Crazy Edge
A suite of AI agents for ecommerce operators — built, shipped, and largely sunset apart from some internal use.
The Problem We Thought We Were Solving
Ecommerce merchants pay agencies significant fees for Google Ads management, SEO, email marketing, compliance monitoring, and reporting. We had been running a healthy agency in this space for over a decade. We believed wrapping AI around those workflows and packaging them as affordable self-serve products to more clients would be an obvious win — faster, cheaper, and more consistent than human-delivered services.
We were wrong about the timing.
What We Built
Seven products shipped under one platform: an AI-assisted Google Ads agent, an SEO and GEO/AEO optimisation agent, email and SMS tooling, automated monthly deep audits, a compliance monitoring agent, a site testing suite, and a brand clone detection tool. Each was a real, functional product — not a wrapper or a demo. Some — including the email and compliance tooling — are still running for clients today, though we’re not actively developing the platform further.
Why It Didn’t Work
Three things converged against us. Google began building AI tooling natively into their ads platform. Shopify launched Sidekick, covering reporting and admin tasks we’d built products to address. The incumbents moved fast and had distribution advantages we couldn’t overcome. Clients weren’t also ready to hand over control to an autonomous agent even in a test.
The market is also saturated with AI noise. Merchants are pitch-fatigued and sceptical — every tool claims to use AI, most don’t deliver, and buyers have been burned. The irony wasn’t lost on us: we were talking to merchants spending thousands a month on traditional SEO that wasn’t moving the needle, offering them a custom content pipeline with GEO built in for a fraction of the cost — and they still weren’t biting. Cost-reduction mode doesn’t always mean willingness to switch.
Our pricing compounded it. Built for enterprise buyers, it didn’t match the budget tier of the merchants most interested, and the enterprises who could afford it had procurement cycles that didn’t suit an early-stage product.
Tech Stack
Multiple – largely using CrewAI, Claude, multiple API’s and LangSmith. Pipelines were all modelled on real world human tasks and run in parallel for several months with refinement steps/feedback loops.
Outcomes
No meaningful commercial traction on the platform as a whole. Several individual products remain in active use. We made the call to stop investing in the platform rather than continue burning resource on a distribution problem we weren’t positioned to solve at scale.
Lessons Learned
Build closer to where the budget already is. The merchants who needed these tools most were the least able to pay, and the ones with budget were being served by platforms with native integrations we couldn’t compete with on distribution.
Timing matters more than quality. Several of these products were genuinely good. Being good isn’t sufficient when the category is noisy, buyers are fatigued, and the platforms you’re complementing are busy eating your use case.
The agent architecture, orchestration patterns, and hard-won understanding of what merchants actually do day-to-day fed directly into everything we’ve built since. It wasn’t wasted — it was expensive R&D.