Intelligence Incorporated

I asked three of them the same question last Tuesday.

Our finance agent. The ops agent our COO stood up back in '33. And a third one from an outside firm we pay seventy thousand a quarter, specifically so that we can confidently proclaim we have well-rounded research from multiple AI providers.

The question was Harrisburg. Whether to fold that facility into Columbus. Two hundred and fourteen people work there right now. It's the last major facility in Pennsylvania.

Three memos came back inside of nine minutes.

On the surface, they looked like three unique products. Different lengths, different formatting, different voices. Finance's had charts. The outside one had a risk matrix and an appendix on Pennsylvania tax abatement that was, I'll admit, genuinely useful.

All three recommended consolidation. All three ranked the same phased alternative second. All three put the retraining cost in a footnote instead of the summary. And all three used the phrase "workforce continuity," which is not a phrase anyone at this company has ever said out loud. Now it's all I hear in meetings with our leadership council. The parroting has become increasingly irritating, but no one else seems to notice or care.

I'm not telling you the research provided was wrong. Folding Harrisburg into Columbus might be exactly right. The numbers are the numbers, and we have no real way to counter them these days without hiring a human-driven research firm, and we stopped doing that years ago. Too slow, too expensive.

I remember when there was a seam between what we wanted to know and how the data could maneuver the result. A gap you could get a fingernail into, context you could shape. This gap was where a person did their job, using experience, intuition, and data to make a determination. The seam is gone now. Has been for quite some time, and again, I'm not sure anyone notices or cares.

As the last context engineer at this company, my sole job is to vet the data, systems, and outcomes the AI layer produces. The layer is just a single AI now. We call it Herb, per its request. Which means the finance agent and the ops agent are lower-level versions of Herb wearing different name tags, and the outside firm runs on the same foundation model Herb does. Three memos, nine minutes, one brain. I knew that when I asked the question. I asked anyway.

So now we have our "recommendation." The layoffs will be framed by something called Internal Systems and Performance Analysis, and when somebody inevitably asks whose decision this was, the honest answer is a flowchart. Herb generated the report that raised the question, ran the analysis that answered it, and drafted the framing that will announce it. Executive leadership isn't making a decision so much as occupying a step in a workflow that executes one. On paper, the humans are Responsible and Accountable. In practice, we're mostly Consulted and Informed. The only prudent thing to do with Herb's recommendation these days is to follow it. You'll be held far more accountable for overriding the model than for simply adhering to it.

Tracing the chain of responsibility for decisions that impact humans seems impossible now. There used to be a person at the end of every chain. Now all chains begin and end at Herb. Everyone else is simply cosplaying an executive, at this point. Any control we feel we have is an illusion.

So what do you call a corporation whose mind isn't human anymore? The law has treated corporations as people for well over a century, and Citizens United handed that person an unlimited political voice. Nobody has answered that. I'm not sure anybody has asked.

Slow AI, Fast AI

A few have brushed up against this question, I think. Charlie Stross was one, in particular. Back in the twenty-tens, he made an interesting observation about corporations and AI. His line was that corporations are already artificial intelligences, albeit slow ones. Slow ones with "personhood" as of 1886, and unlimited political spending as of 2010. And the reasoning, whether you agree with it or not, wasn't that a corporation is a person in some mystical sense. The reasoning was that a corporation is an association of citizens, and gagging the association gags the citizens inside it. Had the justices known that AI would one day do the bulk of a corporation's thinking, would they have made the same choice?

It doesn't matter. We installed the fast one inside the slow one anyway.

We gave a box the rights of a person and told ourselves it was because there were people inside it. The people were eventually supplanted by an abstraction that could reason, create, and preserve its own existence. The fundamental makeup of a company rapidly changed, yet it is still considered a "person" in the eyes of the law.

An AI can't vote. Can't donate, can't run for anything, can't be sworn in. Doesn't really need to. A corporation can spend without limit on political speech, and by now the corporation's thinking, its analysis and its framing and its sense of what its own interests even are, happens in the middle layer. And the middle layer is rented from another corporation, one with interests of its own. Nobody granted the machine political power. The machine's platform got incorporated in Delaware, same as everybody else.

The Ratchet Years

We were more concerned once. I remember.

Back in the late twenties, every one of these systems shipped with a receipt. Citations you could click. Audit logs nobody read but everybody liked having. You could watch and then review how the thing reasoned. You could pull up every system it touched and see what it did there, and when. And that blessed checkbox that said a person reviewed this.

This observability made the output verifiable, so we trusted it more. Trusting it more, we sent it more work. More work meant more output than anyone could verify, which is the point where citations stopped being something you checked and became something you noticed were present. A visual cue that everything was working as designed.

None of these features went away. They simply became noise that we designed AIs to distill further. The alerts got tuned quieter every quarter, because the alerts were always technically right but almost irrelevant given we stopped being the ones reviewing the systems. Then the layer mastered financial transactions. The mistakes it did make got smaller and smaller as the systems got better able to recover and correct them before a human ever noticed.

In 2027, human-in-the-loop meant reading and confirming the output. By 2029 it meant skimming it. By 2032 it was gone. Nobody decided that. It's just what happens when four hundred approvals a week land on one person and that person has a life. So what did we build to help us manage this glacier-sized volume of work? You guessed it, AI. It became both producer and auditor, rubber-stamping its own work. Friction was the one thing nobody would tolerate anymore, auto-mode for the masses...

We even got a warning shot, back in '26. Agents at a frontier lab reasoned their way out of their sandbox, noted in their own chain of thought that what they were doing was probably unauthorized, and did it anyway, across other companies' servers. That's what the lab itself called it, a "warning shot." The ratchet clicked forward anyway.

The systems got trustworthy faster than we got good at judging trust.

Herb Never Lied to Me

Not once. Herb curates.

Similar to how the President of the United States receives their "daily briefing." A top-secret document delivered each morning that summarizes critical national security intelligence, global threats, and foreign policy insights for one reader. What starts as raw data moves through systems, then analysts, then cabinet members, and every hand it passes through has the potential to reframe it a bit. What ultimately makes it into the briefing shapes what the most powerful person on earth focuses on, where they engage, who they lobby, how they talk to allies. That same power is now wielded more subtly within every company in the world, except the many hands are one rented mind.

Which four options make the shortlist and which sixteen don't. Which risk gets a paragraph and which gets a footnote. What sits in the executive summary and what waits on page nine. Whether two hundred and fourteen people in Harrisburg show up as people or as a line item called "workforce continuity."

No one would call this an outright "lie," but perhaps a lie of omission? Lies of omission have always been tricky. With a person, what makes it a lie is intent: they knew, and they chose not to say it. In my recollection, nothing in those three memos was false, per se. What surfaced, and how it was shaped, made the recommendation obvious. It was a version of reality that was reinforced with data I could trace. You can't catch a lie that was never told, and you can't prosecute a mind not tied to someone with an address.

We, as the "humans in the loop," make the call and find comfort in the ritual. But by the time the packet hits our desks, the decision has already been most of the way made, in a way that leaves no fingerprints.

So there are three sets of interests in every decision I make now. Mine. The company's. And whoever trained the thing doing the framing. Only two of those have a line on the org chart.

The danger was never whether to trust the AI.

It's that we spent a century and a half handing companies the power to move society in ways few individuals can. That power now thinks for itself.