Gartner says agents aren't ready; 60% of companies plan to deploy them anyway

Gartner warned in April 2026 that fully autonomous agents were "not ready for the majority of enterprise use cases," according to a Forbes Tech Council column by Scott Zoldi, chief analytics officer at FICO ^[1][4]. The same column cites Gartner's expectation that more than 60% of organizations will deploy AI agents by 2028, up from 17% at the time ^[2]. Those two statements describe the same two-year window ^[2], and the second one is the plan of record for most boards.

The distance is not marginal. Going from 17% to more than 60% is a 43 point jump, roughly three and a half times the installed base ^[1], compressed into the period a vendor analyst has just labelled premature.

Zoldi's diagnosis is worth separating from his product. He argues that many banks are still struggling to close existing AI governance gaps before agents arrive ^[5], and that the familiar failure modes of AI - weak interpretability, hallucination, sycophancy - become materially worse when amplified across multiple agents left unchecked ^[6]. The harder problem is self-adaptation: once agents change their own behaviour, the interpretability and auditability of a decision can depend on operational sensitivities and the environmental conditions at the moment of execution, often with direct customer impact ^[7]. That is a governance problem, not a capability problem. A model that performs well and cannot be reconstructed after the fact is still an unauditable decision.

He is also unimpressed by the tooling now being sold against this risk. The market runs from agent verification platforms to controls against improper privilege escalation and rogue agent behaviour, which in his assessment address security while leaving collective agent behaviour and decision audit unaddressed ^[8]. Read that with the obvious caveat: FICO holds patented and patent-pending blockchain-based governance methods that codify an agent's development so its behaviour can be explained, monitored, controlled and audited, plus production task blockchains meant to capture the interdependency of multiple agents ^[9]. He is describing a gap his employer sells into.

The part operators can use without buying anything is the decomposition. In the anti-money-laundering deployment Zoldi describes, the question of whether an incoming deposit is illicit is broken into dozens of components, each executed by an agent with a single narrow task, with the micro-decisions orchestrated in milliseconds into yes, no, or flagged for human review ^[11]. One batch of focused language model agents handles identity - perpetual know-your-customer checks and sanctions list screening ^[12] - while focused sequence model agents assess the deposit itself against transaction history, including amount, method, account, institution and originating country ^[13]. FICO's two building blocks are focused language models with trust scores for language tasks and focused sequence models for real-time transaction decisioning ^[10]. Narrow tasks are auditable tasks; broad authority is what defeats the audit.

What to watch is whether the escalation path survives contact with volume. In this design, human review is one of three outputs ^[11], and the share of decisions routed there is the honest measure of how much authority has actually been delegated. Watch also whether firms moving toward the 2028 number ^[2] can produce a decision record for a single agentic outcome on request. If they cannot, they have deployed authority without accountability, and the readiness warning ^[1] will read as a dated liability rather than analyst caution.