Agentic AI: judgment over hype, again

Everyone plans to deploy agents; few have. That gap is rational — and it's your window to get governance in place before the pilots multiply.

Two numbers from this spring's survey season are worth holding side by side. According to Gartner's CIO survey, more than 60% of organizations expect to deploy AI agents within two years. The share that actually has them in production today: 17%.1

The vendors will tell you that gap is your competitive window closing. We'd suggest a different reading: the gap is rational. It's what sensible organizations look like when the technology is genuinely promising and the operating model for it doesn't exist yet. Gartner itself predicts that more than 40% of agentic AI projects will be canceled by the end of 2027 — over cost, unclear value, and inadequate risk controls.2

So the honest question for a mid-market executive isn't "what's our agent strategy?" It's "what would have to be true before an agent is allowed to act on our behalf?" That question has an actual answer.

An AI agent differs from a chatbot in one decisive way: it does things. It doesn't draft the email; it sends it. It doesn't suggest the purchase order; it places it. Software has been doing things autonomously for decades — that's what automation is — but agents act with discretion, in situations nobody explicitly programmed, which is both the value and the risk.

Here's what doesn't change: accountability. When an agent acting for your company sends the wrong contract, mis-quotes a price, or emails a customer list to the wrong place, "the AI did it" is not a sentence your customers, regulators, or insurers will accept. Delegation without accountability isn't a technology risk — it's the oldest management failure there is, newly automated.

And the environment is already probing for the gaps: this winter brought the first large-scale supply-chain attack on an AI-agent skills marketplace — hundreds of malicious add-ons delivering malware through the very mechanism that makes agents extensible.3 The attackers, as usual, are early adopters.

The temptation is to run pilots first and write rules later. The cancellation statistics suggest that ordering is exactly backwards. Before any agent touches production anything, five questions need written answers:

None of this requires new technology. It's a page of policy and a set of access decisions — precisely the kind of unglamorous work that separates the 17% who are running agents calmly from the 40% headed for the cancellation statistics.

Within those guardrails, the mid-market wins right now are narrow, boring, and real:

Internal knowledge work with a human at the end. Agents that assemble the draft — the account summary, the first-pass reconciliation, the meeting brief — and hand it to the person who was going to do it from scratch. High volume, low blast radius.

Structured back-office flows. Invoice matching, data entry between systems that never got a proper integration, tier-one internal helpdesk triage. The agent operates where the rules are clear and the exceptions escalate.

Development and IT operations assistance. Watched closely by professionals equipped to catch its mistakes — the safest audience an agent will ever have.

What to wait on: anything customer-facing with commitments (pricing, contracts, promises) and anything touching money movement without hard approval gates. Not forever — just until the audit trail and the limits have a quarter of operating history behind them.

We wrote in our AI Ethics & Human Cognition whitepaper about the deeper risk of these tools: outsourcing judgment itself. Agents raise the stakes on exactly that point. Adopt them — the value is real — but adopt them the way you'd onboard a talented, overconfident new hire: clear scope, limited authority, close review, expanding trust.

Where to start, what to ignore, and how to manage the risk — that's been our AI & ML strategy brief all along. Judgment over hype. Especially now.