Why I'm Pausing
My AI Influencer

In June I built a fully AI-generated influencer called Blythe Sterling. Part 1 covered how I made her. Part 2 covered the first month of running her. This is the last part, for now.

I am pausing the account. Not because the technology failed. Because the audience never came, and the time it took to keep going no longer made sense.

From the start, this was an experiment. I wanted to see what is actually possible with AI content today, not what the demos promise. On that measure, it delivered. On the social media measure, it did not. Both are worth writing down honestly.

The Honest Numbers

Every post was logged. Every follower count was read weekly. Here is where it ended.

The pattern that matters is the trend. Production quality went up every month. Follows went down. The account got better at making things and worse at being followed.

What Happened After Month One

At the end of Part 2, I promised to rebuild everything around one number: follower growth. I did. Then I did it again. And again.

Over the next eleven weeks the strategy changed six times. A new styling direction. A dating storyline. A push for TikTok reach. A switch to Instagram first, after research showed the AI accounts that grow are Instagram and YouTube-led. Then a final pivot to short opinion pieces, Blythe talking to camera about being AI.

Each change was based on the data. Each one made sense on its own. Together, they were part of the problem. The audience was never given one thing to follow. Every few weeks, Blythe became a slightly different account.

July was the best month, with 69 follows. After that, every change produced better-looking content and fewer new followers. In September, a post about how AI renders skin and ageing reached 1,192 people on Instagram, the best Instagram result in three months. It brought in one follower.

Reach was never the problem. People watched. They just did not stay.

What Cannot Be Automated

The whole point of the project was to find the line between what AI can own and what it cannot. Four months gave me a clear answer, and it is not where I expected.

• The ideas. This was the biggest disappointment. The system drafted concepts every fortnight, and they were always competent and always cliché. Safe, expected, forgettable. My best-performing posts were my own ideas. The creative spark is the one thing I most wanted to hand over, and the one thing I could not.
• The posting. Everyone hates posting by hand. But the platforms favour content posted natively, so automated posting risks quietly burying the account. Posting stayed manual from the first day to the last.
• The quality check. Every clip needed a human. Did the character drift? Is the sound clean? Does the edit land? AI cannot reliably answer those questions about its own work yet.

What AI Did Brilliantly

The surprise ran the other way too. Some of the work I expected to be painful turned out to be the strongest part of the system.

• Editing. AI assembled the finished videos: cuts, captions, audio levels, timing. Work that would take an editor hours ran as a script.
• Connecting the tools. Running video generation directly from Claude, through the Higgsfield connection, worked very well. One conversation could plan, generate and assemble.
• Video and voice quality. By the end, the videos and the cloned voice were genuinely good. Lip-synced, consistent and convincing enough that people sent love letters to a woman labelled as AI on every post.

The Real Cost of Quality

In Part 2, I estimated my time at about fifty minutes a fortnight, plus a few minutes a day posting. That was the early version. As the standard went up, so did the hours.

By the end, Blythe took me 4 to 5 hours every two weeks. Five or six times my first estimate.

Most of that was supervision. High-quality video generation still goes wrong often. A face drifts between shots. A hand comes out with the wrong number of fingers. The audio jumps at a cut. Every clip had to be watched, checked and often regenerated. The machine did the making. I did the checking, and the checking was the job.

The money tells the same story. I started on a monthly Higgsfield plan at around £38 a month. In August I moved to the annual Plus plan, $372 for the year on a promotional price, against a full price of $708. Quality video uses credits fast, and every rejected clip still costs the credits it took to make.

It Can Be Done. Just Not Like This.

It would be easy to conclude that AI influencers do not work. That is not what the data says. Some AI accounts are growing fast. One London AI account, in the same niche and on the same platform, carrying the same AI label, went from under 800 to over 4,800 Instagram followers in a single month while Blythe stood still.

So the AI label was not the obstacle. The platform was not the obstacle. What we were missing was a story.

Blythe had a beautiful look, a clear voice and a good premise. What she did not have was a storyline viewers wanted to come back for. The accounts that grow give people a reason to return tomorrow: longer videos, real narratives, characters and stakes. Short, polished, standalone posts get watched and forgotten.

Building that would take significantly more time and significantly more money. Longer formats mean more generation, more credits and more supervision per minute of video. To make an AI influencer genuinely entertaining, the investment goes up, not down.

Why I'm Pausing

Blythe takes a great deal of time to maintain, and right now she gives nothing back. Four to five hours a fortnight, every fortnight, for an account that is not growing. That time has a better use.

So I am pausing. The account stays visible. Nothing is deleted. Parts 1 and 2 stay up, because what they describe is still true.

I may come back. If video generation becomes cheaper, more available and needs less human supervision, the economics change completely. When that happens, I will know exactly what to build, and exactly what not to waste time on.

The Findings That Travel

Most of what I learned applies well beyond one AI persona. If your team is putting AI into real work, these are the five worth keeping.

• Automate production, not ideas. AI is excellent at making things once you know what to make. It is poor at deciding what is worth making. Keep the creative brief human.
• Quality moves the human cost, it does not remove it. The higher the standard, the more of your time goes into checking. Budget for review, not just generation.
• Count the hours, not just the subscription. The tools cost a few hundred pounds a year. My time cost far more. Most AI business cases leave the second number out.
• Strategy churn is its own failure. Six well-reasoned changes in eleven weeks did more damage than any single wrong choice. Pick a direction and give it time to work.
• Set the stop rule before you start. I had a written target and a date to check it. When the date came, the decision was already made. That is the only reason pausing felt easy.

Pausing is not the same as failing. The experiment did what it was designed to do. It showed me, with real data and real hours, where AI stands in October 2026. That is worth more than a few hundred followers.

Last updated: October 2026.