People surprise you. When the conditions let them.

Information, authority, and tools — placed close to the work.

The theory, stated plainly: the more information, the more decision right, and the more tools you place close to where the work actually happens, the more people will surprise you. It applies especially at the customer interface — because the operator there holds the most immediate signal, and because the decisions they make are usually the ones that matter most for the outcome. It applies elsewhere too.

This is a claim about where operational energy seems to come from. It is not a claim about what any particular organisation gets wrong. Operating models carry constraints I cannot see from outside, and the shape of any given setup reflects history, regulation, and choices I was not part of. What I can say is that in the settings where I have watched this work, people thrive when the conditions let them — and the conditions are more designable than most of us are trained to believe.

What follows is less an argument than a description. Of the conditions under which this seems to hold; of the pattern that sits against them; and of the research, some of it sixty years old, that has named both.

Systems do the lifting. Not speeches.

When an operator at the front under-performs, the temptation is to read it as a character issue — they were not ready, not motivated, not committed. My reading runs the other way. Performance at the front is almost entirely a function of the surrounding system: what the operator can see, what the operator can decide, what the operator can say upward without risk. When the system supplies those, capable people appear. When it withholds them, the same people can look different.

The distinction matters because character problems call for replacement, and system problems call for redesign. The second is harder to start and cheaper to finish. Most attempts to push authority downward run aground not because the intention was wrong, but because the intention was stated and the surrounding system was not changed. A stated intention without a structural change tends to produce either cynicism or bravado. Neither is useful.

Douglas McGregor made the same point in The Human Side of Enterprise in 1960. His Theory X / Theory Y framing argues that a leader's assumptions about people shape the systems they build, and those systems then produce the behaviour the assumption predicted. The book is sixty-five years old and keeps being right.

The self-built trap.

One pattern is worth naming, because it recurs often enough to be worth the name. People who have something to propose — a fix, an observation, a reframe — sometimes conclude, often without having tested the claim, that whoever sits above them will not act on it. So the form their ideas take degrades. First into suggestions framed as complaints. Then into complaints without suggestions. Then into silence. When they do propose, they propose apologetically, with the confidence already conceded.

This is a textbook case of learned helplessness — a phenomenon first documented by Seligman and Maier in 1967, reproduced in organisational settings many times since. Structurally, it is what Amy Edmondson calls the absence of psychological safety: a stable belief that speaking up carries more personal risk than staying quiet. Her twenty-year research programme at Harvard (summarised in The Fearless Organization, 2018) is the definitive source.

The trap is self-built in the strict sense that the person builds it. It is not self-inflicted. It is a rational response to a system that has not, so far, rewarded the proposals they did make. The way out is not a motivational speech. The way out is to change the observable consequences of proposing.

A simple rule, and the test that goes with it.

One rule I work with openly: bring me solutions, not complaints. Not because problems are unwelcome — they are welcome — but because a leader's bandwidth is finite, and the person closest to the work almost always knows more about a workable fix than the leader does.

When a team member brings a solution, I test it by a single criterion: how hard they will fight for it. Not whether I like it. Not whether I would have designed it that way. The conviction of the person who will execute tells me more about the probability of the fix landing than the elegance of the proposal.

This is, in operational dress, the same pattern venture capital has been operating on for decades. Y Combinator's long-stated preference for the founders over the idea is the public version of the same heuristic: the will to take down walls is the scarce input, not the concept. Kauffman Foundation work on early-stage angel investing points in a similar direction — team-quality factors weigh at least as heavily as idea-quality factors in outcome prediction.

The rest falls out of that. Strategic alignment has to hold — that much is not negotiable. Inside the alignment, the idea an operator owns outperforms the idea a leader hands down and the operator accepts. The former lands. The latter gets executed halfway and quietly buried.

From permission to intent.

The mechanism I use is close to what Captain L. David Marquet described after turning around the USS Santa Fe in 1999 (Turn the Ship Around!, 2012). Marquet's "ladder of leadership" runs from "tell me what to do" at the bottom, through intermediate rungs — "I think", "I would like to" — up to "I intend to" and beyond. At the lower rungs, language requests permission. At the higher rungs, language declares intent.

