Last year I got access to OpenAI's GPT-3 API and started using it in real workflows, not demos. This year I added Jasper on top for content work. After more than a year of running this in production across our community platforms, some honest notes.
What actually works
First drafts. Event descriptions, session summaries, newsletter sections, social variants. The model produces a competent draft in seconds. A human still edits everything, but starting from 70 percent beats starting from zero, and across hundreds of pieces of content per month that compounds into real hours.
Repetitive transformation. Take this long thing and make it short. Take this English text and adapt the tone. Take these bullet points and make them a paragraph. This is where the model is most reliable, because the input already contains the facts.
What does not work
Facts. The model invents things with total confidence. Names, numbers, dates. Anything factual has to come from us and be checked by us. We learned to treat the model as a writer, never as a source.
Anything unsupervised. Every output passes a human. The moment we tried skipping review for low-stakes content, quality drifted within days. The leverage is real but it is supervised leverage.
Why I am writing this down
Most people I talk to still file this under toys. My read is different. I have seen this movie twice: with open source in the early 2000s and with cloud a decade later. A substrate shift looks like a toy, then suddenly it is the default and everyone claims they saw it coming. I want a written record that we were building on it while it was still quiet.
If you run a content-heavy operation and you are not experimenting with this yet, start small: one workflow, one editor who reviews everything, measure the hours. That is all we did.