Most "AI for business" content lists what AI can do. This one lists what it can't. These are the use cases growing teams keep trying, getting burned on, and then quietly abandoning while the AI vendor still charges them.
1. Customer-facing decisions where being wrong matters
What it looks like: an AI chatbot that handles substantive customer questions without a human gate. "Should I cancel my service?" "Is my procedure covered?" "Will this product work for my situation?"
Why it fails: AI hallucinates. Confidently. The customer asks a substantive question, the AI gives a confident wrong answer, the customer makes a decision based on it, and now you have a service-recovery problem or worse, a legal one.
The right pattern: AI handles routing, scheduling, FAQ-style questions, and routine confirmations. Substantive decisions go to a human. The chatbot is the receptionist, not the doctor.
2. Anything that requires real-time accuracy
What it looks like: using AI to look up current prices, current inventory, current availability, current news, current legal status.
Why it fails: AI's training data has a cutoff date. It doesn't know what happened yesterday. It will confidently give you yesterday's prices, last month's inventory, last quarter's regulations.
The fix isn't "ask AI for current info." The fix is to connect the AI to a real-time data source — your inventory system, a live legal database, today's pricing API. Without that connection, AI is a snapshot of the past.
If a vendor is selling you "AI that knows the current state of the market," ask them where the live data feed is coming from. If they can't name one, the AI is making things up about the present.
3. Tasks where the cost of getting it wrong exceeds the cost of doing it manually
What it looks like: using AI to send marketing emails, post on social, send press releases, or respond to journalists.
Why it fails: a wrong email blast goes to thousands of people. A bad social post is screenshotted. A press release with a hallucinated quote becomes a story. The downside on these is large enough that the time savings don't matter.
The right pattern: AI drafts. A human sends. The drafting saves time. The send-gate prevents disasters. Don't automate the send for anything that goes to more than one person at a time.
The exception: confirmation emails, transactional notifications, and standardized follow-ups where the content is essentially a fill-in-the-blank template. Those are fine to send automatically.
4. Hiring decisions
What it looks like: using AI to screen resumes, score candidates, or filter applicants.
Why it fails: two reasons.
Reason 1: legal exposure. AI hiring tools have shown documented bias by gender, race, and age. The EEOC has issued guidance. Several states have laws requiring disclosure and audit. Your malpractice on this isn't covered by your standard insurance.
Reason 2: it's actually bad at it. AI is very good at pattern-matching to past hires. If you've historically hired a certain demographic, the AI will preferentially recommend that demographic. This is exactly the bias problem.
What's safe: using AI to draft job descriptions, write rejection emails, schedule interviews, take interview notes. The mechanical work around hiring. Not the hiring decision.
The bigger reason to be careful: hiring is one of the few decisions a growing team makes that has 5–10 year consequences. The 30 minutes of resume screening AI saves you isn't worth the 18 months of wrongful-termination exposure if the screening was biased.
5. Anything that requires sustained accountability
What it looks like: making the AI the "owner" of a recurring task, project, or process. "The AI handles our weekly reporting." "The agent is in charge of customer follow-ups."
Why it fails: AI doesn't have accountability the way a person does. When something falls through the cracks, there's no one to ask why, no one to fix the process, no one whose job is on the line.
In practice: when the AI agent owns a workflow without a human owner, drift sets in. Quality slowly degrades. Edge cases accumulate. After 8 months, the agent is doing a half-broken version of the original task and nobody noticed because nobody was responsible.
The fix: every agent should have a human owner. The owner doesn't do the work — the agent does, but the owner reviews accuracy weekly, handles edge cases, and decides when something needs to change. "The agent is the worker, the human is the manager" is the right pattern.
What this looks like in your business
Five tests for whether something belongs to AI:
- Is the cost of being wrong recoverable? If yes, AI candidate. If no, human gate required.
- Does the task need real-time data the AI doesn't have? If yes, either pipe in the live data or don't use AI.
- Is the work going to a single recipient or many? Single recipient is safer for AI. Mass-send needs human review.
- Is there meaningful legal or regulatory exposure? If yes, human in the loop, possibly with disclosure.
- Is there a human owner accountable for the workflow? If yes, AI can do the work. If no, the agent will quietly drift.
If a use case fails any of these, either redesign it or use a human. AI isn't free. Wrongly-applied AI costs more than no AI at all.
What this means for you
The 90% of tasks at a growing team that aren't on this list are fair game for AI. Drafting, summarizing, classifying, routing, scheduling, most of the busywork. Just don't try to automate the things on this list. Other people have. They've burned the money. Learn from their bills.
Get started
Want a real number for your specific situation?
30-minute audit call walks through your workflows and outputs a fixed price for the 2-3 things worth automating first.