Knowledge work is structured to make the human the bottleneck.
For most of the last twenty years, that was fine. Creation moved at human speed. Coordination moved at human speed. The work around a decision and the decision itself ran on the same clock, which meant the person making the call could keep up with the work piling up on either side of it. Email, spreadsheet, approval doc, contract, redline, PDF, folder. Slow, but the two halves moved together.
Computers and the internet made knowledge work faster, but they did not break the rhythm. Email made sending faster and asking faster. Search made finding faster and sharing faster. The two halves of the work moved at the same rate even as the absolute speed climbed.
AI is different. It speeds up one half of the work and not the other.
Creation got faster. Research got faster. Drafting got faster. Volume got faster. The judgment in the middle did not get faster (reading the room, knowing which counterparty is bluffing, deciding whether silence means leverage or fear, taking accountability when something goes wrong). It cannot. Not because humans are magic, but because that part of the work is not information processing. It is reading incentives, managing trust, and making calls that someone has to answer for. None of that compresses without changing what the work is.
Picture the operator. Ten years on the job. They read the room. They know which counterparty is bluffing and which one is actually walking. They have built workarounds for the systems that were supposed to help them: a personal spreadsheet that tracks what the CRM doesn't, a folder of templates, a Slack channel they use as memory. Most of their value lives in the calls they make. Most of their week is spent on the work around the calls.
That person is about to be squeezed harder than they ever have been. The decisions they make will take the same amount of time they always did. The work surrounding those decisions will multiply tenfold. They will be doing the same job at the same speed, and they will look slow. The company will lose money. Deals will stall. Quality will drop. Trust will fray. Everyone will read the failure as a problem with the person, because the person is the visible part.
It is not a problem with the person. It is a problem with how the work is structured.
This is not theory. Eighteen months ago I was pitching an agency executive on AI in her industry. She cut me off. "I don't care about AI. We're trying to negotiate three hundred creator deals simultaneously and our entire team is just going back and forth on email. We're dying here in valueless admin." She was not asking about AI. She was telling me what was already happening.
One creator deal takes 60 to 80 emails from first outreach to signed contract. A 300-creator campaign produces 18,000 to 24,000 messages before anyone makes any content.
Nobody puts this work on a resume. Sending thirty "just following up" emails. Moving information from a spreadsheet into a CRM. Hunting through three threads to find what was agreed to last week. I was that person.
And the person who looks like they're failing at their job is mostly failing at infrastructure. The colleague who missed the email looks like they don't care. The hire who didn't follow up looks unprofessional. The vendor decision that went to the loudest one looks like bad judgment. None of those are character problems. They're tax. Infrastructure failure shows up as moral failure. We blame the person standing closest to the thing that broke.
The dominant conversation about AI splits into two camps and both of them are wrong. One camp says the human goes away. AI handles the call, AI handles the deal, the human in the middle was a relic. The other camp says keep the human in every loop, respect every existing step, do not let the machine make decisions. One produces software that treats the human as a bug. The customer discovers, painfully, that the human was the feature. The other produces software that inherits every existing delay. Different mistakes, same outcome. The bottleneck stays where it was.
The version that works is narrower. The decision stays. The work around the decision goes.
To see why the decision can't go, picture producing a Luke Combs stadium show. Artist team, venue, promoters, sponsors, city permits, everyone needing to align. The artist's manager spends the call negotiating terms. Underneath, he's managing a client who wants certainty and protection. The venue talks about availability. What it's actually doing is juggling other bookings and protecting relationships. The sponsor might not actually have board approval yet, but can't admit it. Everyone is saying one thing while protecting something else.
That isn't information processing. It's reading incentives, managing egos, and building trust. You're constantly judging when to push and when to wait. When the other side goes silent, you have to figure out what it means. Hesitation. Leverage. Fear. Sometimes nothing at all. The harder version is judging your own gut. When it tells you something is off, is that real information or noise?
Could AI eventually automate more of that? Maybe. But not without massive structural change to how humans coordinate. Change that takes years, not months. Maybe longer.
AI also can't be accountable. If the outcome is wrong, no one can answer for the algorithm in a way that satisfies the counterparty, the regulator, the patient, the board. That alone keeps the human present anywhere accountability is real.
So the human stays. The question is what they spend their time on.
Most of what slows them is not their job. They are doing the work systems should do. The person in the middle is the API, moving information from one system to another, holding the memory of what was agreed, manually transferring terms across templates that should already know each other.
The judgment was the job. The rest was overhead.
That is the work that has to go. Not because humans are bad at it. Because nobody became a campaign manager, a lawyer, a recruiter, a producer, or a deal lead because they wanted to be a piece of plumbing.
The architecture matters less than the principle, but the principle has to be picturable. Memory has to travel. Context has to carry between deals, so the second time you negotiate with someone, the system already knows what got asked for and what got conceded. The veteran's judgment has to become the platform's default, captured and reusable, so the new hire's first deal arrives loaded with the lessons of the last hundred. None of that is AI. The AI parts are useful only if the layer underneath them actually works: templates, intake, permissions, workflow state.
Then AI handles what carries: drafting, surfacing, cross-referencing, spotting contradictions, comparing against history. None of those are decisions. They are the work around decisions humans currently do by hand.
The decision belongs to the human. The work around the decision belongs to the system.
What I am describing is not productivity software. Productivity is what you sell when you do not have a thesis. The thesis is that attention is the scarce resource. The existing infrastructure is wasting it. Volume keeps going up. Attention does not. The companies and operators that protect the human's attention will compound. The ones that try to remove the human, or pretend the tax doesn't exist, will not.
I do not know exactly where the AI/human line should sit. When does an algorithmic recommendation become an algorithmic decision? What do counterparties tolerate when they realize they are being routed through a system rather than a person? What is a 90 percent accurate AI worth in a domain where the 10 percent costs you a relationship, a regulator, a child's medical chart?
Open questions. Every company in this space, including mine, is running a different version of the same experiment. Most of us will be wrong about parts of it.
What I am confident about is the direction. The human cannot be the bottleneck. Not because the human should be removed, but because the parts of the work that are actually theirs are not what is slowing them down. The parts that are slowing them down were never theirs in the first place.
I spent fifteen years doing valueless admin. I want it back.
