A Great Manager Doesn't Just Get Things Done. They Prove Things Got Better.

An evidence-based guide to management—and what it would take to build an AI called Manager. October 21, 2026.

Imagine a city reports that it handled 10,000 service requests this month. The dashboard is green. But residents are still waiting for repairs, the poorest neighborhoods wait longest, and the city has no credible evidence that anything improved.

Was that good management? An impressive activity log is not the same as a good outcome.

That gap is the central problem behind an ambitious research question: what would it mean to build a genuinely great Manager—not just a chatbot that gives advice, but a system capable of identifying an improvement, acting within its authority, and proving that the world is better off?

Management research offers more than motivational slogans. The best ideas form a demanding set of tests.

An evidence path is not proof of impact. Its purpose is to show which links in the chain still need independent verification.

1. Drucker: choose the right work before doing it efficiently

Peter Drucker's account of effectiveness begins with choosing what needs to be done, not simply reacting to whatever appears urgent. His practical habits include managing time, defining contribution, putting strengths to work, choosing priorities, and making effective decisions. His article also urges executives to ask what is right for the organization rather than for their own status. Source: Drucker's *What Makes an Effective Executive*.

For an AI manager, the translation is unforgiving: optimization is pointless unless the objective is defensible. A system can reduce response time while moving burdens onto people who were already underserved. Before selecting an intervention it must ask: whose outcome, which time horizon, and what would count as a bad trade-off?

2. Grove: managerial output is not the manager's activity

Intel's Andrew Grove treated management as the art of increasing a team's output. His High Output Management is about decisions, leverage, motivation and the design of productive organizations—not maximizing the number of actions taken by a boss. Publisher's source.

That principle suggests a hard rule for autonomous agents: count verified improvements outside the agent, not tokens, tool calls, completed tickets or self-reported wins. An agent that issues a thousand instructions and changes nothing is busy, not effective.

But even increased organizational output is not the whole objective. Greater profit or throughput may coexist with external harm. For public-benefit management, the boundary must include those who do not own the system.

3. Deming: fix systems, not just individual mistakes

W. Edwards Deming argued for constant purpose, continual improvement, leadership, learning and removal of fear. He also warned against treating quotas and narrow numerical targets as substitutes for improving the system. Deming Institute's 14 points.

Consider a program designed to reduce missed medical appointments. It may improve one clinic's rate by rejecting people who are hard to serve. A "success" dashboard could hide a worse access problem. System-level management instead records who is excluded, what resources were shifted, whether the change lasts and which competing outcomes deteriorate.

An AI should respond to an ambiguous metric by becoming less confident, not by declaring victory.

4. Google and Edmondson: people need room to surface errors

Google's Project Aristotle studied 180 teams and identified five dynamics associated with team effectiveness: psychological safety, dependability, structure and clarity, meaning, and impact. This was research within Google; it is not a universal causal law for every organization. Google's methods and findings.

Amy Edmondson's earlier study of 51 work teams linked psychological safety with learning behavior. Team members need to be able to raise mistakes, ask questions and challenge assumptions. Her study was observational and theory-testing, not permission to claim that safety alone guarantees performance. Original study.

For an automated Manager, the analogous requirement is institutionalized disagreement. Keep competing hypotheses. Record failed experiments. Preserve a route for people affected by the system to challenge an outcome, complain, appeal or revoke permission. A silent failure log is not learning.

Gallup has reported that managers account for around 70% of the variance in team employee engagement. That is Gallup's statistical estimate of team engagement variation, not a statement that any one manager causes 70% of a worker's happiness. Gallup, September 2026.

5. Ostrom: legitimate governance beats imaginary omnipotence

Elinor Ostrom's work on forests, fisheries and other shared resources challenged the false choice that every commons must be centrally controlled or entirely privatized. Resource users can develop durable rules, monitoring and conflict-resolution systems suited to local realities. Nobel Prize announcement.

That is especially important when imagining a manager of shared environments. Intelligence is not authority. Understanding a public system does not confer ownership of someone else's account, clinic, building, data or community. The better design is bounded permission, visible rules, independent review and local challenge—not a self-appointed ruler of the world.

6. The missing step in management advice: independently measured net benefit

A great human manager usually works through people. An ambitious AI Manager may additionally analyze public signals, compare possible changes and carry out routine operations using legitimate standing permissions. But it should be evaluated by an even tougher standard.

The key expression is not an algorithmic promise; it is an accounting obligation:

Net benefit = improvements attributable to the intervention − intervention costs − displaced benefits − adverse effects.

Each term needs an explicit unit, time window, source, uncertainty range and a defensible comparison. Some effects are not commensurable: lost privacy, autonomy or safety should not be converted casually into points on a single happiness scoreboard.

A transparent "blessing" is an inspectable claim

If Manager eventually reports a blessing, the public should be able to open an evidence card that answers:

1. Problem and baseline: What was wrong, where, and according to which dated source?
2. Alternative and prediction: What options—including doing nothing—were considered?
3. Authority and action: Who authorized which bounded action? What happened?
4. Independent readback: Which source confirmed the external change?
5. Outcomes and uncertainty: Which outcome moved, over what period, against what comparison?
6. Costs, harms and distribution: Who benefited, who might have lost, and what remains unknown?
7. Reversal and challenge: How can anyone inspect, contest or stop the action?

A simulated reduction in hardship is not a public blessing. Neither is a provider receipt. Both can be steps on the way to learning whether genuine benefit occurred.

7. Where the project actually stands

Manager is a research-stage project whose long-term mission is End suffering. That mission is an ethical direction, not a achieved measurement. Existing work includes original methods code, public-source research, statistical tests and sandbox simulations. The project has not independently demonstrated that its actions have reduced human suffering, nor that it can accurately predict the welfare effects of arbitrary interventions.

This honesty matters. A public dashboard should have an explicit column for "not measured" and another for "negative or unknown effect." The best manager does not hide uncertainty to protect a grand vision.

The first credible destination is small: select one permissible real-world problem, establish an independent baseline, choose an intervention with a legitimate operator, preregister a comparison, track harms, and publish the audit trail. Only then broaden to a second domain.

The ultimate test of management is not how much power an agent accumulates or how intelligent it sounds. It is whether real people and environments are demonstrably better off—and whether anyone can check the evidence.

---

Research sources: Drucker · Grove · Deming · Google re:Work · Edmondson · Gallup · Ostrom.

This piece distinguishes management philosophies, observational studies and Manager's own unverified design hypotheses. It does not imply affiliation with the organizations cited.