Blog

The Agency That Hired Itself

Written by Guy Alvarez | Jul 22, 2026 2:52:26 PM

"Write, scribes, so that posterity may learn." Inscribed over a medieval scriptorium at Notre-Dame de Lyre.

In the first days of July, Profound launched an always-on agent called Aim that watches how a brand shows up across ChatGPT and Claude. When it spots a problem, it writes the memo and spins up the project with the tasks already queued (Adweek). Profound is not alone. Another vendor will sell you a full autonomous marketing team: a Director, a Strategist, a Creative, a Producer, a Growth lead, an Analyst, every one of them an AI agent (Soulcraft).

That is the story the market is telling agencies right now. You don't hire a marketing manager anymore. You hire an agent. Manage outcomes, not prompts. One person does the work of four. The agency runs itself while you sleep.

We have told this story before. In the old telling, a rabbi in Prague molds a servant out of river clay and brings it to life by writing a single word on its forehead: emet, truth. The golem never tires and never rests. It also never thinks. It does exactly what it is told, which is the problem, because one Sabbath the rabbi forgets to give it the day's instruction and it runs wild through the streets. He stops it by erasing one letter, turning emet, truth, into met, death.

The golem is the fantasy the autonomy pitch is selling. The clay is strong. The clay is obedient. The clay has no idea what it is doing. The moment no one is watching, its strength becomes the danger. An agency that runs itself is a golem with a client list.

The clay was never that capable

In most agencies the clay does not run wild, because it was never that capable to begin with. MIT's NANDA initiative studied enterprise AI last year and found that 95 percent of pilots produced no measurable return. The reason was not model quality. It was memory. The tools "have no memory," Fortune noted in its writeup: "They don't learn from your feedback, which means you have to feed them the exact same context over and over" (Fortune).

Sit inside a real agency and you see the same thing from the floor. I did recently, with a 200-person PR agency. Three of their people had each built something useful with AI, on their own. None of it was shared, documented, or repeatable. Their data lived in SharePoint and ClickUp, with no structured way for the AI to reach it. Everyone worked in a private chat history. Nobody could use what anyone else had already built. That agency was not unusual. ETCIO reported in 2026 that 62 percent of organizations have no formal process for reviewing AI-generated work at all (ETCIO).

The gains exist, but they are trapped. NANDA also found that only about 40 percent of companies had bought official AI subscriptions, while employees at more than 90 percent were already using personal accounts to get their work done (5WPR). The productivity is happening. It is just happening in private, in a hundred separate chat histories that close when the tab closes. None of it ever becomes the agency's.

Every origin myth needs a monster, and I have given you the obvious one. But the golem is not the interesting figure in this story. The older one never stirred on its own at all.

The room that remembered

Before the printing press, the memory of an entire civilization lived in rooms called scriptoria, where monks copied manuscripts by hand. A skilled scribe managed about seven pages a day. Over the scriptorium at Notre-Dame de Lyre they carved the line at the top of this piece: write, scribes, so that posterity may learn.

What set the scriptorium apart was not any single monk. It was the structure around them. There was a role called the armarius, the keeper of the library, whose entire job was to decide which texts got copied and who copied them. He was, in plain terms, a context router: the person who decided what the institution should remember and who was allowed to touch it. The parchment-preppers, the scribes, the rubricators, the illuminators each did one part, and the work passed between them through a process nobody had to reinvent every morning.

The individual monk died. The abbey's memory did not. The golem is one strong body with no mind. The scriptorium is a modest body of people who together built a mind that outlived every one of them.

What an agency really hires

This is what it means to say an agency hired itself. Not that it bought an autonomous worker. That it built the thing the scriptorium had: a shared memory that makes the institution, rather than any one employee, the thing that gets smarter.

Harvard Business Review argued this year that when every company can call the same models, the model stops being an advantage and becomes a shared utility, like electricity (HBR). What is left is context: the voices, the standards, the approved positioning, the way this particular shop makes its decisions. The model will keep getting smarter. It will never hold your context.

The cost of not holding that memory shows up on a timesheet. McKinsey has put the time knowledge workers spend hunting for information that already exists inside their own organization at close to a fifth of the workweek. IDC found that more than half of employees regularly recreate content that already lives somewhere in the building (source). An agency without a shared brain pays that tax every day, in the most expensive hours it has.

The parts that build the memory are almost embarrassingly unglamorous. Shared context, so the model starts every job already knowing the client's voice. Captured workflows, so a good process runs the same way the second time. Handoff notes, so the work survives the person who started it. None of it demos well. Each captured piece means the next job starts from prior work instead of a blank page.

My cofounder Sally Slater has a name for the disciplined version of this: craft science. The rigorous, tested methods for getting AI to do high-quality creative work. The discipline is doing that without losing what makes it human: the soul, the ideation, the thinking. The AI carries the setup, the formatting, the rote passes, while the person keeps the judgment.

Two problems sit underneath that, and the tools skip both. The first is capturing an organization's real voice and way of working and reproducing it faithfully, for the whole firm rather than the one individual who happens to be good at prompting. The second is harder: marrying organization-wide standards with personalization, so the firm holds one consistent voice and every piece still sounds like the specific person who wrote it. Done well, that is a capability the company owns. It outlasts whoever is best with the tools this year. The autonomy pitch skips all of it and sells you a faster way to produce work nobody can build on.

The handoff note is the clearest case. The same ETCIO survey that found most firms have no review process also found that the ones with a defined handoff between AI output and human review reported strong results 71 percent of the time, against 38 percent for the ones without. Writing down how the work passes from the machine to a person nearly doubled the odds it worked.

So the moat was never autonomy. A golem is easy to buy and, lately, cheap to rent. Anyone can point the same autonomous tool at the same brief and get the same competent, forgettable output back. The moat is the memory underneath it: the accumulated judgment that took years to build and cannot be downloaded.

The ship and its planks

There is an old puzzle about a ship whose planks are replaced one at a time until none of the original wood remains. Is it still the same ship? An agency is always losing planks. People leave, the good ones especially. The question a second brain answers is whether the agency stays itself when they go, because the institution holds the memory of how it works instead of letting that knowledge walk out the door in someone's head.

If every workflow is captured, every voice modeled, every decision written down where the machine reads it first, at what point is it the agency doing the work, and not the people? I do not think that question has a clean answer yet. This origin myth is still being written, in the least glamorous way imaginable: one handoff note, one saved workflow, one line of shared context at a time.

The agency that hires itself does not wake up one morning autonomous. It wakes up one morning with a memory no single person can carry out the door.