AI Is Not Your Employee. It Is an Intern With a Perfect Memory and Bad Judgment.

Stop treating AI like a replacement employee. Treat it like a very fast intern that still needs someone responsible in the room.

Roughly 6 in 10 U.S. small businesses now use generative AI in some form, according to the U.S. Chamber of Commerce. That number was 23% just two years earlier. Adoption is no longer the question. The question is whether anyone is checking the work.

Here is the mental model that answers it: AI is not your employee. It is an intern with a perfect memory and bad judgment.

The intern rule: useful, fast, but not unsupervised

Think about what a brilliant intern brings on day one. Speed. Endless energy. They have read everything. They never complain about boring work, and they can produce a draft in minutes that would take you an hour.

Now think about what that intern lacks. Judgment. Context about your business. The instinct to say "I'm not sure, let me check with someone." And critically, an intern who wants to impress you will sometimes give you a confident answer instead of a correct one.

That is a near-perfect description of every major AI tool on the market today. The "intern" framing isn't ours alone. Wired founding editor Kevin Kelly called generative AI a "universal intern" back in 2023, and Wharton professor Ethan Mollick described AI tools as interns that "work infinitely fast and sometimes lie to make you happy." Three years later, the description still holds. The tools got smarter. They did not get judgment.

No business owner would let a first-week intern email clients, quote prices, or file paperwork with a court without someone reviewing it first. Apply the same standard to AI and most of the risk disappears.

Where AI is safe to use first

The good news: when AI is used on the right tasks, the gains are real and measured. A randomized study of 758 knowledge workers run by Harvard Business School and Boston Consulting Group found that people using AI on well-suited tasks completed 12.2% more work, finished 25.1% faster, and produced results rated about 40% higher in quality.

The tasks where AI performs well share a pattern: you give it the material, and you check the result.

• First drafts. Emails, job postings, social captions, service descriptions. You edit before anything goes out.
• Summarizing documents you provide. This is one of AI's most reliable skills. On controlled summarization benchmarks, top models add unsupported content less than 2% of the time when summarizing a supplied document.
• Brainstorming and organizing. Marketing angles, meeting agendas, turning messy notes into a clean outline.
• Formatting and reformatting. Turning a list into a table, a table into an email, an email into a checklist.

Notice what these have in common. The cost of an error is low, a human sees the output before it matters, and the AI is working from material you handed it rather than pulling "facts" from memory.

Where AI should never act alone

The same study found something most AI vendors will not put in their marketing. On tasks outside the AI's real capability, workers using AI were 19% less likely to get the right answer than workers with no AI at all. The plausible-sounding wrong answer did more damage than no answer.

The failure cases are not hypothetical. They are court records.

• In Mata v. Avianca, New York lawyers were sanctioned after ChatGPT invented six court cases that never existed and nobody checked before filing.
• In 2025, a California attorney was fined $10,000 after a brief he filed contained 21 fabricated quotations out of 23. He admitted he never reviewed the AI's output.
• Air Canada's customer service chatbot invented a refund policy that did not exist. A court ruled the airline had to honor it. The "it was the bot, not us" defense lost.

And this is with specialized tools, not just free chatbots. Stanford researchers tested purpose-built legal AI products and found they produced false information 17% to 33% of the time. General chatbots answering legal questions failed 58% to 82% of the time.

The pattern across every failure: the AI produced something specific, confident, and wrong, and no human looked at it before it reached a customer, a court, or the public.

Why review is not optional

Two reasons, one practical and one legal.

The practical reason: AI errors do not look like errors. A typo announces itself. A fabricated statistic, an invented policy, or a wrong price looks exactly like the correct version. The only way to catch it is for someone who knows the truth to read it. That is review, and it takes minutes, not hours.

The legal reason: regulators have already decided whose problem this is. The FTC's Operation AI Comply made the position plain in 2024: existing consumer protection law applies fully to AI output. If your chatbot misquotes a price or your AI-written ad makes a false claim, that is your false claim. The Air Canada ruling says courts agree.

Most businesses have not caught up to this. A 2026 American Arbitration Association survey of 500 senior leaders found that 87% of organizations claim to have AI governance, but only 22% say it actually works in practice. A policy in a drawer is not supervision. A named person who reads the output before it ships is.

The green / yellow / red list for local businesses

Here is the intern rule turned into something you can post next to the coffee maker. Three categories, based on one question: what happens if the AI gets it wrong?

Sources are linked inline. Key research: Dell'Acqua et al., Harvard Business School / BCG randomized trial of 758 knowledge workers (2023, published in Organization Science 2025); Stanford RegLab legal AI hallucination study (2024); American Arbitration Association governance survey of 500 senior leaders (2026); U.S. Chamber of Commerce small business technology report (2025). The "AI as intern" framing traces to Kevin Kelly (SXSW 2023) and Ethan Mollick (One Useful Thing, 2023).