Now, this isn’t another rant about how magical or useless AI is. My stance is pretty mild: it handles baseline work well but often falls short when the output really needs to perform. Instead, I want to talk about understanding and accurately communicating the capabilities of LLMs.
A pretty standard situation now is management asking if a particular content pipeline can be automated with AI. The issue I keep seeing is that the specialist says “yes,” the manager hears that “yes,” but they have pretty different ideas of what that “yes” means.
Typically, the specialist means something like “AI can produce drafts or handle the predictable parts of the pipeline.”
The manager hears that the pipeline can now operate automatically and reliably without requiring human hours.
The issue is that however well AI performs, someone still has to provide input and context, evaluate the output, catch plausible-looking bullshit, edit it, handle exceptions, and take responsibility when the pipeline fails.
My general takeaway from communication about AI is that precision is crucial here. If you say, “Yes, AI can handle that,” but then spend 3 hours a day on this task, is the question “How come you’re working on this if we have AI?” really out of the blue?
When answering, it is helpful to consider what the other person knows and what they don’t. If your manager doesn’t have years of experience in your field or with AI-assisted workflows, you need to fill that gap in your answer.
Instead of saying, “Yes, we can use AI here,” I have found it helpful to be very specific about:
What exactly AI can handle
What human work will remain
What failure rate we are prepared to accept
The better answer might be: “AI can handle these parts, but this part still requires human hours for these reasons. The AI-generated parts also require quality checks for these reasons. We could automate another part, but it would produce problematic results at a certain rate, so we need to decide whether the additional speed is worth the lower quality.”
And I personally prefer to be really explicit about that last decision. If the company knowingly chooses speed over quality, the resulting problems should be accepted as a business tradeoff. The people working on the pipeline should not be expected to somehow produce perfect quality from a process designed to tolerate bad output.
—————————————
Мне интересно узнать про опыт технических специалистов с ИИшкой. У вас есть такое же ощущение, что простой, понятный дефолтный код ИИшка напишет плюс-минус без проблем, но если нужно, чтобы код реально перформил и был хорошо оптимизирован, то генерить его – себе дороже, исправлять потом придется дольше, чем самому написать?
И как вы обсуждаете с компанией интеграцию ИИ? Получается согласовать, что можно, а что не нужно отдавать ИИ, или сталкиваетесь с завышенными ожиданиями от аутпута ИИ?
Пишите ваши мысли в комментариях на английском – и я, как всегда, по некоторым комментам дам подсказки или фидбек в плане английского языка.
