Rethinking Work: AI Partners, Not Replacements

Start with a confession: McKinsey is right that humans, agents, and robots will be stitched into new skill partnerships. But right now the stitches look hand-sewn — messy, uneven, and happening in the back rooms of a small set of companies that can afford the tailor.

Where the article earns its keep is the pivot away from the tired binary — “AI replaces humans” — toward a more realistic tripod: human judgment, autonomous agents, physical robots. That reframing matters. Skill stops being a line on a resume and becomes a property of systems: interfaces, data flows, incentives. Let’s be real: most executives still talk like they’re buying a tool, not redesigning a workflow.

The piece’s big miss is treating “skill partnerships” as a template you can roll out like a software patch. You can’t. Healthcare, semiconductor manufacturing, logistics, and call centers don’t just have different cultures; they have different failure modes, liability exposure, and regulatory shackles. A misrouted customer email is not the same as a miscalibrated medical device. Yet the language flattens those differences into one smooth narrative of “augmented work.”

The real bottleneck is not intelligence, it’s plumbing. Everything unglamorous: approvals, liability contracts, retraining plans, line layouts on factory floors, union agreements, cyber insurance riders. When I was at Goldman, the models were almost never the cost center; integration was. That hasn’t changed just because someone slapped “agentic” on the PowerPoint. The math doesn’t lie — the budget lines that blow up are the ones labeled “implementation” and “change management,” not “cloud credits.”

McKinsey nods at value creation but glides past who pockets it. Ownership of the data, the fine-tuned agents, and the deployment pipelines won’t be evenly spread. Big, well-capitalized players already hold the best training data, the engineering muscle, and the risk appetite to rebuild business lines around human–agent teams. Smaller firms will be handed SaaS agents and told they’re now “AI-enabled,” while the real economics sit with whoever can refactor entire workflows, not just bolt a chatbot onto the front end.

That’s not just a productivity gap; it’s an access gap. When performance gains depend on proprietary data, custom workflows, and tightly integrated stacks, you get quasi-monopolies in outcomes: more accurate decisions, lower marginal labor costs, and much stickier customers for the firms that got there first. Markets will reward that concentration unless someone deliberately re-opens the playing field through standards, interoperability, or different data-sharing rules.

The article also treats training like a corporate L&D problem, when it’s actually an institutional design issue. If the productive unit is a human–agent–robot team, then training cannot stay siloed by job title or tied to single employers. We need portable, modular credentials that certify people to operate inside specific classes of workflows — not just “data analyst” but “claims analyst on audited agent workflows,” for instance. Without that portability, companies will hoard training, workers will carry more risk than upside, and the supposed “partnership” becomes another word for dependence.

Measurement is even trickier. Traditional metrics obsess over individual throughput and error rates. But once agents are involved, the unit of analysis should be the loop: who checks what, how often escalation happens, where the handoffs fail. If you don’t measure the loop, you end up rewarding the person who shoves work to the agent fastest, not the team that catches the subtle but catastrophic mistake.

On labor rules, McKinsey is gentler than reality will be. Job descriptions, bargaining frameworks, and even benefits are still built around humans as discrete units of labor. Once agents shoulder large parts of a role, the question “who gets paid for what” becomes non-trivial. Do workers participate in the gains when their job is redesigned around an agent, or just absorb the extra monitoring and accountability when things go wrong? Without explicit bargaining around workflow redesign — not just headcount — the upside goes to capital, the risk stays with labor.

Here’s where some history helps. When industrial robotics hit auto plants, companies that treated robots as bolt-on cost savers got modest margin improvements. The ones that redesigned the entire production system — Toyota being the obvious reference — pulled away. So, today, a company like Amazon doesn’t win because it bought more robots; it wins because it rebuilt warehouses, software, and jobs around tightly choreographed human–machine systems. That’s the scale of change “skill partnerships” actually demand.

McKinsey’s piece also underplays the frictions that will shape a two-speed economy. Procurement cycles, sectoral error tolerance, and regulatory overhead are not annoying side notes; they are gating functions. A bank can’t just slip an agent into credit workflows without audit trails, explainability mechanisms, and model risk sign-off. A hospital can’t quietly automate triage without confronting malpractice exposure and credentialing bodies. Fast adopters will be the ones that can afford entire governance teams; slow adopters will wait for pre-certified stacks and hope they’re not obsolete by then.

The practical moves here are less glamorous than keynote slides. Companies should define their smallest human–agent work loop and measure it ruthlessly. They should write auditability into procurement contracts by default, not as an add-on. They should treat cross-organization training credits as an asset, not a cost to be minimized. Regulators, for their part, should stop fixating on “AI systems” in isolation and focus on certifying specific classes of hybrid workflows, with clear accountability built in.

McKinsey is right that skill partnerships will sit at the center of how work is redesigned. The live question is whether these partnerships become a standardized, shared infrastructure — or a patchwork of bespoke, high-performing islands owned by the few firms willing to rewire themselves fast enough.