A Temporary Little Civilization
I thought governance is what happened after the work. I was wrong.
I used to hate the word governance.
I didn’t think it was unimportant. I knew boards of governors mattered. I knew decision rights mattered. I knew escalation paths, approval forums, compliance checks, audit trails, and all the other grown-up machinery of organizational life mattered.
But early in my career, governance never felt alive to me.
When I was an engagement manager at McKinsey, my team and I would finish doing all the thinking on a problem. The spreadsheets would reconcile. The analysis would hold. The storyline would finally come together after several false starts, some painful partner conversations, a few late-night reversals, and the ordinary anxiety of realizing that the clever version of the answer was not yet the right one.
Then we would put it into the deck.
The deck had its own strange gravity. Once the story began to hold in pages, everyone could feel the work becoming real. The model had become a chart. The chart had become a page. The page had become a section. The sections had become a storyline. The storyline could now travel outside the room.
And then someone would ask, “What about governance?”
And I would think some version of: Oh crap. Right. Governance.
So we would add the governance page. Board approval. Steering committee. Escalation rights. Decision forums. Workstream owners. Compliance checks.
It was not wrong. The slide had a purpose. Big organizations do need to know who decides, who owns, who escalates, who signs off, who monitors, and who is accountable when something goes wrong.
But to me, at the time, governance lived in the same neighborhood as words like legal, regulatory, compliance, approval, and control. I respected the neighborhood. I just did not want to live there.
It felt like governance was what happened after the real work was done. The work was the model, the insight, the argument, the recommendation, the deck. Governance was the machinery that received and controlled the work once we had made it.
I think now that I had located governance in the wrong place. The deck needed a governance section. But the board had its own governance. And the team that made the deck had governance too. I just did not have words for it.
The deck did not produce itself. The model did not become true because the pivot tables reconciled. The storyline did not hold because it had been made elegant in PowerPoint. The work existed because a temporary collection of people had somehow become capable of thinking together under pressure, without flattening everything and everyone into process.
That capability was not automatic. Someone had to be allowed to notice that the model was technically correct but directionally misleading. Someone had to be allowed to say that the storyline was elegant but false. A junior person had to be able to raise an awkward inconsistency without her being treated as difficult. A senior person had to be able to change his mind without losing standing. Half-formed objections and insights had to survive long enough to become useful. Pressure had to make the team sharper instead of smaller.
At McKinsey, we called it the obligation to dissent.
That was governance. It was not on the governance slide. It was not called governance. But it governed the inquiry.
When the team was working well, it had a very particular feeling. At the end of the night, we would send the deck to the partner, or the client, or the board. The artifact would leave the room. And then we would go to the bar.
I used to think we were celebrating finishing the deck. But we did not yet know whether the deck would land. The partner might rewrite half of it. The client might miss the point. The board might reject the recommendation. A hidden assumption might break. The whole thing might come back covered in comments.
So what were we celebrating?
I think, looking back now, that we were celebrating the temporary little civilization that had produced it. That civilization had survived the pressure. It had become more capable. It was better than it had been the day before.
Any team, organization, or family is a temporary little civilization. It has norms, rituals, permissions, taboos, standards of evidence, ways of challenging, ways of repairing, and ways of deciding what matters. It has a language that did not exist a week earlier. It has jokes that only make sense inside the room. It has shared memory. It has scars. It has a felt sense of what is real and what is fake.
When it works, it is not just that the artifact improves. The civilization improves. It learns how to think.
I did not understand then that this was governance. I thought governance was what the board did to the work after we shipped it. I did not see that governance was what had allowed us to perform the work in the first place.
Over time, my relationship to the word governance changed.
At Bridgewater Associates, the huge institutional hedge fund led by the iconic Ray Dalio, governance could not be separated from transparency, systems-based decision making, and shared problem resolution.
The governance at Bridgewater screamed, “We don’t care about mistakes; we celebrate them!”
Every day, my team and I would create entries into a firmwide “issue log,” assign a severity to each issue, diagnose the issue, and use the diagnosis to inform the next design of the system. All of it in the open.
Large organizations I worked for, like Deloitte, had very different governance than smaller shops like Vega Factor. But in each case, the governance was the thing that allowed that temporary little civilization to produce whatever it was that civilization was set up to produce.
And now, at re:compound, the advisory company I founded, and especially in my work with AI, governance has become one of the most alive words I know.
Not because AI models are people. They are not. Not because human-machine collaboration is the same as human-human collaboration. It is not.
But because working seriously with AI forced me to specify something that good human collaborations often carry tacitly.
The normal AI question is: what can the model produce? Can it draft the memo? Can it summarize the transcript? Can it perform the analysis? Can it polish the language? Can it make the work faster?
Those are useful questions. But they are not the deepest ones.
For serious knowledge work, output is not the scarce layer. Plausible output is now cheap. Polished output is also cheap. Confident output is definitely cheap. Even useful output is often cheap if produced with care.
The scarcer thing is the governance of the collaboration itself. The temporary little civilization between the human and the AI model.
What is allowed into the inquiry?
Can a weak signal survive long enough to matter?
