Ireland writes the human handoff into financial services rules

The mechanism worth understanding here is not the wording of the rule but what it does to a number. Containment counts the interactions that ended without reaching a human, and many contact centres have measured their automation that way for years ^[5]. A right to request human intervention makes part of that number determined by the customer rather than by the design of the flow ^[11]. For journeys inside the rule's scope, an escalation that fails to arrive stops being only a service defect.

The adoption figure offered as context deserves a second look. Eurostat is quoted as finding that 20% of EU enterprises with 10 or more employees used AI technologies in 2025, "up by 13.5%" on the prior year ^[2]. Read as percentage points, the base was 6.5% and adoption roughly tripled; read as relative growth, the base was about 17.6% and the year added a little over two points ^[16]. Those are different stories about how fast this is arriving, and the sentence as written does not settle which one holds.

The operationally useful part of Liam Dunne's argument is that the failures he describes sit outside the model. Authentication breaks, a back-end connection returns the wrong status, routing sends the call to the wrong team, and the agent receives none of the context the customer already supplied ^[7]. A vendor's intent-recognition accuracy is therefore not evidence that a right to reach a person is being honoured, because the chain that delivers the person is a separate system with its own failure points.

A skeptic would say this is one rule, in one country, for one sector, argued by a man who sells call path testing for a living ^[4]. Both observations are fair, and Dunne concedes the legal scope is specific while arguing the message travels wider ^[1]. The honest reading is that the supplied record does not establish a pattern: no statute, regulator, commencement date or penalty is named ^[9]. What it does establish is a design requirement whose cost does not depend on whether the law spreads.

Public opinion cuts in both directions at once. More than 70% of Europeans told a 2025 Eurobarometer survey that robots and AI improve productivity, while 84% said AI requires careful management to protect privacy and ensure transparency in the workplace ^[3]. The distance between those two figures is at most about fourteen points ^[10], and it reads less as opposition to automation than as a condition attached to it.

This quarter's decision is narrow. Either escalation is a built capability that carries the reason for contact, verified identity, steps already completed, relevant history and the point of failure ^[13], or it is an exit hatch that opens only when a customer repeats the word "agent" ^[14]. Next quarter's consequence is a staffing one: Dunne expects human agents to concentrate on exceptions, complaints, fraud concerns, bereavement and financial distress ^[15], which is a more expensive bench than containment targets were built to fund. He calls regulation a floor rather than a service strategy ^[8], and on the evidence supplied that is the right altitude for it.