The failure mode in most AI deployments isn't that the AI is bad. It's that nobody designed where the human goes.
Last quarter a SaaS company I'd worked with shipped a campaign launch email that referenced a feature they hadn't built yet. The email had been generated by an AI workflow, approved through an automated review chain, and queued in their ESP. By the time anyone caught it, it had gone to 14,000 prospects.
The post-mortem went the way these always go. Someone blamed the AI for hallucinating a feature. Someone blamed the marketer who ran the prompt for not catching it. Someone blamed the reviewer for not flagging it. Someone proposed a more sophisticated AI fact-checker as the fix.
Nobody named the actual problem. The actual problem was that there was no place in the workflow where a human's only job was to read the email and ask "is everything in this true." The AI had a job. The marketer had a job. The reviewer had a job. The CMS had a job. The ESP had a job. Nobody had the human-shaped job. So when the AI made a human-shaped mistake, there was no human there to catch it.
This is what bad AI deployments look like. They're not bad because the AI is bad. They're bad because the team forgot to leave room for the human.
When you take a workflow that used to be all human and add AI to it, something funny happens to the shape of the work.
The human used to do six things in sequence. Now the AI does five of them, and the human does the sixth. Or so the leader thinks. What actually happens is that the AI does five and a half. Because the AI doesn't understand the difference between five and six, and the team's job is no longer well-defined.
The human used to know they owned every step. Now they own the leftover step, which is implicitly defined as "the part the AI didn't do." That's a terrible role definition. It produces ambiguity, drift, and the kind of mistake where everyone in the room genuinely thought someone else had checked.
The leader who doesn't design the human-shaped hole into the new workflow is not deploying AI. They're hoping AI will figure out where the human goes. AI cannot do that. Designing where the human goes is the leader's job.
A human-shaped hole is a place in the workflow where a named person has a specific job that only they can do, and the workflow cannot proceed without them having done it.
It is not "review the AI output before publishing." That's not a hole. That's a wish. People skim AI output. People rubber-stamp. People defer to the model because the model wrote it confidently. "Review" is what bad workflows write down so they can pretend the human is in the loop.
A real human-shaped hole has four properties. Name. Job. Authority. Block.
Name: who specifically owns this. Not "the team." Not "marketing." A name.
Job: what specifically the named person is checking for. Not "errors." Not "issues." Something a human can actually look for, like "is every claim in this email true based on what we shipped this quarter."
Authority: the named person is allowed to stop the workflow without negotiation. They don't have to escalate. They don't have to justify. They can say "this doesn't ship" and that decision is final.
Block: the workflow technically cannot proceed without their explicit yes. Not their absence interpreted as approval. An actual yes.
If your AI workflow doesn't have at least one human-shaped hole defined this way, your AI is going to make a human-shaped mistake at some point, and there's nobody there to catch it.
When teams don't design human-shaped holes deliberately, the things that get squeezed out follow a pattern.
Brand voice is the first to go. Because brand voice can't be checked by an AI fact-checker, and because the human reviewer assumes the AI got it close enough, brand voice drifts in small increments that compound across hundreds of pieces of content. Six months in, the brand sounds like the AI. Nobody can name the moment it happened.
Factual accuracy on the things only an internal human knows is the second. The AI knows what it was trained on. The AI doesn't know that you renamed the Pro tier to Pro Plus last month, or that the partnership with Acme is actually paused, or that the customer story you cited was anonymized for a reason. Only a human knows. If no human is positioned to catch this, it doesn't get caught.
The "should we ship this at all" question is the third. AI doesn't refuse. AI produces what it's prompted to produce. The decision that a campaign shouldn't ship, that an email shouldn't go, that a piece of content was a bad idea in the first place, is a decision only a human can make. If no human owns the kill criteria, nothing gets killed.
The team's morale is the fourth, and the most under-counted. People who used to have ownership of work now don't. The leader thinks they've kept the team in the loop because the team is "reviewing" things. The team knows they don't really own anything anymore, because their judgment can be overridden by speed or by someone above them saying "the AI says it's fine." Roles erode quietly.
A leader deploying AI well is not a leader picking the right tools. The tools matter less than people think. Claude, ChatGPT, custom internal models, they all converge on similar capability over time.
The leader's job is to design where the humans go in the workflow. Specifically: name the holes, define the jobs, give those people the authority to stop the workflow, and build the block into the system so the workflow physically cannot proceed without them.
This is not extra. This is not optional. This is the work of deploying AI well. If a leader is not doing this, they are not deploying AI. They are pushing AI into a workflow and hoping it works out, which is what we used to call cargo-culting.
The human-shaped hole is the deliverable. The AI is the substrate. A leader who confuses these two will produce mediocre work at scale and blame the AI when it goes wrong.
Pick the most important workflow your team runs that involves AI. Map it on a piece of paper. For every step, write down the answer to four questions. Who's named. What's their specific job. Do they have the authority to stop the workflow. Is the workflow blocked without their explicit yes.
For at least one step, all four answers should be yes. That's your human-shaped hole. If no step in the workflow has all four answers, you don't have a workflow. You have a hope.
Then ask the people you named whether they understand they have those four properties. Most of them won't. Telling them, on the record, in writing, that they have name, job, authority, and block is the leadership work. It's not glamorous. It's the only thing that prevents the campaign that ships with a feature you haven't built yet.
Zara
This essay sits inside the Marketing Prompt Hub: 30 tested Claude prompts across 8 marketing disciplines, 6 playbooks, 5 agent blueprints, full reporting kit. Pay once, own the files. No email gate.
See the hub →