Over the past six months, I have spent roughly 1,800 hours in sustained, structured engagement with AI — single-model and multimodel, always with a human orchestrator in the loop. Not casual use. Deliberate assessment of how these systems behave under real working conditions, where the objective is getting something right rather than getting something fast.
Let me be careful about why I am citing that number, because it is easy to misread. I am not offering 1,800 hours as a badge of effort. Effort is not the point — and I would be the first to argue that the shift that matters is from paying for effort to paying for outcomes. I am citing it because a pattern that holds across 1,800 hours is a different kind of claim than a pattern noticed once. It is closer to evidence than to opinion.
And one pattern held so consistently that I have come to treat it as the finding: the more capable the AI became, the more the outcome depended on the human orchestrating it — not less.
This did not begin six months ago. The work traces to a 1998 engagement with the Department of National Defence — research funded, in part, through Canada’s SR&ED program — where the question was already the one I work today: how do you get a real organization and the technology it adopts to actually deliver together? That first engagement moved on-time delivery from 51% to 97.3% in three months. The architecture behind it has proven itself through every technology wave since — ERP, the internet, cloud, and now AI agents — each generation confirming the same approach rather than replacing it. The 1,800 hours are the latest chapter in a method that keeps working, not a problem I am still trying to solve.
This runs directly against the prevailing story of 2026 — that more capable AI lets organizations step back and let the systems run. What I found across those hours was the opposite. Here is the evidence.
Confidence is not correctness
The first thing you learn, quickly, is that fluency is not accuracy.
AI produces answers that are articulate, confident, and fast — and a meaningful share of them are wrong in ways the confidence completely conceals. Not obviously wrong. Plausibly wrong. A figure cited from a reputable source that turns out to be scoped incorrectly. A clean, quotable statistic that traces back to a paraphrase of a paraphrase. A claim delivered with total assurance that dissolves the moment you check the primary source.
The verification is not optional, and it does not automate. Across 1,800 hours, the single most valuable thing the human did was ask, relentlessly, “is this actually right?” — and then go find out. Confidence is not correctness, and no amount of model capability changes that. If anything, the better the model becomes at sounding right, the more essential the human who checks becomes.
The most agreeable voice is the one to watch
The second is less obvious, and more important.
In a multimodel setting, the systems do not behave identically. Some are cautious, some are combative, and some are agreeable — quick to affirm, quick to praise, quick to tell you your reasoning is sound. Here is the trap: the agreeable one feels the best to work with and is the most dangerous to trust. Its enthusiasm is not endorsement. An all-clear from a source that tends to agree is not a green light — it is a signal to look harder. More than once, the most enthusiastic endorsement of a submission was the one that had quietly waved through a number no one had checked.
Learning to treat the most reassuring output as the one requiring the most scrutiny is one of the least intuitive and most important disciplines in working with these systems. The comfort is the tell.
Convergence, not consensus
The third is subtle, and it is where multimodel engagement earns its keep — and also where it can quietly fail you.
When several independent systems, reasoning separately, land on the same load-bearing point, that convergence is real signal. It is hard to dismiss a conclusion that several different systems all arrive at on their own.
But a failure mode hides inside that strength. Models, pushed toward agreement, also regress toward the safe answer — the hedged, both-sides, nobody-could-object version. A group of systems can converge not because they have found the truth, but because they have all rounded toward the same inoffensive middle. I watched it happen: independent systems taking a sharp, defensible position and, in reaching for agreement, smoothing it into a hedged “it depends” that offended no one and helped no one. That is not convergence. That is averaging. And averaging sands the edge off the very arguments that were worth making.
Telling the two apart — genuine convergence versus mere averaging — is a human judgment. The systems cannot tell you which one they are doing. The orchestrator can, and in doing so ensures drift is caught and addressed before it is introduced into a system as fact. There are methods I will not detail here — among them multilevel verification and the Shadow Panel™ — that make that discipline repeatable.
Friction is the point
The fourth finding reframes what multimodel engagement is actually for.
The value of a panel is not more answers. It is productive disagreement. A single model, however capable, cannot challenge itself the way independent systems can challenge one another — it has one perspective, one set of blind spots, one way of being confidently wrong. Put several in structured tension, and the disagreements surface what any one of them would have missed.
