TrenchOps 🐎 
The AI Not-To-Do List
Ten Things Your Company Should Never Delegate to a Machine, Ranked by Blast Radius
Classic scenario: an executive pasted an entire acquisition contract into a free public chatbot. He wanted a summary because reading twelve pages was, quote, "not a good use of his time." He got a lovely summary. He also got a confidential M&A document uploaded to a third-party server, which - depending on how you read the NDA he had personally signed - was somewhere between "career-limiting" and "please forward all future correspondence to my lawyer."
The best part: the deal was so straightforward that his own legal counsel could have summarized it in the elevator. He didn't save time. He converted a five-minute conversation into a compliance incident.
Everyone asks what AI can do for their company. Almost nobody asks the more profitable question: what should it never do? So let's invert the problem. If I wanted to guarantee an AI disaster, what would I deploy? Here's the list. Count how many your org is already doing. Two or fewer is excellent. Five or more - well, it was nice knowing you.
The hard no's: where lawyers get involved
1. Feeding confidential or regulated data into unvetted tools. GDPR data, health records, trade secrets, customer contracts. Pasting them into a public chatbot is publishing with extra steps. There's no attorney-client privilege for your chat history - it can be subpoenaed, and courts have already gone fishing in exactly those waters. The chatbot will never sit in the deposition. You will. That's the whole test right there: never delegate a decision to something (or someone) that can't be fired, fined, or sued for it. The AI has no skin in the game. Yours is the only skin available.
2. Drafting or reviewing contracts unsupervised. A contract is precision writing where one wrong clause costs six figures. An LLM is a confidence machine that occasionally decorates its confidence with facts. Lawyers have been sanctioned in real courtrooms for filing briefs with AI-invented case citations - complete with realistic case numbers and fictional judges. The machine didn't lie. Lying requires knowing the truth. It just autocompleted plausibly, which is somehow worse.
3. Final decisions about people. Hiring, firing, promotions, credit, discipline. AI pre-sorting resumes with human review: fine. AI as the last word on someone's livelihood: congratulations, you've automated your discrimination lawsuit intake. Bias in the training data becomes bias at scale, with your company's letterhead on it.
4. Talking to customers without a leash. A chatbot as one support option is fine. A chatbot as the only wall between a furious customer and your brand is how you become a screenshot with 40,000 retweets. A big airline learned this in court after its chatbot cheerfully invented a refund policy, and a tribunal ruled the airline had to honor what its own bot promised. The airline argued the chatbot was "a separate legal entity responsible for its own actions." The tribunal was not amused. Neither was anyone else.
The soft no's: where money quietly leaves the building
5. Tasks where the acceptable error rate is zero. Financial reports, safety documentation, court filings. If every line must be verified by a human anyway, the verification cost eats the efficiency gain and then bills you for dessert. You haven't automated the work. You've added a very eloquent intern who requires full-time adult supervision.
6. Tasks where checking takes longer than doing. Bulk data extraction is the classic trap. The errors are invisible at scale, and you discover them three systems downstream when the quarterly numbers look like modern art.
7. Work where learning IS the point. If your junior engineer has AI summarize every incident postmortem, you get excellent summaries and zero engineers. Fast forward five years: the big outage hits, the AI has never seen it, and the only human who could have handled it spent those years as a prompt courier. You didn't buy productivity. You bought the slow atrophy of your own bench on a subscription model. Meanwhile, your future power horse 🐎 never grew past pony size.
8. Anything a dumb automation solves cheaper. Decision trees, templates, cron jobs. Deterministic, auditable, no hallucinations, no token bill. Using a frontier LLM to select one of ten standard NDAs is like hiring a philosopher to operate a vending machine. He might do it. He might also question whether the customer truly wants the snack.
9. Judging humans from their faces and voices. Some vendors sell AI that scores "trustworthiness" or "leadership potential" from interview videos. This is phrenology with a REST API. Even where it's legal, the reputational blast radius isn't worth whatever it claims to measure - which, spoiler, it doesn't.
10. Anything where authenticity IS the product. The condolence message. The layoff announcement. The personal thank-you to your oldest customer. People can smell machine-generated warmth, and the moment they do, every genuine word you ever wrote becomes suspect retroactively. Some things are expensive precisely because a human spent time on them. Remove the human, and you've removed the product.
The meta-pattern nobody puts on the slide
Here's the insight that ties all ten together: AI doesn't fail like humans do, and your entire review process was designed for human failure.
A tired paralegal makes plausible mistakes - a missed clause, a typo in a date. Your checks were built over decades to catch exactly those. An LLM invents a court case with a realistic citation, a real-sounding judge, and perfect formatting. Your process has no antibody for that because no human ever failed that way. It's not that the machine makes more errors. It makes alien errors, delivered with the confidence of a senior partner.
That's why "human in the loop" isn't a compliance checkbox. It's an engineering requirement - and the human in that loop needs to know what alien failure looks like, not just what human failure looks like.
Notice something about this whole list: it's pure subtraction. Not one item asks you to buy anything, integrate anything, or attend a single vendor webinar. The fastest way to improve your AI strategy is to remove the ten dumbest uses before optimizing anything. Deletion is free. Cleanup is not.
👉 Never delegate to something that can't be fired, fined, or sued. No skin, no game, no decision authority.
👉 If verifying the output costs more than producing it yourself, you've automated nothing.
👉 Protect the tasks where learning happens - or enjoy your engineer-free future.
👉 AI fails in alien ways. Your review process only catches human ways. Close that gap before scaling anything.
👉 Your AI strategy improves faster by deleting bad uses than by adding good ones.
The companies that win with AI won't be the ones who deployed it everywhere. They'll be the ones who knew exactly where not to.
👇 Which of the ten is your company currently doing - and does anyone above you know?
This article is also available in German.
0 comment(s)
No comments yet. Be the first to comment.
Leave a comment