Which rung a person operates on is almost entirely a function of the system above them — whether their earlier "I intend to" was honoured, whether the proposals they volunteered produced visible consequences, whether dissent was treated as information or as insubordination. Marquet frames this as the source of energy in a performing organisation. Self-determination theory (Deci & Ryan, 1985; 2000) frames the same phenomenon psychologically: autonomy is one of three necessary conditions for intrinsic motivation, alongside competence and relatedness. Two literatures, the same lever, different dialects.

My observation is that many teams settle between "I would like to" and "I think", and that moving a team to "I intend to" comes down to three moves: consistent respect for intent language when a person uses it; visible protection of those who propose and are wrong; and a firm absence of punishment for good-faith proposals that do not pan out.

A system of conditions.

Delegating authority without the surrounding system tends to produce bad outcomes, correctly identify that the front line was not ready, and return the organisation to central control — usually with the stated lesson that decisions at the front cannot be trusted. What failed in those cases was not the delegation. What failed was the surrounding system. Five conditions, in the order they tend to need to land:

1. Information symmetry

   The operator at the front sees what the customer sees — the order history, the product context, the prior ticket trail, the SLAs at stake. Where they have to ask permission to see it, they will not use it. Usually the cheapest condition to supply and the one most often missing.

2. Decision rights written down

   Not "use your judgement" — a stated ceiling, a stated escalation trigger, a stated class of decisions the person owns outright. The widely cited Ritz-Carlton USD 2,000 per employee per guest is not a slogan; it is an engineered decision right inside a defined boundary.

3. Training that transfers the model

   Not the script. People need to understand why, not only what. Scripted autonomy fails in the first moment that reality diverges from the script — which is, by definition, the moment autonomy is supposed to be doing work.

4. Coaching, not correction

   When a decision at the front goes sideways, the default move is to walk the reasoning back together — not to retract the permission. Retraction teaches the organisation that proposals are revocable, and the trap tightens. The permission stays; the reasoning improves.

5. Observable consequences of proposing

   When a proposal lands, it lands visibly. When it does not, the lesson is surfaced without humiliation. The rule people read off the system: proposing is safe, mediocre proposals get better, and the cost of silence exceeds the cost of being wrong.

Most of this sits, in slightly different language, in Zeynep Ton's The Good Jobs Strategy (MIT Sloan / New Harvest, 2014) — four operational choices (invest in people, standardise the work while pushing latitude down, cross-train, operate with slack) tracked across Costco, Trader Joe's, QuikTrip, and Mercadona. Her counter-intuitive empirical finding: the firms with the highest front-line investment had the lowest prices and the highest customer satisfaction. The usual assumption that the two trade off is simply wrong.

The same conditions, a different workforce.

The same conditions seem to carry into AI-mediated operations, and this is the part of the theory that has sharpened most over the last several months of hands-on agent work. An agent at the customer interface needs what a human operator needs: information access, a defined decision ceiling, a legible escalation path, and a coach rather than a corrector. Harvard Business Review's February 2026 coverage names this as a workforce-management problem with a workforce that happens to be software.

The reading goes both ways. Teams that have built the conditions for human front-line autonomy have a head start in designing good AI systems; the architecture is already there, and an agent drops into it. Teams that have not will tend to build AI systems that inherit the same missing conditions — and fail for the same reasons. The technology changes quickly. The operating principles change slowly, if at all.

If anything, AI makes the operating principles more legible. An agent lacking information, decision rights, or a clear escalation path fails quickly and visibly, in ways a human operator can sometimes mask for longer. The same system that was holding a human front line back becomes observable in the AI front line within a week.

Read the Build tier in full

Where these arguments come from.

A short, opinionated reading list. Every one of these is worth reading in full. None of them should be read uncritically.

Talk it through?

A thirty-minute call usually surfaces the one condition that, once present, would make the biggest difference in how work runs at the front of an operation. The most useful calls come from leaders who want to test this theory against something concrete they are already working on — a team, a role, a service design, a first AI deployment.