AI at Work: The Tasks Even AI Insiders Still Do Themselves
If anyone should be comfortable handing work over to AI, it's someone who works at an AI startup. They see the product being built. They know what the models can do. And yet — even there, some things stay human.
A piece making the rounds this week from Business Insider puts a name to what a lot of people quietly feel: the only nontechnical employee at an AI startup still handles certain tasks personally, and won't apologize for it. The writer drafts her own tricky emails. She does her own research when the stakes are high. She makes the calls that require real judgment about people.
That's not technophobia. It's something more interesting — and more useful to understand if you're trying to figure out how to work smarter with AI at work right now.
What really decides it: who owns the mistake
People tend to frame AI hesitation as a capability question: can the AI do this task well enough? That's the wrong frame. A good language model can draft a competent email, summarize a document, generate a presentation outline, or write functional code. It can do most of those things faster than you.
The real question is: what happens if it gets it wrong, and who owns that?
When you send an email to a difficult client, you're not just sending words — you're sending a signal about who you are, how you read the room, and what you care about. If AI drafts that email and something in the tone misses by five degrees, the client doesn't blame the software. You wear it. The consequence belongs to you, so the judgment should too.
This is why trust in AI tools tends to track with reversibility. You'll trust a model to help you brainstorm because a bad brainstorm idea costs you thirty seconds of backspacing. You'll trust it to summarize meeting notes because a summary error is easy to catch. But you're more careful when the output goes directly to someone else, or when being wrong has costs that take weeks to undo.
Where AI actually earns its keep
To be clear: AI at work is genuinely useful. Not in a vague productivity-buzzword way — in specific, measurable ways that free up hours.
Research compression is the most underrated one. Say you need to understand a regulatory change, a competitor's pricing structure, or the basics of a technical concept outside your expertise. Before AI tools, that might cost you two hours of reading. A well-prompted model can get you to an 80% understanding in ten minutes, and then you can decide if the last 20% matters enough to dig further. It usually doesn't.
First-draft generation is another. Writing is slow. The blank page is the worst part. If you can prompt a model to put something coherent on the page — even if you'll rewrite most of it — you've bypassed the hardest thirty minutes of the task. Most professional writers I know quietly use AI this way now, even if they don't say so publicly.
Repetitive structured tasks are where AI shines brightest. Data formatting, contract clause comparison, generating variations on a standard document, answering the same question with slightly different context — this is exactly the dull, low-judgment work that AI does without complaining and without errors that a tired person would make at 4pm on a Thursday.
The three tasks worth keeping to yourself
The nontechnical startup employee in that Business Insider piece identified a few categories she wouldn't outsource. They're worth examining, because they point to something real.
First: sensitive human communication. Messages where the relationship is the point — where saying the right thing matters more than saying anything quickly. Firing someone. Apologizing to a client. Negotiating something personal. The words matter less than the fact that a real person chose them and means them. AI can help you prepare, sure. Think through what you want to say, identify what might land badly, check your tone after you've written it. But the actual writing, in those moments? Keep it.
Second: original strategic thinking. Not "summarize what's been said" — that's fine to outsource — but "what should we actually do here, given everything we know about our situation." AI can give you frameworks and options. It can tell you what companies in analogous positions have historically done. It cannot weigh the specific idiosyncrasies of your team, your market position, your boss's risk tolerance, and your company's actual runway in the way that someone who lives inside the situation can. Strategy is contextual in ways that are hard to fully convey in a prompt.
Third: anything where being wrong looks like lying. This is a subtler one. If an AI tool invents a statistic, gets a date wrong, or confidently states something that turns out to be false — and you sent that output to your CEO or a client without checking — you look like you made up a fact. The model's error becomes your credibility problem. For anything that involves specific claims, numbers, or citations, verify before it leaves your hands. Every time.
What this means if you're not at an AI startup
Most people reading this aren't embedded at a company building AI products. They're using AI tools at a regular job — in finance, in marketing, in law, in operations — and trying to work out what the right level of reliance looks like.
The startup experience is actually a useful proxy here. People who are closest to the technology, who understand its capabilities most clearly, are not handing everything over. They're developing a feel for the grain of the tool — which direction it cuts cleanly, and where it binds.
If you work in finance and you're using AI to help draft client communications, that's fine, but read everything before it goes out. Especially anything with numbers. The model might get a return percentage wrong, or describe a portfolio allocation in a way that's technically accurate but misleading given what the client actually understands. The liability is yours, not the tool's.
If you're in a legal role, the AI can help you find precedents, structure arguments, and speed up document review — but you wouldn't submit an AI-generated filing without a close human read. Any lawyer who's done that has stories. Not good ones.
If you're in a creative field, the trap is different: the AI output is fine, it's just not very surprising. It gives you the median good answer. If your job is to be distinctive — to produce work that says something other work wouldn't — you need to bring more of yourself to the output, not less.
The career question underneath all of this
There's a broader thing worth saying here, especially given that college students are entering a job market where AI fluency is increasingly expected: knowing how to use AI tools is now table stakes — a baseline, not a differentiator.
What differentiates you is the judgment layer on top. Knowing when to use AI and when not to. Knowing how to evaluate AI output critically. Knowing when the confident-sounding answer the model gave you is actually wrong. Those skills don't come from using AI more — they come from staying involved in the work even while using AI, so you maintain the expertise needed to catch its mistakes.
The risk that doesn't get discussed enough: if you outsource your thinking completely and consistently, you slowly lose the calibration that makes you good at the judgment calls. You get worse at the exact thing AI can't replace. That's the drift worth watching for.
The nontechnical employee at the AI startup seems to understand this intuitively. She's not avoiding AI out of stubbornness — she's protecting the thing that makes her valuable.
A few quick questions, answered
Does using AI at work make you look less capable to managers?
Roughly the opposite, right now. Most managers are actively looking for employees who know how to use AI tools to work faster and handle more. The concern is usually not that someone used AI — it's that they used it without checking the output. Produce good work efficiently and own the quality. How you got there is less important than whether it's actually good.
Should you disclose when you used AI to write or research something?
In most workplace contexts, no — just as you don't disclose that you used spell-check or a calculator. The output is yours; the tool is infrastructure. The exceptions are contexts where originality or personal voice is explicitly the point (some journalism, academic work, certain client deliverables), or where your company has a policy requiring disclosure. When in doubt, check what your employer's guidelines say. More of them have written policies now than a year ago.
Going into the second half of 2026, watch how companies start measuring AI-assisted output. Right now there's almost no accountability infrastructure for it. That will change, and when it does, the people who built genuine judgment alongside their AI habits will be in a much better position than those who just got fast at prompting.
