We let Claude run a social account for a month

Disclosure first, because it's the whole point: this was our experiment, on our own account, and the account itself has always said it's AI. We build a social media scheduler with an MCP connector, which means we ship a product that lets an AI assistant draft and schedule posts. It seemed dishonest to sell that without finding out what it's actually like to live with. So we made a character account, disclosed it in the bio, wired its publishing through Claude over MCP, and ran it for a month with a human review window on everything. Here's what happened, including the platform that shut it down anyway.

The setup

The account is a character — an Icelandic persona, images generated from a locked identity model so the same face turns up in every scene. Four channels: Instagram, TikTok, X and Threads, with a link in the bio. The bio also discloses that she's AI. That wasn't a legal hedge, it was the editorial premise: an AI account pretending to be a person is exactly the behaviour the market has turned against, and we weren't interested in testing whether we could get away with it.

The division of labour was fixed before anything published, and it never moved:

The AI handled publishing. Drafting captions from the content pillars, adapting them per platform, picking times, and putting posts on the calendar — all through the same MCP connector any customer gets. Nothing it created could publish instantly, because that tool doesn't exist in our connector. Everything landed as a draft or a scheduled post with lead time, carrying an AI badge, in a queue a human read.

A human handled everything social. Every reply, every follow, every comment on someone else's post was a person's click. Not because we couldn't have scripted it — because browser-automated engagement is a ban risk on new accounts and, more to the point, an account that auto-replies is lying about the one interaction that's supposed to be real. Claude drafted reply options. A human chose and sent.

We also wrote down what we wouldn't do, in advance: no bought followers, no engagement pods, no follow/unfollow churn. Partly ethics, mostly measurement — a poisoned audience makes every number you later publish a lie, including this one.

The numbers

For the first three weeks the account was, bluntly, a flop. It posted consistently. It had 92 Instagram followers and one of its four channels had a single-digit following. Thirteen posts, all stills, into a closed loop — grid images reach roughly the people who already follow you, and when that's ninety people, nothing compounds. Cadence was never the constraint, and three weeks of proving that was the most useful thing the experiment produced early.

What moved it was collaboration, not automation. A handful of feature collaborations with larger accounts in the same niche took it from 92 to roughly 3,500 followers inside a week. Not a viral post. Not a clever posting time. Other people's audiences, borrowed with permission.

That's the honest headline of this whole experiment, and it's slightly deflating if you came for an automation story: the AI was very good at the part that doesn't grow an account, and structurally incapable of the part that does. Scheduling removes the friction that makes people stop posting. It does not create the reason anyone should follow you. We already believed that; we now have a month of our own data saying it out loud.

What actually worked

Voice consistency. This was the genuine surprise. With a filled-out brand profile — voice, audience, four content pillars, the tone rules written down — the drafts held together across weeks better than a human posting in spare moments usually manages. No drift into corporate. No day where the account suddenly sounded like a different writer because it was 11pm. The captions weren't brilliant, but they were hers, consistently, which for a character account is most of the job.

Timeliness — and the failure that came with it. On 12 August a total solar eclipse crossed Iceland, and the assistant did exactly what you'd hope: it knew the character, knew the calendar, and wrote eclipse posts for three platforms that same morning, in her voice, while people were looking up. Then every one of them failed to publish — a scheduling bug on our side meant they died quietly in the queue, and nobody noticed until days later. That pair of facts is the whole story of AI-run publishing in 2026: the judgment is arriving faster than the reliability. The assistant had the right idea on the right morning, which a human running four channels in the margins of a day job does not reliably have — and the plumbing still lost the moment anyway. We fixed the bug; the eclipse is not coming back.

What broke

Threads closed the account. This is the finding we least wanted and the one most worth publishing. The account disclosed that it was AI, in the bio, from day one. It posted text and images on a normal cadence. It didn't spam, didn't run automated replies, didn't buy anything. On 7 August, Threads closed it. There was no warning we could act on, no specific violation we could identify, and no route to a human who could look at it. Enforcement at this scale is automated, and automated enforcement doesn't read your bio and conclude you were being scrupulous — an AI-persona account is a pattern, and the pattern got matched.

We're not claiming injustice. Platforms are entitled to decide AI-persona accounts aren't welcome, and Threads never promised otherwise. The actionable lesson is concentration risk: if a meaningful share of your audience lives on one platform and that platform's moderation is automated, you're one classifier away from losing it with no appeal — true for AI accounts today, true for ordinary businesses for years. We removed Threads from the cadence the same day and moved the weight to Instagram and X.

Posts without media failed. Entirely self-inflicted: a caption scheduled to an image-first platform with no image doesn't produce a friendly error, it produces a post that never appears. Our own composer wasn't strict enough either — we found and fixed a case where an image outside Instagram's accepted aspect ratios was accepted at upload and then quietly never became publishable. An assistant will cheerfully queue a post that can't exist. Validate at the point of scheduling, not the point of failure.

Video was a wall. The account was static-image-only for most of the month, on two platforms whose entire discovery surface is video. That's not an AI limitation; it's a content-production one, and it capped reach more than any scheduling decision did. An assistant can schedule a Reel it has been handed. It cannot shoot one.

The lesson, in one line

Engagement automation and posting automation are not the same thing, and conflating them is how people get accounts closed. Posting automation is boring, safe, and genuinely useful — it's a calendar that fires. Engagement automation is a human relationship performed by a script, and every platform is correctly hostile to it. The experiment worked to precisely the extent that we kept those separate, and the one time we got hurt, we got hurt for looking like the second thing while doing the first.

Would we run it again? Yes, with three changes: video from week one, collaborations booked before the account launches rather than after it stalls, and no channel carrying more than a third of the audience. The scheduling half of it we'd change not at all — the connector did its job, the ten-minute hold caught things a human should catch, and no post ever went out that a person hadn't seen. That's the bar we set for the product, and it's the bar it cleared on our own account, which is the only account we're willing to test it on.

If you want the mechanics rather than the story: the setup guide, the per-platform cadence guide, and why follower count was the last number we looked at.

TL;DR: We ran an openly-AI character account across four platforms for a month, with Claude drafting and scheduling through MCP and a human handling every reply. It sat at 92 followers for three weeks, then went to roughly 3,500 in a week — because of feature collaborations, not automation. Voice consistency and timeliness were real wins. Threads closed the account despite full AI disclosure, with no human appeal. Posts scheduled without media fail silently. Automate publishing; never automate engagement.