Rethinking AI 2025: Agents Fall Short Without Guardrails

I’ll be honest — praising agents as the mechanical wonder of 2025 feels a little like celebrating a new steam engine before anyone figures out where the rails go. McKinsey’s headline triad of agents, innovation, and transformation is seductive and not wrong; agents are a meaningful inflection point. Funny thing is, that doesn’t mean the rest of the machine is anywhere near ready to run.

McKinsey’s emphasis on agents implies two things: technical readiness and institutional appetite. Both exist. The demos are impressive, and the C-suite is at least pretending to pay attention. But technical capability doesn’t erase internal incentives, regulatory ambiguity, or legacy workflows; those are the friction surfaces agents will meet first.

An agent can draft a policy memo or triage a customer request. It cannot change who gets promoted, who controls the budget, or how risk is really measured inside a large firm.

That gap matters because value from a technology rarely flows automatically to the people building or deploying it. Cloud computing didn’t magically flatten org charts; it mostly advantaged companies that were already good at software. Without redesigned incentives and governance, agents risk sliding into the same pattern: point solutions that amplify existing power structures rather than redistribute efficiency.

Call it the “trust tax.” Every agent deployment requires someone to trust its outputs. Trust, in turn, is bought with process changes, audit trails, and accountability loops that are costly and slow to implement. McKinsey is right to spotlight innovation; my argument is narrower — organization design is the throttling valve on whatever benefit agents can theoretically deliver.

Look, innovation does not automatically imply equitable diffusion. The piece treats innovation as the driver of transformation, and that’s fair as far as it goes. But innovation tends to cluster. Sectors and regions with cleaner data, stronger infrastructure, and more adaptive management cultures will sprint. Others will trudge.

That gap is not an academic abstraction. It shows up as uneven reskilling demands, fractured labor markets, and localized productivity spikes that leave entire professions or regions negotiating from a position of permanent catch-up. If training budgets, certification frameworks, and public-sector capacity don’t expand alongside agent deployments, the headline benefits will be visible only in select firms and cities.

Human capital friction is not solved by better models. Agents rely on reliable data flows and clear ownership. Those are organizational problems, not engineering ones. Fixing them requires bargaining among stakeholders who often have conflicting incentives and different risk tolerances. You can’t federate data governance with a command line.

We’ve seen this movie before. When ERP systems rolled through corporate America, the promise was integrated, real-time insight. The reality was multi-year implementations that exposed how misaligned incentives, messy processes, and interdepartmental turf wars could turn a software rollout into a low-grade civil conflict. Agents threaten to do the same thing, just faster and closer to the “intelligence” layer of the enterprise.

Some will argue that agents will force all these issues to resolve — that market pressure will compel firms to overhaul incentives, clean up data, and invest in reskilling, so the diffusion problem fixes itself. That’s a plausible story. Market pressure is a powerful force.

It is also extremely good at rewarding short-term arbitrage. Firms that extract quick gains without investing in sustainable governance will win the first few innings, entrenching imbalances and making systemic reform harder, not easier. The path that produces fast productivity headlines may also hardwire brittle, opaque systems that are costly to unwind or regulate.

The practical remedy isn’t glamorous: build auditability, oversight, and workforce transition funding into early deployments. That means intentionally slowing down where the payout is politically or socially sensitive — which is precisely the kind of awkward advice consultants tend not to put in the executive summary.

There’s another missing ingredient in the McKinsey framing: shared infrastructure. “Agents + innovation + transformation” reads like a linear progression, but the critical middle layer is standards and public plumbing — interoperability rules, auditing practices, and public-sector capacity that let smaller players plug in without bespoke engineering.

Without that layer, agents become islands with different rules and dialects. Large organizations will pay for custom bridges between them. Smaller firms, cities, and agencies will simply opt out or accept whatever black-box tools a vendor offers on take-it-or-leave-it terms. The risk isn’t that agents fail; it’s that they succeed in such a fragmented way that only a narrow band of actors capture most of the upside.

You can already see the outlines: big tech building end-to-end stacks, consultancies selling highly tailored deployments, and everyone else trying to stitch together commodity tools with limited influence over design decisions. The longer we wait on shared standards and oversight, the harder it becomes to retrofit them into entrenched practices and expectations.

As William Gibson liked to remind us, the future arrives unevenly. Agents are arriving exactly that way — highly capable in pockets, constrained or mistrusted everywhere else.

If McKinsey’s piece is a useful map of where agents might go inside the enterprise, consider this a note scribbled in the margin: the next few years won’t be defined by how smart the agents get, but by who pays the trust tax and who gets stuck living with everyone else’s shortcuts.