I have seen too many versions of the same meeting recently.

Someone says the company needs to “go all in on AI.” A vendor shows a polished agent demo. A manager imagines every employee suddenly having three tireless digital coworkers. The deck looks great. The numbers look even better.

Then the work hits reality.

The data is messy. The process was never clearly defined. Permissions are vague. Nobody knows who verifies the output. The agent produces something plausible, someone trusts it too early, and now the team is spending more time cleaning up after AI than doing the work itself.

So when a Hacker News thread titled “I believe there are entire companies right now under AI psychosis” reached 1,870 points and 1,055 comments, it did not surprise me at all.

“AI psychosis” is a harsh phrase. I do not love the medical framing, but I understand why it resonated. It points to something real: some companies are no longer evaluating AI as a tool. They are being pulled around by the fantasy of what AI might do.

My view is blunt: AI is useful, but collective AI intoxication inside a company can be more dangerous than not using AI at all.

Not using AI may make you slower. Misusing AI can break workflows, blur responsibility, and make bad decisions travel faster.

What AI intoxication looks like

The first symptom is treating demos as products.

In a demo, an agent reads email, writes a proposal, modifies code, creates a report, and looks magical. The demo world is clean. The input is clear, the permissions are generous, the edge cases are invisible, and nobody is accountable if the output is wrong.

Real companies are not like that.

Customer data is incomplete. Contracts have exceptions. Old code has hidden assumptions. Internal documentation is stale. A person working in that environment at least has some sense of what they are risking. AI makes mistakes differently. The dangerous part is not that it can be wrong. The dangerous part is that it can be wrong in a very convincing tone.

The second symptom is confusing generation with responsibility.

AI can generate a polished market analysis. It can summarize a meeting into action items. It can draft a customer email in a calm, professional voice. But generation is not judgment. It is definitely not accountability.

This is easy to forget because AI output often looks complete. It feels like thinking has already happened. Most of the time, what has happened is expression.

Expression is cheaper now. Decision-making is not.

The third symptom is trying to turn every unresolved process into an agent project.

Recruiting needs an agent. Sales needs an agent. Support needs an agent. Engineering needs agents too. On paper, it sounds like a productivity revolution. In practice, it often means handing an already broken process to a system that is harder to debug.

A bad workflow plus AI does not become a good workflow.

It becomes a faster bad workflow.

Why companies are so vulnerable to this

There are three forces behind the AI rush.

The first is anxiety.

AI is moving fast enough that many leaders feel they must visibly do something. That fear distorts normal technical evaluation into a signaling contest. Whoever announces “AI transformation” first looks more future-facing, even if the actual plan is thin.

The second is metric temptation.

AI is easy to put into a KPI slide. Response time went down. Automation rate went up. Meeting summaries saved hours. Code output increased. These numbers are attractive because they are easy to report.

The hard numbers are less flattering: who checks the errors? Who owns damage from a bad customer reply? How much time do employees spend cleaning data, writing prompts, reviewing output, and fixing subtle mistakes?

If you count the time AI appears to save but ignore the human time spent verifying it, you are just lying to yourself with better charts.

The third is that the vendor story is too smooth.

“Give every employee an AI assistant” is a great story. “Use agents to reinvent your business process” is also a great story. The problem with great stories is that they make the boring details sound like negativity.

Permissions, audit logs, rollback, data boundaries, human review, failure plans. None of this looks exciting in a demo. All of it determines whether AI becomes a tool or an incident multiplier.

An agent is not magic. It is a new coworker.

I prefer to think of agents as new coworkers rather than magic buttons.

This coworker is unusual. It reads quickly, writes quickly, handles repetitive work, and does not mind running all day. But it has obvious weaknesses: it does not reliably know when it is wrong, it does not automatically understand your organizational context, it can mistake similarity for correctness, and it cannot be accountable for consequences.

So the right question is not only: can it do the task?

You also need to ask:

  • What happens when it is wrong?
  • Who can detect the mistake?
  • Can we roll back the action?
  • What data did it use?
  • Was it allowed to see that data?
  • Does a human approve the output before it enters the real workflow?

These questions are not sexy. But the teams that ask them early will outperform the teams that only obsess over prompts.

A practical checklist before adopting agents

If a small team wants to introduce AI agents, I would start with this checklist.

1. Begin with low-risk, high-repetition work

Do not start with contracts, finance, medical advice, key customer communication, production deployments, or anything that can cause serious damage quickly.

Better starting points include research collection, meeting notes, internal knowledge-base search, low-risk code refactoring, customer-support drafts, and competitor monitoring. If the AI makes a mistake, a human can usually catch and correct it.

AI’s first job should be assistant, not boss.

2. Budget for verification

Many AI projects fail not because the model is too expensive, but because verification is too expensive.

If AI saves two hours drafting a report but the team spends three hours checking the data, citations, assumptions, and conclusions, the automation did not save time. It created a prettier version of the same work.

Every AI workflow should answer one question early: how expensive is it to verify the output?

If verification costs almost as much as doing the work manually, that task is probably not ready for automation.

3. Keep human checkpoints

The riskiest step is not generating text. It is taking action.

Reading, drafting, summarizing, and suggesting can be relatively safe. Sending emails, changing prices, deleting data, placing orders, or deploying code should move more slowly.

A simple rule works well: AI accelerates preparation; humans own final action.

Once the workflow has proven stable, you can gradually give it more autonomy. Do not hand over the keys on day one.

4. Log what the agent does

If nobody can review what an agent did, it should not be in production.

What did it read? Which tools did it call? What did it generate? Who approved it? When did it execute? These records matter. Without them, every incident becomes guesswork, and guesswork quickly becomes blame-shifting.

Logs are not bureaucracy.

For agents, logs are seat belts.

5. Let the team say “not worth it”

This may be the most underrated rule.

Not every process deserves AI. Some tasks take ten minutes manually and three days to automate badly. Some workflows become more fragile after an agent is inserted. Some work is simply not repetitive or valuable enough to justify the risk.

A mature AI culture is not “use AI everywhere.”

It is knowing where AI helps and where it should stay out.

My real position on the AI boom

I am not anti-AI. Quite the opposite. I think agents will reshape a lot of work.

You can already see the infrastructure forming. GitHub’s monthly trending list is full of skills, memory, and code-context projects. Product Hunt has tools such as Agentmemory getting attention. The direction is real: agents are becoming less like chatbots and more like work infrastructure.

That is exactly why we should be careful.

The stronger the tool, the more damage a mistake can cause. One employee writing a bad email is a contained problem. An agent connected to CRM, email, internal docs, and ticketing systems can make a mess at a much larger scale.

Good AI adoption should not feel like a religious movement. It should feel like engineering: small pilots, clear boundaries, serious acceptance criteria, logging, review, and iteration.

That sounds less exciting than “AI transformation.”

I trust it much more.

The point: be excited, but stay sane

The most fascinating thing about this AI wave is that it really does make previously impossible workflows possible. A small team can now build automation that used to require a much larger company. An individual can connect research, writing, coding, design, and operations into a personal workflow.

Excitement is reasonable.

But excitement cannot replace judgment.

If a company starts treating “we use AI” as the answer, instead of asking what problem AI solves, what risk it introduces, and who owns the outcome, it is already drifting into dangerous territory.

My advice is boring and useful: treat AI as a powerful new coworker that needs supervision. Give it tasks. Give it boundaries. Give it logs. Give it review. Do not worship it, and do not demonize it.

Teams that can do that are using AI.

Everyone else may just be getting high on it.

References