A head of marketing I know sent me a deck a few days ago. Slides, copy, layout, narrative arc. Designed and written end to end by Claude. He hadn't opened a single slide before forwarding it. He told me proudly, "I didn't touch a single thing."
It's a pattern now. Inboxes are filling with emails written by Claude and not read by their senders. Decks are getting shipped that nobody glanced at. The senior people are doing it too.
The Second Wall
The companies that adopted AI fastest are running into a second wall. The first one was getting AI working in the building. The second is everything that breaks once it's in there and nobody is checking.
Marc Benioff said something on the All In podcast that didn't fully land at the time and won't leave my head now. "None of this stuff works if you don't have context."
Satya Nadella made a related point this week. "You can offload a task, or even a job, but you can never offload your learning."
Two of the most powerful software executives on the planet, pointing at the same gap. The idea applies to a deck a head of marketing ships without reading. It applies to a prompt passed around a Slack channel until it reaches a client. It applies to every place AI now sits where humans used to add context, and don't anymore.
Three Things We Keep Seeing
Three patterns are showing up at the companies that adopted AI fastest. We're seeing each of them play out in our own client work. All three are the same problem wearing different clothes. Context, the thing that turned the model from generic into useful, gets stripped out somewhere along the way, and nobody puts it back in.
The first pattern is the email phenomenon, scaled. Someone uses Claude or Copilot to draft a message. They glance at it. They send it. The receiver opens it and immediately senses something is off. They spend the time the sender saved trying to decode what the sender actually meant.
When a chef serves a dish he didn't taste, you call it disappointing. When he didn't even cook it, just pulled a freezer pack and warmed it up, you call it something else. That's the energy a lot of internal AI output now carries. Microwaved. It looks like food. The chef's hand isn't in it anywhere.
The asymmetry runs in one direction. The sender saves ten minutes. The receiver loses fifteen. Multiply across every message in the company. Then across every message between companies. The time the sender saved becomes the time the receiver loses, with interest.
The second pattern lives inside the company. Employees are building their own prompts, sharing them in Slack channels, and pushing the outputs toward client work without anyone reviewing what those prompts actually do. A recent conversation with the CEO of one of our clients made it concrete. Every team in his organization has its own way of using Claude. Nobody has standardized. Nobody is gatekeeping. Prompts are passed around like memes. Some of them are excellent. Some of them are garbage. All of them are reaching client deliverables.
A law firm would never treat a contract template this way. The template gets written by a partner, reviewed, version-controlled. It becomes a firm-wide artifact. A prompt that touches client work has the same blast radius. The output goes out on firm letterhead, metaphorically. Yet inside most companies right now, the prompt drafting the client email was crafted by whoever felt confident enough to share it that morning, with no review.
The third pattern is the quietest and the most dangerous. Employees with ten years of expertise are starting to defer to Claude. The reason is simple. Claude sounds confident. A tired human, after a hundred meetings telling them to use AI more, lean on it harder, justify their salary against its output, sounds less and less so. Quality is high enough that the model feels more impressive than the person who's been doing the job for a decade.
The same CEO has been watching it happen across his organization. Someone with deep tacit knowledge starts second-guessing themselves because the model produced something that looked smart and assertive. The model didn't have a fraction of the context the person carries. It can't know that the pricing in this proposal is the exact number that blew up the last negotiation, that this contract has a history with a counterparty going back three years, that this deliverable goes to a board member who reads every word.
The model sounds just as confident when it's wrong as when it's right. The operator has texture. The texture is what gets stripped out when the operator steps aside. The task still gets done. The learning behind it walks away.
What Governance Actually Means Here
All three patterns trace to the same root. Companies bolted AI onto their workflows faster than they built the discipline that comes with it. Feeding the model context. Validating what it produced. Deciding where humans still own the call.
That discipline has a name at company scale. Governance. The internal, operational kind. The version that decides which prompts can touch clients. The version that draws the line between where Claude leads and where the human still owns the work. The version that says your name on it means you read it. Different beast from the regulatory governance the conferences spend their time on.
It comes down to a few things. Approval before a prompt touches client work. Spend controls. A clear line on where the model leads and where the human still owns the call. Accountability for what actually ships.
Then the Bill Arrives
Governance has been an academic conversation until now. That changes when the bill arrives. Marty Kausas posted recently that his company's Anthropic spend jumps from $400,000 to $1.4 million annually the moment they cross 150 seats. Engineering teams produced clear ROI. Non-technical roles with low adoption did not. Same bill, very different value per dollar.
Token-maxxing, letting employees use the model freely because nobody is watching, is what happens when nobody owns the spend. It's also what takes a bill from $400,000 to $1.4 million.
The Best Governance Is the One You Do Yourself
There's a temptation to dump work into Claude and pocket the saved time. We've all hoped the rabbit will come out of the hat, and the hope hasn't aged well.
The people leaning on the model as a shortcut are doing something specific to themselves. They've stopped reviewing. They've stopped editing. They've stopped adding the context that made them valuable in the first place. Their job shifted from doer to validator, and they read it as doer to spectator. A spectator has already eliminated their own job and forgotten to notice.
That's the part no policy document fixes. A model is potent in both directions, for good and for bad. What decides which one you get is whether the person at the keyboard knows how to use it well. Context about them and the company. A tight, well-crafted set of instructions, call it a prompt. And the understanding that the job moved from doer to validator, not from doer to spectator.
The rails still matter. Approval, spend controls, a clear line on who owns what. But rails don't validate the work. People do. The best governance at company scale is a building full of people who govern themselves.
Tools are starting to ship for this, shared context layers that put the same prompts in everyone's hands. They solve distribution. They don't solve judgment. You can install the library. You can't install the reader. That part has to be taught, which is why every engagement we run starts with discovery and training, before anything gets automated.
My bet: the companies that build that first will keep their teams sharp. The rest are about to find out what it feels like to run a business from a row of microwaves. Nobody at the stove. Looks busy though.
This is what we do at Exponential Partners. We help leadership teams get AI into their companies without ending up with a kitchen full of microwaves. If your company is hitting the second wall right now, let's talk.
