A consultant I spoke with had started drafting her proposals with AI and was pleased with the speed, right up to the afternoon she noticed a statistic in one of them that did not exist. It was formatted beautifully, attributed to a real-sounding institute, and completely invented. She caught it on the third read: the tool had not saved her time, it had moved the time to the end, where it was more stressful. This is not a story about a careless founder. It is the ordinary behavior of the technology. In September 2025 OpenAI published research describing hallucinations as "plausible but false statements" and arguing that language models produce them because "standard training and evaluation procedures reward guessing over acknowledging uncertainty." The model is built to give you an answer, not to tell you when it does not have one.
This article is about working with that fact instead of being surprised by it. I will explain why AI gets things wrong in the first place, the four kinds of mistakes you should actually expect, why "just check everything" quietly turns you into a full-time editor, how to set AI up so it makes fewer mistakes before you review anything, a three-tier review routine that fits into a founder's week, and where AI must never be the last pair of eyes. By the end you will have a review system you can write down on one page and hand to anyone who works with you.
Why does AI get things wrong in the first place?
The short answer: because it is trained and graded to produce a confident answer, not to admit that it does not know. That is a design consequence, not a bug that the next version will simply remove.
The OpenAI research uses a school analogy that I find clarifying. On a multiple-choice test, a wild guess might be right; leaving the question blank guarantees zero. If a model is asked for someone's birthday and guesses "September 10," it has a 1-in-365 chance of scoring; saying "I don't know" scores nothing. Over thousands of test questions, the guessing model looks better on the leaderboard. The systems we use every day were shaped by scoreboards that penalized humility. The paper gives a concrete comparison from OpenAI's own evaluations: an older model abstained on 1% of questions and gave wrong answers 75% of the time, while a newer model abstained on 52% and was wrong 26% of the time. Their accuracy was almost identical, 24% against 22%. The difference was entirely in how often each one admitted uncertainty.
The second cause is how these models learn. They are trained to predict the next word across enormous amounts of text, and, as the same research notes, consistent patterns like spelling get learned reliably, while arbitrary, low-frequency facts cannot be predicted from patterns at all. A model can be flawless at grammar and hopeless at your competitor's founding year, for the same reason. The practical rule I give founders: the more specific and the less commonly repeated a fact is, the more suspicious you should be of it. A definition of coaching will usually be fine. The exact percentage from a 2024 report, the name of a journalist, the date of a rule change: that is where the model fills a gap with something plausible.
Which kinds of mistakes should you actually expect?
The short answer: four kinds: invented facts, reasoning that sounds right but is not, drift away from your instructions, and a confident tone that hides the other three. Once you can name them, you can check for them separately, which is faster than a general sense of unease.
The confidence problem is the one that costs founders the most. In March 2025 the Tow Center for Digital Journalism at Columbia Journalism Review ran 1,600 queries across eight AI search tools, asking each to identify the source of a real article excerpt. Collectively they answered more than 60 percent of the queries incorrectly, from 37 percent for the best performer to 94 percent for the worst. What struck me more than the error rate was the manner: the tools rarely used qualifying language, and one of them incorrectly identified 134 articles while signaling a lack of confidence just fifteen times out of two hundred responses, and never declined to answer. Premium versions were more confidently incorrect than the free ones, because they were more willing to give a definitive answer at all. Price does not buy you honesty about uncertainty.
So here is where errors cluster in the work a service business produces. Names. Dates and deadlines. Statistics. Direct quotes. URLs and email addresses. Prices, especially your own, if the model was working from an old document. Anything legal or tax-related. And one more that founders miss: drift. You asked for a summary of a client's intake form and the model added a sensible-sounding goal she never mentioned. Nothing is false about the world, but it is false about her. The check for drift is simple: compare the output to the source, not to your memory of the source.
Why does "just check everything" turn you into a full-time editor?
The short answer: because unstructured review costs more than the drafting saved. Reading every line with equal suspicion is slow, exhausting, and still misses things.
There is now a name for AI output passed along without that review. Researchers from BetterUp Labs and the Stanford Social Media Lab called it "workslop" in a September 2025 piece for Harvard Business Review: AI-generated content that looks good but lacks the substance to move a task forward. BetterUp's survey of 1,150 full-time U.S. desk workers found that 40% had received workslop in the previous month, that each incident took about two hours to resolve, and that the cost worked out to roughly $186 per employee per month. In a company, that cost lands on a colleague. In a business of one to ten people, you are usually the sender and the receiver, so you pay it twice: once to generate the draft and once to untangle it, often at nine in the evening.
This is where the difference between a tool and a system matters. A tool needs you at the keyboard, and again afterward, as the editor. A system decides in advance what gets reviewed, how closely, and by whom. Review is a design problem, not a discipline problem. If you feel like a full-time editor, it is not because you lack rigor. It is because nobody has built the sorting step that tells you which outputs deserve two minutes and which deserve twenty.
How do you set AI up so it makes fewer mistakes before you review anything?
The short answer: give it your own material, explicit permission to say it does not know, and a job small enough to check. Most errors I see were invited by the prompt, not produced in spite of it.
The clearest guidance here comes from the model makers themselves. Anthropic's documentation on reducing hallucinations lists the same handful of moves I would give a client. Explicitly allow the model to say "I don't know" rather than fill the gap. For long documents, ask it to extract word-for-word quotes first and base its answer only on those. Ask it to cite a supporting quote for each claim, and to remove any claim it cannot support. Restrict it to the documents you provided. The same page ends with a line I quote often, because it comes from the people who build the thing: these techniques "don't eliminate them entirely. Always validate critical information." The vendor is telling you to review. Believe them.
