On the morning of August 1, 2012, a trading firm called Knight Capital switched on new software. One of its eight servers was still running a dormant piece of code from 2003. In the 45 minutes it took to find and stop the problem, the system pushed more than 4 million unintended trades into the market and lost the firm around $440 million. Knight was sold to a competitor within months.

No AI was involved. That was ordinary automation. But the shape of the disaster is the thing to memorize, because it’s the exact shape AI agents inherit: software that acts, at machine speed, with no human in the loop, in a place where actions can’t be taken back.

An AI agent is software that doesn’t just answer questions the way ChatGPT does. It does things. It reads your inbox, fills in forms, moves records between systems, approves or rejects requests, and decides its own next step along the way. That’s the appeal. It’s also the risk, because a wrong answer on a screen costs you nothing, and a wrong action in your billing system costs you real money.

Right now nearly every software vendor is pitching agents. So it’s worth sitting with one number. Gartner predicts that more than 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and risk controls that never quite materialize. The cautionary piece almost nobody writes is the one about when to say no. Here are the five situations where no is the right answer, and a short test you can run before committing budget.

When the action can’t be undone

Some actions have an undo button. A draft email, a suggested price change, a proposed schedule. And some don’t. A wire transfer that left the building. A production database row that got deleted. An email that landed in 40,000 customer inboxes with the wrong discount code.

The rule is blunt: an agent should never be the last pair of hands on an irreversible action. There has to be a human gate, a checkpoint where a person approves before the action fires. Banks figured this out decades ago with the four-eyes principle, where no single employee can move large sums alone. Your agent deserves less trust than that employee, not more.

Here’s the catch that kills many projects. If the approval gate takes as long as doing the task by hand, the agent saved you nothing. When that happens, it isn’t a sign to remove the gate. It’s a sign the task was never a good fit.

When a 2 percent error rate quietly compounds

An agent that gets each step right 98 percent of the time sounds excellent. It isn’t, once the steps chain together. The math is unforgiving: 0.98 multiplied by itself 20 times is about 0.67. So in a 20-step workflow, roughly one in three runs contains at least one error. Stretch it to 50 steps and nearly two out of three runs are contaminated.

Think of a photocopy of a photocopy. Each copy looks almost identical to the last, and the tenth is unreadable.

This isn’t hypothetical. Researchers at Carnegie Mellon built a simulated software company staffed entirely by AI agents and gave them 175 realistic office tasks, from writing code to filling out HR forms. The best-performing agent completed about 30 percent of them. Not 30 percent badly. Thirty percent at all. Agents did noticeably better on programming tasks than on administrative ones, which is roughly the opposite of what most companies want to automate first.

Long multi-step workflows also fail in a sneakier way: one agent’s small mistake becomes the next step’s input, and the error propagates like an infection. That’s why the teams doing this seriously build monitoring that works like an immune system, catching bad outputs before they spread. If you’re not prepared to build that layer, don’t deploy the long workflow.

When the agent costs more than the person

Agents bill by the token (the small chunks of text a model reads and writes, which is how AI providers charge). A single agent task is rarely a single model call. It’s a loop. The agent reads context, reasons, tries something, checks the result, retries, and each lap around the loop costs money. Then add the part almost everyone forgets to price in: the salary of the person reviewing the agent’s output.

Now put that against a thin-margin process. Say your operation clears 30 cents of margin per processed claim, per verified listing, per routed ticket. If the agent’s fully loaded cost per task, retries and review included, comes to 45 cents, you’ve automated your way into losing money on every unit while the dashboard shows a green “success” rate. Volume makes it worse, not better.

The fix is boring and effective. Before scaling anything, compute cost per completed task, not cost per API call, and compare it against the human doing the same work. The broader arithmetic of what it costs to build an AI product in 2026 follows the same pattern: the model calls are rarely the expensive part.

When a script or a form would win

Here’s a sentence that will save you real money: if you can write the rules down completely, write them down as code.

A script is a fixed set of instructions a computer follows exactly, the same way every time. It costs a fraction of a cent per run, never improvises, and when it breaks, it breaks loudly and identically until fixed. An agent, by contrast, is paying a very smart generalist to rediscover your rules on every single run, with some chance it rediscovers them slightly wrong.

Password resets. Invoice matching against a purchase order number. Renaming 10,000 files. Copying fields from one system into another when the fields always map the same way. None of these need judgment, so none of them need an agent.

The same logic applies to collecting information. A web form with four required fields will outperform a conversational agent at intake almost every time, because the form cannot forget to ask the third question. Agents earn their cost only where the input is genuinely messy and the path genuinely varies. Everywhere else they’re an expensive way to add randomness to a solved problem.

When you have to explain the decision

In February 2024, a tribunal in British Columbia ordered Air Canada to compensate a passenger after its website chatbot invented a bereavement refund policy that didn’t exist. The airline argued the chatbot was responsible for its own words. The tribunal disagreed and awarded the passenger roughly 650 Canadian dollars in damages plus interest and fees. Small money. Enormous precedent: you own what your software does.

Now scale that up to decisions that are actually regulated. Credit approvals, insurance claims, hiring screens, medical triage. Europe’s GDPR, in Article 22, gives people the right not to be subject to purely automated decisions with significant effects, and the EU AI Act layers documentation duties on top for high-risk uses. When a regulator or a rejected applicant asks “why was this decision made,” the answer “the agent decided” is not an answer. It’s an admission.

Language models are poor at reconstructing their own reasoning after the fact. So in compliance-heavy contexts, the workable pattern is narrow: agents that gather and check evidence against written rules, with a person making the final call. Stripe’s approach to compliance agents shows how far you can get with that discipline, and where the line sits.

The four-question test

Before any agent project gets budget, run the task through four questions.

  • Reversible? Can every action be undone, or gated by a human before it fires?
  • Explainable? Could you defend the outcome to a customer, an auditor, or a judge?
  • Cheaper? Is the fully loaded cost per completed task, retries and human review included, below the cost of a person or a plain script doing the same work?
  • Measured? Have you run a pilot on real volume and observed an error rate you can live with once it compounds across the full workflow, rather than assuming one?

Four yeses, proceed. Any no means fix that dimension first or pick a different tool, and “a different tool” will often be a 200-line script that never asks for a raise.

None of this is an argument against agents. They’re genuinely good at drafting, at research, at triaging the messy front end of a process before a human takes over. The argument is against defaults. Gartner’s 40 percent won’t fail because the technology was fake. Most will fail because someone pointed a probabilistic tool at a task that needed certainty, and nobody asked the four questions while it was still cheap to ask them.