Guardrails first: resisting a rush to agentic AI

Deloitte wants us to sprint into an era of “human with agentic AI.” Sounds sleek. Sounds inevitable. But when you frame adoption as the problem to solve, you’ve already picked a side.

To be fair, their central claim has gravity: humans plus AI that can act on our behalf is where the corporate world is heading, whether executives feel ready or not. Hesitation, they argue, is a competitive liability. That’s not hysterical futurism; that’s a reasonable read of how fast tools are landing on people’s desks.

But follow the money.

Deloitte is in the business of telling companies how to move fast — then selling the maps, the seatbelts, and the fuel. When a consulting giant urges rapid rollout of “agentic” systems, that’s not neutral guidance. Vendors sell more platforms; consulting firms sell more integration blueprints, training programs, and governance frameworks. Clients sign transformation contracts. The incentives stack up like neatly aligned dominoes.

That doesn’t automatically make the advice wrong. It does mean we should read it as both argument and sales collateral.

What the piece sells most aggressively is a particular framing of control. “Human with agentic AI” sounds like human-led, reassuringly so. Yet in practice, when systems move from recommendation to action, power flows toward whoever designs the defaults. Do organizations retain meaningful oversight, or do people become approvers of pre-baked, vendor-optimized workflows?

I’ve watched this movie before.

When enterprise resource planning systems swept through corporate back offices, executives were told they’d gain visibility and discipline. They did — but they also quietly adopted the vendor’s idea of how finance and procurement should work. The software’s structure hardened into company policy. Agentic AI has the potential to do that again, only faster and in more domains.

Deloitte nods to governance, risk, and oversight. Sensible words. But in their telling, these look like items on an implementation plan, not hard boundaries backed by power. Governance frameworks help only when they can actually stop a deployment, not just generate a slide deck.

Who inside a company can say no? Does an internal audit team get real veto power over an agentic system that might shave points off operating costs? Or does “oversight” become a compliance tick-box taped on after the strategy team has already promised the savings to shareholders?

Here’s what they won’t tell you: in most large organizations, the people tasked with guarding risk and ethics are badly outgunned by the people chasing quarterly targets.

Then there’s the quiet assumption beneath the adoption drumbeat: that the data and infrastructure feeding these systems are ready to be trusted with agency. Agentic tools act — not just suggest — based on training data and live signals. If your data is skewed, stale, or incomplete, you’re not just making bad dashboards. You’re operationalizing those errors at speed.

Deloitte’s vision presumes a level of data hygiene and internal expertise that many firms simply don’t have. Getting there is possible, but it’s expensive, slow, and talent-intensive. Large corporations might manage it with dedicated teams and legal departments. Smaller businesses, pressed for time and cash, are more likely to buy “AI in a box” and accept the defaults. That’s how decision-making quietly migrates from local judgment to distant platforms.

Accessibility and cost barely register in the cheer for adoption. When only well-capitalized firms can afford to build and govern their own agentic systems, everyone else rents capability on someone else’s terms. A few platforms write the rules; the rest of the economy plays along. Convenient, isn’t it.

There is a strong counter-argument, and it deserves more than a line of polite acknowledgement. Rapid deployment can unlock genuine efficiencies, surface new patterns in operations, and cut drudgery that burns people out. Companies that experiment early often do learn faster, discovering failure modes in time to fix them. Waiting for the perfect framework can become a sophisticated way of doing nothing.

But speed without teeth is how risk multiplies quietly. Learning-by-doing is tolerable in sandboxes and low-stakes workflows. It turns ugly when the “learning” phase involves customers denied services, workers mis-allocated or disciplined by opaque systems, or public bodies automating decisions they don’t fully understand.

The practical path isn’t to slam the brakes on AI. It’s to separate how fast you deploy from how much authority you delegate. Roll tools out widely if you must, but keep high-stakes actions under explicit human control and insist on transparent logs that show who — or what — made which call. That requires enforceable contracts and regulatory clarity, not another glossy advisory template.

History gives us one more warning label. When algorithmic trading systems took hold in finance, banks hailed them as a way to enhance human traders. The story was “human with automated systems.” Over time, markets bent to the logic of the machines; strategies compressed into milliseconds. Then came flash crashes — sharp reminders that when you give systems the ability to act at scale, you inherit new kinds of failure, whether or not anyone “adopted” them carefully.

Deloitte talks adoption and glances at governance. It has less to say about who pays to build the enforcement muscle, or what happens when vendors promise “turnkey agency” and then point to clients when things go sideways. That silence is not accidental; it’s where liability and profit quietly part ways.

Agentic AI will find its way into mission-critical processes because big firms see clear upside and consultants see billable hours. The real story won’t be the slogans about “human with AI,” but the contracts, defaults, and logs that decide who actually steers when the system acts and the outcome stings.