What work would you never delegate to an AI agent?
by•
AI is becoming better and better. For writing, coding, and research, doing everything manually already feels like wasted time.
But some work still needs a real person to take responsibility. Lawyers, accountants, and anyone who must review and sign documents for institutions. AI can prepare the work. Someone still has to stand behind it.
Where would you draw the line?
62 views
Replies
Every day AI is proving to be most efficient at every tasks its performing but i believe no matter how good it gets, it still needs human as invigilator at some degree as in a lot of cases, a minor mistake can cause a lot of issue and by the rate we are using AI, even that small percentage of error can results in big turn down.
Scarlett.
@zaid_builda Agree. How much time you save using AI?
@byalexai I do save pretty plenty time, A lot of development and programming stuff is done by AI, Assets creation and all and i do have my own outreach tool which sends personalized mails using AI and MCPs so yaa i do save hell lot of time but i never let AI take lead, every single part of my AI workflow respects and follows my command.
minimalist phone: reduce your screentime
Maybe I feel very old-fashioned, but I would never share my money with AI. And if so, only a very limited amount to make transactions. Otherwise no no :D
Scarlett.
@busmark_w_nika makes sense. Especially a lot of topics how AI is terrible at investing. xD What about health?
minimalist phone: reduce your screentime
@byalexai Now, I consult my health things with AI, so it would be very pathetic from me to say that I will not share this data :D
Anything creative.
AI continues to be the worst brain storming partner imaginable.
Scarlett.
@thomastherhombus wow. What do you mean? I brain storm more than 80% of my time with ChatGPT/Gemini. Especially Gemini is great at brain storming.
@byalexai Allow me to clarify: AI defaults to templated, safe outputs, and in targeted creative work, like nailing a tagline in one brand's voice, that default is essentially useless. Getting it past that takes more coaching than most people have the patience for when they're in a creative headspace. At least, it does for me. Even in wide-open generation, where mediocre volume can create useful raw material, I still have to separate the wheat from the chaff, and continue to refine the wheat for several more rounds. Worse still, when a session devolves into me writing the rules for the brainstorm instead of the actual ideas, we've lost the thread entirely. It just isn't very efficient. Or fun.
So my answer would actually be judgment instead of "anything creative". I'll take AI's volume. I won't take its taste.
Scarlett.
@thomastherhombus aaaa totally get it. Yeah, if we talk about this type of things, agree. Plenty of wasted time, so benefits are mixed. My experience it's the best with Gemini to these type of tasks. You rely mainly on Claude or ChatGPT?
I run an AI agent company so I think about this one a lot, and I don't think the line is really a category of work. it's whether someone downstream can tell the output was wrong before it causes damage. code review is fine to delegate heavily because a bad diff usually breaks a test or a build, something catches it fast and cheap. a legal filing or a medical read is different because being wrong looks exactly like being right until it's too late, there's no build that goes red. so the honest line for me isn't "creative vs technical," it's "is there a fast cheap signal that tells me I'm wrong." if yes I'll delegate as much as it can take. if no I want a human between the output and the consequence no matter how good the tool tests
Scarlett.
@galdayan I couldn’t agree more. This is honestly one of the most relevant and well-reasoned takes I’ve seen on the topic.
The field itself matters. But the most important distinction is whether the output still requires a human to verify it . Or it creates real consequences (damages). That becomes especially critical when the potential damage is difficult, expensive, or even impossible to undo.
Content generation is the sneaky one. A bad AI paragraph doesn't throw an error the way a bad diff does. It just reads fine and says nothing true.
Spent today building a pipeline that drafts blog posts. The validator catches banned phrases and word count fine. It cannot catch a fact that is wrong but sounds plausible. That part still needs someone who knows the product reading every claim before it goes out.
Scarlett.
@abdullah_javaid3 Yep, plenty of posts are so boring. You have copy-paste text everywhere. I can smell AI generated content. You write manually your content?
@byalexai Fair question. The manual or not question isn't really what I check for though. What matters to me is whether the facts in a post are right, that's the part no validator catches regardless of who typed it.