Let an agent decide alone only where the decision is low-ambiguity and reversible, MIT Sloan's routine quadrant, and only after a named human owner has reviewed its track record. KPMG's Q3 2026 AI Pulse (September 24) found 49% of leaders now define no-go zones for autonomous agents. Sir John Whitmore's Coaching for Performance supplies the promotion rule: responsibility is granted as awareness is demonstrated, never on confidence.

Three reports landed this week on a question I get in almost every coaching session with a founder who runs agents: which decisions can the thing make without me? I run a handful of agents in my own company, and the honest answer used to be "whichever ones I haven't been burned on yet." The reports gave me something better than that, and one number in them worried me.

The facts first. KPMG's Q3 2026 AI Pulse, published September 24, surveyed 314 US C-suite and business leaders at companies with revenue of $1 billion or more between July 24 and August 25. Sixty-two percent are building or deploying AI agents, up from 53% last quarter; the share developing multi-agent systems jumped to 25% from 6%; and 49% of leaders have defined high-risk use cases where they do not allow autonomous decision-making by agents. Confidence in governance rose to 73% from 57%. The number that worried me sits next to it: the share of organizations building controls into their agents alongside monitoring and evaluation fell to 30% from 43% two quarters ago, because, KPMG says, many are relying on guardrails already in place. ISG's State of Enterprise AI report, published September 23 from a survey of 400 senior decision makers, puts the current reality in proportion: less than 7% of AI-enabled work is performed autonomously today, 55% is human-led, and companies expect the autonomous share to nearly double to 13% by the end of 2027, with decision quality up an average of 4.3% where AI is used. And on September 22 MIT Sloan published the AI Decision Matrix from Ina Sebastian, Peter Weill, Thomas Haskamp and Jan vom Brocke at MIT CISR, built from interviews with 30 executives. It sorts decisions on two axes, ambiguity and risk, into routine, consequential, exploratory and strategic, and splits every decision into framing, acting and learning. Their case company, One New Zealand, runs more than 50 agents; a segmentation agent cut marketing time by 60% with a human reviewing the plan before launch, pricing decisions always keep a human in the loop, and every deployed agent has a named human owner.

Here is what the three say together, translated for someone with four agents rather than fifty. Confidence going up while built-in controls go down is the exact shape of every delegation failure I have watched in coaching, and the book that explains it is not an AI book. Sir John Whitmore's Coaching for Performance argues that performance equals potential minus interference, and that you raise it through two things, awareness and responsibility, granted in that order. Telling someone what to do creates dependence; asking them to see reality clearly and then choose the action creates ownership, and "when I truly choose to do something, my commitment and my performance go up." Whitmore's Performance Curve moves a culture from dependent through independent to interdependent, and the leap is earned by demonstrated awareness, never by how confident anyone sounds. An agent is a new hire you are coaching up that curve, and the MIT matrix is simply the curve drawn with two axes. Framing is Whitmore's Goal and Reality, and it stays human. Acting is where Options live, and an agent can own more of it as the decision gets more routine. Learning is Will, and it belongs to the named owner.

The protocol I now run, which I call the decision-rights ladder, has four rungs:

  1. Classify before you configure. For each decision an agent touches, answer two questions in writing: would two smart people on my team reach the same answer from the same data (ambiguity), and can I reverse it within a day (risk)? Routine decisions get automated action with a human reading a sample for the first month. Consequential ones get "agent proposes, owner approves," with the non-negotiable constraint written down the way One New Zealand wrote "no autonomous pricing." Exploratory ones get autonomous action inside a small budget with a human reading results weekly. Strategic ones get agents that gather and analyze while a human decides.
  2. Keep framing human. The agent never writes its own goal or its own definition of success. If I cannot state the Reality in a sentence, the agent is not ready for the decision either.
  3. Name an owner who chose the job. Not the person nearest the tool, the person who said yes to owning the outcome. No named owner, no autonomy. This is Whitmore's responsibility pillar made operational, and it matches the MIT case exactly.
  4. Promote on track record, demote on the first surprise. An agent moves up one quadrant only after a run of reviewed decisions with nothing the owner would have vetoed, and it drops back the moment it does something the owner did not expect, even if the outcome was fine. This rung is the one KPMG's 43% to 30% drop tells me companies are quietly skipping.

Where this breaks. The matrix comes from 30 executive interviews and one telecom; KPMG's respondents run billion-dollar companies with governance teams, and a founder with three agents has less oversight but also a smaller blast radius, so the ladder is a discipline, not a compliance program. ISG's 4.3% decision-quality gain is an average across enterprises whose outcomes vary widely, and ISG itself warns that human validation becomes a bottleneck when agents produce results faster than people can review them, which is the failure mode of rung one done badly. And the mistake I still catch myself making is classifying a decision by how confident the agent's output sounds rather than by whether I can undo it tomorrow. Whitmore would say the confidence is interference. The reversibility is the reality.

Sources: KPMG, "AI's Value Story Sharpens as Organizations Gain Confidence in Governance, Accountability and Workforce Adoption," Q3 2026 AI Pulse Survey, September 24, 2026 (kpmg.com/us/en/media/news/q3-ai-pulse-2026.html); MIT Sloan Ideas Made to Matter, "A framework for determining when AI can make decisions," September 22, 2026, on the MIT CISR AI Decision Matrix (mitsloan.mit.edu); ISG, "AI Is Changing How Work Gets Done, but Business Value Still Lags," 2026 State of Enterprise AI: The AI Value Gap, Business Wire, September 23, 2026; welcome.ai summary of the KPMG survey's value and decision-speed figures, September 24, 2026; Sir John Whitmore, Coaching for Performance, 5th edition (Nicholas Brealey, 2017).

Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · Why Exploration Is Important for Success