An MIT Sloan/BCG panel of 50+ experts agrees at 72%: governing an agent as an 'autonomous decision maker' dissolves accountability — and a reader comment points at the model's own crack
According to the MIT Sloan Management Review and Boston Consulting Group expert panel published 8 September 2026 (Elizabeth M. Renieris, David Kiron, Steven Mills, Anne Kleppe), 72% of the 50+ experts surveyed believe governance that treats AI agents as 'autonomous decision makers' will fail: it would let companies 'launder responsibility through the machine'. Recommendation #3 is blunt — name a human accountable for every decision. A comment posted under the article on 12 September, by Srikanth Devarajan, raises the question that concerns us directly: if a company runs fifty agents, who reviews all those checkpoints?
The panel — more than 50 practitioners, academics and regulators gathered for the fifth year running — was reacting to a deliberately provocative statement: 'responsible governance that treats agents as autonomous decision makers will fail'. 71.5% of respondents (17.9% strongly agree, 53.6% agree) confirmed it. The core argument, summed up by Simon Chesterman (National University of Singapore): 'the more we speak as if agents 'decide', the easier it becomes for firms and governments to launder responsibility through the machine: the model recommended, the agent acted, the human shrugged.' The panel's recommendation #3: 'regulators, courts, and boards need someone to hold responsible, and pointing to 'the AI agent' is unlikely to cut it' — name a person accountable for every decision, before deployment, not after something has gone wrong.
The most useful comment isn't in the article itself: it's a reader response, from Srikanth Devarajan, posted three days later. He accepts the recommendation, then asks the question the article leaves open: 'if an organization runs fifty agents, how many checkpoints does that require, and who actually reviews them all?' Naming an accountable human doesn't remove the review work, it just moves it — and agents get deployed precisely to cut headcount, not add it. That is exactly the question a domiciliation network's director should ask any agent vendor before signing: who, on my staff, will review what, with how much spare time?
For AppH
- Pack Accueil doesn't create fifty checkpoints: the volume of decisions to review is bounded by how many visitors actually walk up to the counter that day, not multiplied by an agent's speed. The person who approves is already there, at their post — approval adds to work they already do, it doesn't create a separate review role.
- The panel's recommendation #3 ('name an accountable human, before deployment') is already Pack Accueil's structure: every conversation stays tied to the agency's file and to the person who validates it, never to 'the AI decided'.
Against / the honest limit
- The panel says this too (recommendation #4): govern 'the system, not the agent' — accountability shifts to the company deploying the agent, not the one that built it. That means the review burden still lands, ultimately, on the domiciliation agency's own staff, not on AppH.
- Devarajan's objection still holds if a customer deploys AppManager to automate far more than a single front-desk flow — ten branches, several modules at once. AppH hasn't hit that case at the scale described (fifty agents) yet, and has no magic answer for the day it does.
This panel confirms an intuition we've repeated since day one: 'autonomous decision' is a convenient fiction until the incident happens. But Devarajan's comment is the test the article itself never puts its own recommendation through: naming a human only works if that person has time to actually review. Our answer isn't 'we've solved this' — it's more modest: at an agency counter, the person who approves already exists, physically, before AppManager ever shows up. We don't have the fifty-agent case yet; the day a customer asks us for it, we'll ask them Devarajan's own question back.
Reviewed by a human at AppH