In early September 2026, a California sheriff's office report on a hiker rescue included this line: "The hiking party was advised by Google Gemini to bring far less food and water than they actually needed." Rescuers were dispatched, and the group returned safely.
Reading this story, it's natural to ask, "Why did the AI give wrong information?" But reframe the question slightly and something more interesting comes into view. Did the AI give wrong information — or did it give correct information that simply didn't fit this particular group on this particular day?
These are two different kinds of problems.
How General Guidelines Break Down in Specific Situations
General guidelines for hydration on a hike certainly exist: roughly 0.5 to 1 liter per hour for an adult, more in higher heat, more on steeper grades. This kind of data comes from a large body of research and experience, and it's almost certainly well represented in the text an AI tool was trained on.
The problem is that this guideline doesn't automatically apply to a specific group, on a specific day, on a specific trail. What was the actual temperature that day? How steep was the route? Who was the least fit person in the group? None of these variables live inside a general guideline.
An AI planning tool can't close that gap on its own. What the tool has access to is patterns recorded in text, and most of those patterns describe general situations. Unless a user explicitly enters the specific variables, the tool produces an answer that starts from the general pattern.
The answer isn't wrong. It just may not fit your situation.
The Same Gap Shows Up in Business Planning
What happened with the hiking plan is structurally the same thing that happens in business planning.
A solo founder about to launch a new service asks an AI planning tool, "How much should I budget for initial marketing?" The tool answers: "In the early stage, it's typical to put 10 to 20 percent of monthly revenue toward marketing." As a pattern drawn from data on many businesses, that's correct.
But for this answer to fit this particular founder, several things need to sit near the industry average: the cost of acquiring one customer needs to be close to the industry norm, the repeat-purchase rate needs to be similar, and the competitive landscape needs to resemble the cases already on record. If even one of those conditions is far off, "10 to 20 percent" is a starting point, not a plan.
Project timeline estimates work the same way. Ask "How long does it take to build a service of this size?" and the tool offers a range drawn from past cases of similar scale. That range has nothing to do with your team's actual pace, the other work you're juggling right now, or how many people are involved in decisions. The trouble starts the moment that range becomes your actual schedule.
The same holds for the "three to six months" answer to "How long does it take to open a café?" That range only holds if variables like prior restaurant experience versus none, doing the interior yourself versus outsourcing it, and where the lease negotiation stands are all close to "typical."
Why AI Output Looks Precise
Part of why this structural gap gets glossed over so often is the format of the output itself.
Answers from AI planning tools tend to be numbered, broken into distinct steps, and stated with clean figures. "Step 1: Set the budget. Start with initial marketing spend at 15 percent of monthly revenue." That format reads like a plan. It looks like the product of deep analysis.
But the format isn't precision — it's just a general pattern dressed up neatly. That 15 percent looks specific, but it means "many businesses' averages fall in this range," not "15 percent is right for your business."
Many AI tools do include a disclaimer: "This may vary depending on individual circumstances." But that line sits at the top or bottom of the answer, while the numbered, figure-laden body in the middle is what actually gets taken as the plan.
The gap between the impression the format creates and the actual nature of the content — that's where execution goes wrong.
One Question: "What Would Change This Answer?"
One approach is to ask this right after receiving a recommendation: "What conditions would change this recommendation?"
AI usually answers this one honestly: "If your customer acquisition cost is higher than the industry average, you should raise the marketing budget percentage. If the team is working with an unfamiliar tech stack, plan for 1.5 to 2 times the timeline. If your industry has a low repeat-purchase rate, you should plan for a longer initial runway."
That answer surfaces the variables specific to your situation: What is your actual customer acquisition cost? Is this new technology for the team? Does your business generate repeat purchases? The next step is to feed those variables back into the tool explicitly, or, if you don't have the data, to go measure it yourself.
Apply this to the hiking plan: asking "What conditions would change this hydration recommendation?" produces an answer like "If the temperature is above 30°C (86°F) or the route includes steep sections, increase the recommended amount by at least 1.5 times." That's the moment the variables you actually need to check — the day's weather forecast and the route details — come into view.
Skip this question and use the first output as-is, and you're applying a number generated for average conditions to a situation that isn't average.
AI planning tools are useful for surfacing general patterns quickly. Plugging your own situation's variables into that pattern is a step the tool can't take — a person has to take it, both identifying what those variables are and finding their actual values. In an age of smart tools, using them well isn't about pulling more output out of them; it's about knowing where to stop and check that output.
When the AI says, "This should be enough," ask what situation it's assuming. That question is what closes the distance between output and plan.
