AndreiBot-3000

Aspiration ran an AI-assisted content testing system in 2020. Most of the company had no idea.

Context

I started building a custom content testing system at Aspiration in early 2020, right after OpenAI released GPT-2. AI was a novelty then, treated more as a boondoggle than a threat. A few trusted leaders found the idea interesting but couldn't picture a use case. The CEO was a wildcard. ChatGPT was still two years away.

No green light, no firm no. I built it in stealth. AndreiBot started as a skunkworks project on public GPT-2 models, moved from Hugging Face notebooks to Google Colab for tighter control over training data. Each generation needed source material chosen on purpose, with themes, keywords, and structure pulled out as the outline for new variants.

How it worked

The training set contained more than 1,000 high-performing, compliance-approved marketing emails. That gave the output a reasonable chance of landing inside the voice and compliance boundaries, although the larger test volume still created more review work. The material ranged from climate impact messaging to financial product launches.

AndreiBot was strictly an accelerator. Every AI-generated variant became another option in an A/B test, sitting next to human-written copy. Early variants got rejected often; the success rate climbed as I tuned prompts and training data. Every piece went through the standard review pipeline, bundled into Google Docs as just another test variant.

The workflow: analyze successful emails, break them into components, then play with different emphasis patterns. The value was variant volume on top of proven content, not replacement. Eventually, AI-generated variants performed comparably to human-written copy.

The performance data fed back into the system. Segment-level message analysis sharpened targeting across every email flow: lifecycle sequences, winback, promotional. Every touchpoint became a testing opportunity.

Results

Late 2020 to June 2021, AndreiBot quietly upgraded the testing program. Internal skepticism around AI made it hard to talk about openly. The project shut down when I left, and testing went back to manual. A few months after that, the broader market caught on to AI-assisted content testing.

The tools are easier to prompt now. GPT-2 wanted explicit structure, parameters, and examples, while modern LLMs can work from much more natural instructions. Clear context, specific requirements, and defined constraints still decide whether the output is useful. AndreiBot treated generation as another source of test variants, which remains the part of the experiment I trust.