In practice this becomes a three-line preamble you paste above your prompts: Use only the material below. If something is not in it, say so instead of guessing. After you answer, list every specific fact, number, name or date you used and where in the material it came from. That last sentence turns review from reading into checking a short ledger. The other habit is scope. "A full proposal" produces a document you have to read end to end. "The three deliverables section, from these notes" produces something you can verify in a minute. Microsoft's 2026 Work Trend Index, surveying 20,000 knowledge workers who use AI at work, found that 86% treat AI output as a starting point rather than a final answer, and that the most advanced users were far more likely to pause before starting work to decide what AI should do versus a human: 53% against 33%. The pause is the system.
What does a review routine look like that fits into a founder's week?
The short answer: three tiers of review matched to risk, instead of one tier applied to everything. The tiers are decided once, written down, and then they run without a decision each time.
Tier one is internal and low-stakes: a first draft of a social post, a meeting agenda, a brainstorm of workshop names. Skim it for drift and tone. Two minutes, because the worst case is an odd sentence only you see. Tier two is anything a client or prospect will read: proposals, onboarding emails, session summaries, blog posts. Here you check the ledger. Every number, name, date, promise and price gets verified against the source, and you read the whole thing aloud once, because the ear catches invented confidence that the eye slides over. Tier three is money, law, health and contracts. Here the order flips: you write and AI reviews for gaps, or AI drafts and you check line by line and a second person reads it before it leaves. There is no tier where nobody reads.
Two details make this run in real life. The tiers are attached to types of work, not to your mood, so the decision is already made when you are tired. And the ledger step means tier-two review is a checklist of eight to twelve facts, not a re-read of nine hundred words. The same Microsoft index found that when people were asked which human abilities matter most as AI takes on more of the work, the top answers were quality control of AI output at 50% and critical thinking at 46%. That skill is not "reading more carefully." It is knowing which things to read carefully. The way I wire this into a real workflow is described on the AI systems page: review steps sit inside the process, in your own accounts, documented in plain language so a VA or a future hire can run them exactly as you do.
Where should AI never be the last pair of eyes?
The short answer: anywhere a relationship is on the line. A wrong date in a reminder is a logistics error. A wrong tone in a repair message is a relationship error, and those cost far more than two hours.
My rule is to automate the logistics around relationships and never the relationships themselves. Booking, reminders, intake, sorting, follow-up: logistics, and AI can draft most of it with a tier-one or tier-two check. First conversations, a tense scope discussion, the email after a session that went badly, the message when she lands the thing you worked on together: connection, and it stays human even when AI helped with a first draft. Nobody has ever felt seen by a paragraph that was reviewed for accuracy only. As more of the text in your client's life is generated, a message that is clearly from you becomes rarer, and rarer things are worth more. Warmth is the asset AI makes scarcer.
A related fear I hear often: that reviewing AI output means you do not trust it, and that not trusting it means you are behind. The opposite is true. If the anxiety itself is the obstacle, I wrote about overcoming AI fear separately. A review system is what trust looks like when it is earned rather than assumed.
How do you start this week without rebuilding everything?
The short answer: write your three tiers on one page, add the ledger line to your prompts, and test it on one real piece of work.
- List the ten kinds of text you produce most and put each one in tier one, two or three. Fifteen minutes, and it is the whole system.
- Save the preamble where you can paste it in five seconds.
- Pick one tier-two document this week. Draft it with the preamble, review only the ledger, read it aloud once. Time it against your last full re-read.
- Mark the no-go list: the three or four messages that stay handwritten no matter what. Repair, scope, first contact, and whichever one is personal to your practice.
- Put the page in your own account, in plain language, so it works for the next person too.
Founder to founder: the goal was never to stop checking AI's work. The goal was to stop checking all of it equally, which exhausts you and still lets the invented statistic through. The model guesses because it was built to guess. You are the part of the system that knows a guess from a fact, and once that job is sized properly, it takes minutes, not evenings. Pick one document this week and run it through the tiers.
Frequently Asked Questions
Why does AI make up facts that sound so real?
Because it is trained to predict plausible text and graded in ways that reward a confident guess over "I don't know." OpenAI's 2025 research describes hallucinations as plausible but false statements that arise from those incentives, so fluency is not a sign of accuracy.
Will the newest AI models stop hallucinating?
They hallucinate less, but the model makers themselves say the problem is not fully solved. OpenAI notes that its newer models still hallucinate, and Anthropic's guidance says its techniques reduce errors without eliminating them, which is why a review step stays in any serious workflow.
How do I check AI output quickly without reading every word?
Ask the model to list every specific fact, number, name and date it used and where it came from, then verify that list instead of re-reading the whole draft. Add a read-aloud pass for anything a client will see, and reserve line-by-line review for money, legal and health topics.
What kinds of AI mistakes are most common in a service business?
Invented statistics, wrong names and dates, made-up quotes or links, outdated prices, and "drift," where the model adds something your client never said. Errors cluster around specific, rarely repeated facts, so check those first.
Should I stop using AI for client-facing work?
No, but match the review to the risk. AI can draft the logistics of client work, such as reminders, summaries and proposals, as long as facts are verified before sending; the messages that carry the relationship itself should stay written by you.
The Scale Atelier builds AI systems for women-led service businesses with the review steps designed in, living in your own accounts and documented so anyone you hire can run them. If you would like to know where AI is quietly costing you review time, book a free 20-minute audit: no pitch, just a prioritized plan that is yours to keep either way.