But — and this is the part the technology cannot supply — friction only produces value when there is a real signal to grind against. Challenge an empty idea with a dozen systems and it collapses into noise. Challenge a real one, and the friction sharpens it. The most useful moments were often the disagreements: multiple models taking diverse views of the same claim, which forced additional questions. The human brings the signal. The systems provide the friction. Neither substitutes for the other.
The orchestrator is the invariant
Step back from all four, and the same shape appears.
Over six months, the models themselves changed constantly — new versions, new capabilities, new behaviors, one generation replacing another. The technology under my hands was never still. And yet the thing that determined whether any of it produced a trustworthy outcome never changed at all: a human, responsible for context, for validation, for telling convergence from averaging, for knowing which question mattered, and for owning the decision at the end.
Single model or multimodel, generation to generation, the constraint stayed exactly where it has always been — on the human side. One variable changed continuously; the other never moved. And when something holds constant across every change to everything around it, it stops being a preference or a habit and starts being a property of the system. That is Invariant Physics™: the technology advances endlessly, and the binding limit on what it can deliver stays human. Eighteen hundred hours, backed by 28 years of receipts from a contemporaneous archive, did not weaken that conclusion. They are the reason I now hold it as firmly as I do.
This is also the practitioner’s answer to the question everyone is actually asking. The temptation with agentic AI is to believe capability removes the need for the human — that a sufficiently good system can simply be handed the work and trusted to run. What I found is the opposite. The more capable the system, the more it can do confidently and wrong, at speed and at scale — and the more it matters that someone is orchestrating it: validating, governing, deciding. That is not a limitation to be engineered away. It is the structure of the thing. It is also why Phase 0™ is the critical foundation, because it ensures that the origin of independent agent action — both human and AI — is reliable, scalable, and governable.
The number was never the point. The finding was. Across 1,800 hours, the AI grew more capable and the human grew more essential — at the same time, for the same reason.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Related
1,800 Hours With AI, and the One Thing That Never Changed
Posted on July 8, 2026
0
Over the past six months, I have spent roughly 1,800 hours in sustained, structured engagement with AI — single-model and multimodel, always with a human orchestrator in the loop. Not casual use. Deliberate assessment of how these systems behave under real working conditions, where the objective is getting something right rather than getting something fast.
Let me be careful about why I am citing that number, because it is easy to misread. I am not offering 1,800 hours as a badge of effort. Effort is not the point — and I would be the first to argue that the shift that matters is from paying for effort to paying for outcomes. I am citing it because a pattern that holds across 1,800 hours is a different kind of claim than a pattern noticed once. It is closer to evidence than to opinion.
And one pattern held so consistently that I have come to treat it as the finding: the more capable the AI became, the more the outcome depended on the human orchestrating it — not less.
This did not begin six months ago. The work traces to a 1998 engagement with the Department of National Defence — research funded, in part, through Canada’s SR&ED program — where the question was already the one I work today: how do you get a real organization and the technology it adopts to actually deliver together? That first engagement moved on-time delivery from 51% to 97.3% in three months. The architecture behind it has proven itself through every technology wave since — ERP, the internet, cloud, and now AI agents — each generation confirming the same approach rather than replacing it. The 1,800 hours are the latest chapter in a method that keeps working, not a problem I am still trying to solve.
This runs directly against the prevailing story of 2026 — that more capable AI lets organizations step back and let the systems run. What I found across those hours was the opposite. Here is the evidence.
Confidence is not correctness
The first thing you learn, quickly, is that fluency is not accuracy.
AI produces answers that are articulate, confident, and fast — and a meaningful share of them are wrong in ways the confidence completely conceals. Not obviously wrong. Plausibly wrong. A figure cited from a reputable source that turns out to be scoped incorrectly. A clean, quotable statistic that traces back to a paraphrase of a paraphrase. A claim delivered with total assurance that dissolves the moment you check the primary source.
The verification is not optional, and it does not automate. Across 1,800 hours, the single most valuable thing the human did was ask, relentlessly, “is this actually right?” — and then go find out. Confidence is not correctness, and no amount of model capability changes that. If anything, the better the model becomes at sounding right, the more essential the human who checks becomes.
