A center of excellence is a coordination structure: shared standards, a review path, expertise that individual teams cannot staff alone. A named owner who signs for a specific approval limit is a different object altogether, and the Forrester figure of 39% operating a center of excellence at scale measures the first thing [2]. A board that reads the gap as a budget problem can move that number without touching the question Jason Kurtz of Basware puts at the center of his argument, which is who answers when the agent approves the wrong payment [4].
The two figures share a respondent base, which makes the gap arithmetic rather than rhetorical. If 67% of enterprise finance and AP decision-makers already use AI for targeted AP use cases [1] and 39% run a center of excellence at scale [2], then at least 28 points of that base are using AI in accounts payable without one [16]. That is a lower bound, not a precise estimate, since the groups need not overlap neatly. Inside the AI-using population the floor is 28 divided by 67, or at least 42% [17]. Kurtz calls this a governance problem rather than a capability problem [6], and on his own numbers the description holds.
A skeptic's objection here is about provenance: the study was commissioned by the company Kurtz runs [1], and a finding that finance teams adopted AP automation faster than they built governance around it is a convenient one for that company to publish. The answer is that the second load-bearing number belongs to Gartner, whose forecast is that more than 40% of enterprises will demote or decommission autonomous agents because of governance gaps identified only after production incidents [7]. Kurtz treats that gap as evidence of an explanation problem: the agents may have performed fine, but nobody could reconstruct why they were allowed to do what they did [8]. It remains a forecast. Neither source supplies a count of payments agents have wrongly approved, or a case in which liability was allocated after one, so the exposure in this story is argued rather than measured.
The most useful part of the case is the line Kurtz draws between variability and accuracy [14]. An accuracy rate is what a board asks for and what a vendor reports each month. Variability is what appears when invoices clear cleanly for months and then an edge case is handled in a way nobody would have signed off on in advance, which Kurtz says erodes trust faster than inaccuracy does [14]. The reporting line a company is most likely to build will track the first and miss the second.
The board-deck version is two numbers and a funding request: adoption at 67, governance at 39 [1][2]. What it omits is that the binding decision sits below the structure. Kurtz's own operating rule, treat each agent like a new hire and widen its authority as it is earned [9], with low-dollar invoices automated and anything involving unusual supplier behavior or regulatory exposure held back [11], is a calibration made repeatedly at the level of individual workflows. The CFO and CIO he describes debating autonomy at a European manufacturer [13] were deciding who carries a delegation that already has consequences for actual payments. His position is that in 2026 leadership carries it, because an agent cannot explain to a board or a regulator why a control failed [12]. Whatever authority a company grants this quarter is the authority someone will have to account for after the first incident.
