of non-performance preferences.

And this leads to my next point: how would AI identify soft-skills behaviors at work?

Performance vs. Culture

For many years HR leaders have pushed the idea that it’s the “how” not the “what” that matters. While we want people to perform well, they also need to align with our values. (Juniper, a high-performing tech companies in the early 2000s, decided that “job output” was only 25% of a performance appraisal. Learning Agility, Relationships, and Alignment with Mission made up 75%.)

Here is how Jack Welch (GE CEO) modeled performance. High performers (upper row) who are not aligned with culture should not be in the company.

As he explained it, the box 3 (high-performance, poor culture fit) people undermined everyone else.

We’ve all experienced this situation: the over-ambitious hard-driver who ignores the team, manages upward, and refuses to share. How would Rippling measure this?

Well the system is intended to measure the “how” as much as it can. For sales and support it listens to calls and looks at notes (and eventually will likely read emails and meeting transcripts). So these types of AI systems are quite capable of measuring “culture fit.”

SHL, the leader in performance assessment, actually has 96 “soft skills” that correlate statistically to job performance. One could imagine Talent Signal reading this model and evaluating people against this highly researched rubric.

Despite its early stages, I’m very positive on this offering. Talent Signal could give managers and individuals powerful developmental feedback, and that, at a minimum, is a huge win.

Think for a minute about how managers coach and evaluate people today. They use data just like this (performance metrics, alignment with culture, teamwork, etc.), in a much more imperfect way. The whole movement toward OKRs is an attempt to improve such accountability.

Talent Signal is an AI Trailblazer, bringing forward an idea we’ve been waiting for. Look for many more tools like this coming soon.

The History of Performance Management. (From The Josh Bersin Academy.)