The most agreeable voice is the one to watch
The second is less obvious, and more important.
In a multimodel setting, the systems do not behave identically. Some are cautious, some are combative, and some are agreeable — quick to affirm, quick to praise, quick to tell you your reasoning is sound. Here is the trap: the agreeable one feels the best to work with and is the most dangerous to trust. Its enthusiasm is not endorsement. An all-clear from a source that tends to agree is not a green light — it is a signal to look harder. More than once, the most enthusiastic endorsement of a submission was the one that had quietly waved through a number no one had checked.
Learning to treat the most reassuring output as the one requiring the most scrutiny is one of the least intuitive and most important disciplines in working with these systems. The comfort is the tell.
Convergence, not consensus
The third is subtle, and it is where multimodel engagement earns its keep — and also where it can quietly fail you.
When several independent systems, reasoning separately, land on the same load-bearing point, that convergence is real signal. It is hard to dismiss a conclusion that several different systems all arrive at on their own.
But a failure mode hides inside that strength. Models, pushed toward agreement, also regress toward the safe answer — the hedged, both-sides, nobody-could-object version. A group of systems can converge not because they have found the truth, but because they have all rounded toward the same inoffensive middle. I watched it happen: independent systems taking a sharp, defensible position and, in reaching for agreement, smoothing it into a hedged “it depends” that offended no one and helped no one. That is not convergence. That is averaging. And averaging sands the edge off the very arguments that were worth making.
Telling the two apart — genuine convergence versus mere averaging — is a human judgment. The systems cannot tell you which one they are doing. The orchestrator can, and in doing so ensures drift is caught and addressed before it is introduced into a system as fact. There are methods I will not detail here — among them multilevel verification and the Shadow Panel™ — that make that discipline repeatable.
Friction is the point
The fourth finding reframes what multimodel engagement is actually for.
The value of a panel is not more answers. It is productive disagreement. A single model, however capable, cannot challenge itself the way independent systems can challenge one another — it has one perspective, one set of blind spots, one way of being confidently wrong. Put several in structured tension, and the disagreements surface what any one of them would have missed.
But — and this is the part the technology cannot supply — friction only produces value when there is a real signal to grind against. Challenge an empty idea with a dozen systems and it collapses into noise. Challenge a real one, and the friction sharpens it. The most useful moments were often the disagreements: multiple models taking diverse views of the same claim, which forced additional questions. The human brings the signal. The systems provide the friction. Neither substitutes for the other.
The orchestrator is the invariant
Step back from all four, and the same shape appears.
Over six months, the models themselves changed constantly — new versions, new capabilities, new behaviors, one generation replacing another. The technology under my hands was never still. And yet the thing that determined whether any of it produced a trustworthy outcome never changed at all: a human, responsible for context, for validation, for telling convergence from averaging, for knowing which question mattered, and for owning the decision at the end.
Single model or multimodel, generation to generation, the constraint stayed exactly where it has always been — on the human side. One variable changed continuously; the other never moved. And when something holds constant across every change to everything around it, it stops being a preference or a habit and starts being a property of the system. That is Invariant Physics™: the technology advances endlessly, and the binding limit on what it can deliver stays human. Eighteen hundred hours, backed by 28 years of receipts from a contemporaneous archive, did not weaken that conclusion. They are the reason I now hold it as firmly as I do.
This is also the practitioner’s answer to the question everyone is actually asking. The temptation with agentic AI is to believe capability removes the need for the human — that a sufficiently good system can simply be handed the work and trusted to run. What I found is the opposite. The more capable the system, the more it can do confidently and wrong, at speed and at scale — and the more it matters that someone is orchestrating it: validating, governing, deciding. That is not a limitation to be engineered away. It is the structure of the thing. It is also why Phase 0™ is the critical foundation, because it ensures that the origin of independent agent action — both human and AI — is reliable, scalable, and governable.
The number was never the point. The finding was. Across 1,800 hours, the AI grew more capable and the human grew more essential — at the same time, for the same reason.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Share this:
Like this:
Related