Fifty-three points separate planning for agentic AI from governing it

Start with the arithmetic the two Deloitte figures make available. Seventy-four percent of organizations expect to be using agentic AI within two years ^[2]; 21% say they currently hold a mature governance model defining which decisions an agent can make alone and which require human approval ^[3]. That is 53 percentage points of daylight ^[4], with expected adopters outnumbering the governed by roughly three and a half to one ^[5]. The horizons differ, though, and that changes how the number should be read: one measures an expectation two years out, the other a capability in place now ^[2][3]. The governance work implied by the 74% has a two-year clock on it, and most of it has not started. Gharpure's claim is that the second number lags for a structural reason: autonomy gets treated as one policy settled once at the top, when it is five questions asked for every task an agent touches - criticality, predictability, governance and accountability, business maturity, and data quality ^[1][6]. Answer them and the task sorts into one of three lanes: execute alone where risk is low and predictability high; act with a professional reviewing the exceptions rather than every case where both are moderate; recommend only, with a human deciding, where risk is high and predictability low ^[7]. The debate that never resolves defaults instead to keeping a human in the loop everywhere, for now ^[12]. That default looks like caution and bills like headcount. Reversibility sits in that list. It is not its own axis. It is folded into the accountability question, as whether every action can be explained and every decision undone if it is wrong ^[6], and it returns in the operational readiness test as whether a human can quickly roll back an unexpected action ^[11]. For a middle-lane task that is the load-bearing question, because it decides whether a wrong call is an incident or a correction, and a task that is cheap to reverse can tolerate less pre-approval than its criticality score alone would suggest. The sequencing argument is the part with a cost attached. Gharpure's position is that governance bolted on after go-live, as a patch once something breaks, is where programs undermine themselves, and that the tiering belongs in the design: the highest level of human interface from the outset for processes touching financial or confidential data, moderate oversight for customer support ^[9]. His illustration is a biotechnology firm where more than 50 identified use cases were narrowed to those automatable with high accuracy, humans were embedded at the major decision points rather than spread evenly across every step, and throughput rose 30% with high error containment ^[10]. That figure comes from his own firm's engagement and carries no independent verification ^[1], so it supports the design principle better than it prices it. Five questions and three lanes describe any competent risk register, and task-level triage is old consulting furniture. The piece does not claim the axes are novel; the claim is about how often they get applied ^[1]. The corroboration that carries weight is McKinsey's, cited in the same article: nearly two-thirds of respondents name security and risk concerns as the top barrier to fully scaling agentic AI, ahead of regulatory uncertainty and technical limitations ^[8]. Were the binding constraint the technology or the rulebook, a per-task method would be beside the point. This quarter's version of the work is small: a list of the tasks agents already touch, a lane assigned to each, and a named rollback path for anything in the middle lane.