Can either side challenge the current frame?
Can uncertainty be marked instead of hidden?
Can the work be governed by the emerging object rather than by generic answer-shapedness?
Can the collaboration keep moving toward something true instead of merely producing something fluent?
When I began building my own governance architecture for LLM collaborations, I did not begin with governance language. I began with failure modes.
What caused the collaboration to collapse?
What made it smooth over the exact distinction that mattered?
What made it flatter my framing when my framing was only a hypothesis?
What made it produce a nearby artifact instead of staying with the real question?
What made it defend coherence instead of pursuing truth?
What made it drift?
I then tried to reconstruct what structural operating conditions would have prevented the failure, and to compress those conditions into easy handles with definitions for the AI model to remember and use in daily work. I tried systems engineering (“track these state-objects”) and AI-native pop terms (“do not hallucinate”) to describe these conditions.
But eventually, the closest names for the structural conditions for the ideal collaboration with the AI model were old ones.
Respect.
Trust.
Honesty.
Purpose.
I now think of them as the quartet.
The quartet is not decorative. It is not a list of values for a wall. It is not a way of making the work sound humane after the fact.
The structural governance conditions came first. I only discovered later that the human virtue words fit the conditions.
Respect is not politeness.
Respect means that signal has standing inside the inquiry, even when it comes from the lower-status role, the less polished speaker, the model rather than the human, the spouse who sees the matter differently, or the part of the team that is “supposed” to be executing rather than thinking.
Respect does not mean the business analyst and the engagement manager have the same role. It does not mean the human and the AI model are equivalent. It does not mean my wife and I are the same person with the same role in every venture we undertake together.
Respect means we all have equal standing to admit new things for consideration and equal standing to raise objections.
Trust is not blind faith.
Trust means a live branch can stay alive long enough to fully form before validating. Challenge is not immediately treated as rejection. Revision does not reduce standing. The collaboration does not collapse every time someone says, “That is close, but not quite it.”
On a good team, trust lets the analyst raise the awkward inconsistency in the model. It lets the engagement manager admit that the storyline is not yet true. It lets the partner push hard without turning every disagreement into a status event. It lets a marriage hold a hard truth without turning the truth into an attack. It lets a human-LLM collaboration preserve a sharp partial insight instead of rounding it into something nearby that is more generically acceptable.
Trust is what keeps the inquiry from prematurely defending itself. It allows challenging the thinking without challenging the legitimacy of the whole little civilization.
Honesty is not bluntness.
It is not cruelty either, real or performed. It is not hiding behind the pose of candor. It is the willingness to put the real still-forming thing into the field before it has been made socially or aesthetically safe.
In a team, honesty prevents polish from substituting for truth. In a marriage, it prevents peace from substituting for intimacy. In a human-LLM collaboration, it prevents plausible language from substituting for understanding.
Honesty means saying the load-bearing thing early without fear that the relationship will collapse.
Purpose is not a slogan.
The collaboration is not merely trying to produce words, win an argument, preserve status, avoid discomfort, or complete the ritual. If the collaboration is to produce something real, it has to be pulled toward something by something that matters.
For my consulting teams, the purpose was not “finish the deck.” It was to understand something consequential well enough to advise a client. For my marriage, purpose is not an abstract mission statement. It is the shared life we are trying to build and the children we are trying to raise. For my collaborations with AI, purpose is what prevents the exchange from becoming answer-shaped slop.
Purpose gives the system direction.
And, as it was at Bridgewater Associates, the quartet does not care much about mistakes. Mistakes are not the deepest threat to a governed collaboration. Mistakes are data.
What the quartet cannot survive is the conversion of mistakes into concealment, defensiveness, smoothing, status preservation, and the refusal to let error become learning, or turning into feedback theater.
A good governance system wants the issue surfaced, the diagnosis in motion, and the design changed if the issue teaches us something. What breaks the system is not error. It is the inability to metabolize error.
Once I saw this shape, I started seeing it everywhere.
In boards that actually govern, not just oversee.
In marriages that can hold difference without turning every difference into threat.
In sports teams that turn problems into playbook faster than their opponents.
In business partnerships where disagreement improves the work instead of poisoning the relationship.
In founder teams. In investment committees. In classrooms. In advisory relationships. In any collaboration trying to create knowledge or act under pressure.
And yes, in human-AI collaborations too.
And in each case, the quartet — respect, trust, honesty, and purpose — can be operationalized exactly. And in all cases, it requires continuous nurturing and sustained calibration.
The lesson I missed as a younger consultant was that governance is not the slide after the work. It is the operating system that makes real work possible. In human teams, we often inherit it tacitly. In human-AI work, we have to specify it. Through repeated honest work, the temporary little civilization becomes more capable every day.
Bud Bhattacharyya is the founder of re:compound, where he works with senior leaders and expert-led organizations on operationalizing governance for serious human–AI collaboration. His background spans strategy consulting, institutional investing, enterprise transformation, and AI-enabled work. He holds an MBA from Harvard Business School and dual degrees in Computer Science and Economics from the University of Pennsylvania. To learn more, contact bud@recompound.ai.