I work at Publora — an API and MCP for publishing to social networks from one place, including through an AI agent. So, in a sense, my job is selling automation for posting. Which made the answers I started getting from actual people a little awkward.
I've been doing a small study — nothing fancy, just messaging people directly and asking: what do you use to post, and why do you do it that way? No pitch, no setup. It's an ongoing thing, reaching out to people across different niches one by one, so this isn't a clean statistic — more a pile of honest answers I'm still adding to. A few hundred replies in, and I'm still going.
The answers are all over the place, but one thing keeps coming up. People aren't asking for more automation. Quite a lot of them want to keep the final step for themselves.
A few answers, anonymized.
One person told me: "I use AI rarely, do the core work myself. Sometimes I go into ChatGPT or Gemini as a 'second brain' — when I'm out of ideas or a deadline's closing in. But at the end of the day it's still just a machine."
So AI is useful to think with. That doesn't mean they want it doing everything.
Another: "I use Claude, but not for posting — it sort of sets me up, and I publish myself."
And a third was very direct: "I'm wary of AI tools having access to my account. LinkedIn isn't forgiving about it."
That one wasn't really about whether AI-written posts sound authentic. They simply didn't want another tool having that much access to their account.
Then there was this: "I like hitting publish myself. In a world where everything's polished, it makes the content feel a little more alive."
Different reasons: sometimes security, sometimes control, sometimes just liking the fact that there's still a person involved. But I kept hearing versions of the same thing. Draft for me, help me think, prepare it — fine. I'll publish it.
Nobody was dreaming about an agent happily posting all day while they went for coffee.
At first this was slightly uncomfortable to hear, considering what I do for a living. But the more replies I read, the less it sounded like an objection to automation. It sounded more like people being quite specific about which parts they actually want automated.
Where that caution stops being a quirk
One answer stood out because it came from someone managing several brands rather than just their own account.
They prefer tools with integrations because content and reporting are easier there than doing everything natively. The problem is that not every platform has the integrations they need, and the more channels a brand uses, the messier the whole setup gets.
That changes the problem quite a bit.
Being cautious about giving AI access to your own LinkedIn account is one thing. Being an agency responsible for ten clients' accounts is another. Now you're dealing with reputations that aren't yours, client credentials, several calendars, approvals living somewhere in email, and the possibility of someone putting one client's post on another client's account.
At that point, automating everything sounds terrifying. Doing everything manually isn't much better. Ten clients still have to fit into the same working day.
So you need something in between.
What that looks like in practice
After reading the replies, I kept coming back to three things.
The first is keeping clients separate. Each one needs their own connections, content and calendar, without the possibility of a post quietly ending up on somebody else's account. A posting mistake on your own account is embarrassing. Doing it to a client's account can get expensive.
The second is passwords. Ideally, an agency shouldn't have them at all. The client connects their account through an OAuth link, and the team can work with it without ever knowing the login. The person who told me they were wary of tools having access to their account wasn't being paranoid. That's a perfectly reasonable thing to worry about.
And then there's approval. An agent can draft a post. It can prepare it, schedule it, do the repetitive bits. But publishing can still wait for a person to approve it. In a team, that can also mean the person who wrote the post isn't the person who approves it.
It's basically the same habit I kept seeing in the survey, only with more people involved.
We build this at Publora — it's our workspace for agencies, and the API is documented here. I won't turn this into a tour of the product; if that's what you're looking for, the links are there.
What's been more interesting for me is realizing that some of the things we'd been thinking about as product features map almost exactly to things people told me they were worried about.
What I took from it
Automation is usually sold as: let the machine do this for you.
But after reading these replies, I think there's a pretty big group of people asking for something slightly different: help me do it, just don't take the whole thing away from me.
For someone posting to their own account, that might simply mean wanting to hit Publish themselves. For an agency, the same instinct turns into much more practical requirements: keep clients separate, don't pass passwords around, and make sure a person approves what goes out.
I find it slightly funny that it took working inside an automation company to make me think harder about what shouldn't be automated.
And for the record, I wrote this with AI. It helped me sort through material and clean up the draft. Deciding what was interesting in a few hundred replies was still my job.
I'm keeping that bit.
Where do you draw the line? What are you happy to hand to an agent, and what do you still want to do yourself?
Top comments (4)
Generally, I think people are too afraid to trust AI with credentials, or to execute final actions, because that's a risk there's no coming back from. Eg. like having an AI commit to git, is fine... pushing to git on the other hand, not fine... There has to be a level of certainty in the outcome and if it's irreversible, they'd rather do it themselves.
Take for example, you write your CV into a prompt. Do you trust AI to send it out? Or do you trust AI to produce a pdf you can review? That's the difference between 1 click risk and 2 click certainty. And that's probably what keeps users from using it? If they can see it in their browser and click submit, instead of click and pray, they'd have alot more confidence using the tool?
Ha, you read my mind — we're building an agent right now, and mine cheerfully assures me it's totally fine to spin up a repo, everything will be great. Meanwhile I'm sitting there uneasy. And the funny part: when we test a plugin that isn't published anywhere, zero real risk, it suddenly gets careful and makes me walk the whole flow by hand — approve this, confirm that.
So it's sure of itself right where I'd want it to stop, and nervous right where I'd happily let it run. That instinct for what you can take back is the thing the agent doesn't have. Which is probably why the last click stays ours.
That's why you need to gate the tool executions. If it tries to fire eg. 'post to dev.to', the harness has to prompt you, not the agent. the choice to move on has to be at the tool execution level, not at the model reasoning level.
Yeah, exactly. The gate can't sit in the agent's head — that's just the agent deciding whether to check itself, which is how we got here. Put it at the tool call and the harness stops it before 'post to dev.to' fires, however sure the model managed to talk itself into being.
And honestly this keeps showing up no matter where you poke it: the thing doing the check has to sit outside the thing being checked. Agent suggests, something it can't argue with says go.