What Highwire's CEO Just Got Right About AI and Performance Reviews

Most conversations about artificial intelligence (AI) in performance management begin with the promise of consistency and end with a vendor pitch. Praveen Chandrahas, chief executive officer (CEO) of Highwire, chose a different path: he commissioned research that asks what AI actually gets wrong in the review process, then shared the findings publicly. That is rarer than it sounds, and it deserves direct attention from anyone sitting in a chief human resources officer (CHRO) seat right now.

Who is Praveen Chandrahas, and why does his work matter here?

Chandrahas leads Highwire, a firm focused on performance intelligence. He is not a neutral academic — he has commercial interests — but the argument his team put into the field is pointed enough that it cuts against the short-term interest of many AI vendors who would prefer HR buyers to keep the hype cycle running. His core claim, delivered to HR Dive this week, is precise and quotable on its own: a good manager can tell the difference between a smart risk and a careless mistake, but that takes context — and context is exactly what AI does not have, as HR Dive (2025) reported.

That sentence is not a throwaway. It is a structural critique of how most AI-assisted review tools are built. They are pattern-matchers trained on historical output data — logins, tickets closed, deals won, code committed. What they cannot encode is the organizational situation around those outputs: a team member who hit 80% of target while covering for a colleague on medical leave; an engineer who shipped a cautious fix instead of an elegant one because the elegant one would have broken a client's legacy system. Pattern-matching scores both as underperformance. A manager with context scores neither that way.

Why should HR leaders care about this finding right now?

The timing is not incidental. More than 50% of employers are already offering higher-than-planned salaries to attract candidates with AI expertise, according to HR Dive (2025). That wage pressure will intensify scrutiny of every AI-enabled HR process — because the people receiving AI-generated performance scores in 2025 are increasingly the same people companies are fighting hardest to retain. A flawed review cycle handed to your highest-value talent is not a neutral administrative event; it is a resignation trigger.

The plain-language version of what Chandrahas is describing: AI performance tools work through machine learning models that identify statistical correlations between measurable inputs and defined outcomes. The models do not reason about causation; they detect patterns. When the surrounding organizational context shifts — a reorg, a market crisis, a strategic pivot — the historical pattern becomes misleading, and the model's output degrades without signaling that it has done so. That opacity is the operational risk CHROs need to price into any AI review deployment decision.

There is a secondary signal worth naming. Research on manager-led development published this week showed that structured manager involvement in employee growth directly improves retention outcomes, as HR Dive (2025) reported. That finding, read alongside Chandrahas's warning, points to the same conclusion from two directions: the manager's judgment and relational intelligence are not legacy features to be automated away — they are the retention mechanism itself.

What is the one takeaway that earns a slot on your agenda this quarter?

Chandrahas's research implies a design principle, not just a warning. If you are currently piloting or scaling AI-assisted performance reviews, the question is not whether to use them — it is whether you have built a mandatory human override layer with defined criteria. Most organizations have not. They have imported the tool and left the governance architecture to evolve organically, which in practice means it does not evolve at all.

The decision your leadership team needs to make is not "AI or no AI in reviews." It is "what categories of performance judgment are structurally unsuited to pattern-matching, and who owns those calls explicitly?" Chandrahas has handed you the research basis to put that question on a board agenda. The organizations that answer it in writing, before a wrongful-termination claim forces the answer in litigation, will be the ones whose AI investments actually compound rather than corrode trust. The question is whether your organization builds that governance architecture before the first contested review, or after it.