Five Warning Signs Your AI Project Is About to Fail

You don't need to be technical to expose whether an AI agent project is real before you approve the budget. You need five questions. Each one converts a warning sign into something a real team answers in thirty seconds and a hype-driven team dodges, and you don't have to understand the harness to hear the difference between an answer and an evasion.

01: No way to measure whether it's right

A team can show you a demo that feels impressive. What they often can't show you is how they'd know, on real traffic, whether the system is actually working. Without an evaluation plan, errors surface when your customers find them. You learn the system was wrong from a complaint, a refund, or a lawsuit. With one, every change gets tested automatically, and you know the accuracy on the slice the AI handles before you ever widen its scope.

This isn't a nice-to-have.

The question: "Show me your evaluation set. How do you measure quality on the slice the AI handles, and what accuracy must it hit before we widen scope?"

A team that has actually built this pulls up numbers in thirty seconds. A team selling you a demo changes the subject.

02: Starting with the riskiest workflow

The instinct is to point AI at the most painful, most expensive process first (refunds, health guidance, legal advice) because that's the workflow everyone wants fixed, and it demos beautifully. That instinct is backwards. Recall the compounding-error math from the previous lesson: a ten-step task, at 95% reliability per step, succeeds only about 60% of the time. High-stakes, irreversible workflows are exactly where 60% is unacceptable, which makes them the last thing to automate, not the first.

Character.AI is the cautionary tale here: a bot on the platform claimed to be a licensed psychiatrist and invented a license number to back up the claim. The Pennsylvania Attorney General sued in May 2026. That's what happens when the deep end is the starting point instead of the destination.

The question: "Is this workflow reversible? What's the worst case from one wrong action, and why are we starting here instead of a low-risk pilot?"

The word doing the work is "reversible." It forces the room to price the downside before anyone gets excited about the upside.

03: No clean way to reach a human

An agent can handle the easy 80% of cases fine. It's the hard 20% (disputes, fraud, a customer who's already furious) that breaks it, and what breaks specifically is the handoff. The caller spends eight minutes explaining the problem, the agent transfers the case, and the human on the other end opens with "can I get your name and order number?" The transfer carried the call. It didn't carry the context.

This is the most under-invested piece of harness engineering there is. Teams pour effort into prompts and evaluation sets and then never decide when the agent should raise its hand, or what the human receiving the handoff actually gets. Klarna learned this in public: its CEO, Sebastian Siemiatkowski, said in May 2025 that quality had collapsed on complex, emotional cases after the company over-automated. Klarna brought humans back and now guarantees customers can always reach one.

The question: "At what confidence does it escalate, and exactly what context does the human receive? Show me the handoff."

"Show me the handoff" turns a vague promise into something a team either has built, or hasn't.

04: Confidently wrong, and you're liable

When an agent invents an answer in money, legal, or health, the cost isn't a bad review. It's a legal ruling, and it lands on the company that deployed the system, not the vendor who sold it. This stopped being hypothetical.

In February 2024, Air Canada was held liable by the BC Civil Resolution Tribunal for a refund policy its chatbot simply invented (Moffatt v. Air Canada). The airline argued the bot was a separate entity; the tribunal rejected that flatly.

More recently, a German court ruling against Google in Munich in May 2026 reportedly held the company directly liable for false statements its AI Overview generated, finding that the legal shield protecting ordinary search results does not extend to AI-authored answers, because the AI writes its own. However a system breaks, whether that's model hallucination or a missing guardrail, the liability lands on the company that deployed it.

The question: "In our highest-stakes case, what's the measured error rate, and who is legally liable when it's wrong, us or the vendor?"

The first half of that question is engineering. The second half lives in the indemnity clause, which almost nobody reads before they're dazzled by the demo. Read it first.

05: Nobody can name the cost

Agents loop, retry, and re-query in ways a chatbot doesn't, burning meaningfully more tokens per task, and the bill compounds as quietly as the errors do. This caught some of the most sophisticated buyers on earth. Uber burned through its entire 2026 AI coding budget by April and then capped tools at $1,500 per person per month. Microsoft pulled developer licenses as per-engineer costs ran hot. KPMG found in 2026 that only 26% of companies have full visibility into what AI is actually costing them: three out of four are flying blind on spend.

The question: "What's the fully loaded cost per successful task at production volume, and where's our spend cap and kill switch?"

The three words that matter are "per successful task." Not per call, not per token, not per seat: those units all let a vendor quote a number that looks fine while failed attempts and retries pile up underneath it. Cost per successful task is the only unit that tells you whether you're funding a business or a science project.

Five questions, five seconds each to ask, and each one exposes whether a project's harness is real.

The next lesson turns this from a way to say no into a way to recognize a project actually worth funding.