Sabir Semerkant has been in direct-to-consumer ecommerce since before most of us knew what to call it. He runs Growth by Sabir, and his Rapid2X program is built on AI from the ground up. On this week’s episode he told me he exited about three brands from that program.
Those brands had been soaking up hype. They kept hearing it should all be going faster. Sabir’s answer was blunt: you fix the business side first, then you automate. Otherwise you’re automating garbage.
I’ve watched the same pattern in B2B. A company approves the budget, buys the seats, and expects the tools to sort out a process nobody ever wrote down. When the output is bad, the tool takes the blame.
AI scales what you feed it
Sabir reached for the old computer science line, garbage in, garbage out, and added the part that matters now. AI scales that garbage. He called out data specifically, because for B2B and B2C businesses alike, your life depends on your data. If your attribution data is wrong, the model doesn’t know any better. It guesses with whatever context you gave it.
His analogy was meeting someone at a party. Tell a stranger you’re in fashion and they’ll make assumptions to keep the conversation going. Someone who has worked inside your business for two or three weeks gives you much better answers. AI behaves the same way. The context you put in sets the ceiling on what comes out.
For his members, Sabir has loaded enough context that the AI flags tasks it can’t do itself, because no API, MCP, or CLI supports them, and offers to draft an email or create an Asana card for a named person on the team. That only works because the business was described accurately first.
Build the checks before you trust the output
The second half of readiness is assuming the AI will tell you it’s done when it isn’t. Sabir said it plainly: AI is very confident about completeness.
His audit workflow runs in Claude Code. Sonnet 5.5 on high orchestrates, plans, and makes every fix. It spawns Haiku 5.5 workers for basic operations so it isn’t burning through the plan. When an artifact is finished, two auditors review it: GPT 6.1 Sol and Opus 5.5, both on high, both read-only. They only find problems and report back, partly because output tokens on those models are expensive.
In his example, round one came back at 32 and 42 out of 100. After fixes, round two scored 75 and 82. He treats 85 as a pass and stops at three rounds, which usually lands around 92 or 93. Then the workflow tells him what’s still open and asks whether to keep going.
I explained this on the show as a loop, and loops aren’t new. Organizations in 1950 had human feedback loops for checking work. We use more complicated language for the same idea. Inside our shop we have a skill called Naysayer that code has to pass before moving on, followed by one called Simplifier that cleans up the mess Naysayer created, then Naysayer runs again. These checks stop an error in one step from compounding through every step after it.
Readiness includes your own build hygiene
The garbage isn’t always in the business process. Sometimes we put it there ourselves. I admitted on the show that I found about 600 references to Sonnet 4.6 across 50 repos, hard-coded by coding agents that nobody told otherwise. That model is going away, and those references will start breaking.
The fix is a library, or what Sabir called an LLM config file. Your workflows call a named role, like an orchestrator or a worker, and one file says which model fills it. Change the file and everything follows.
Both of us were firm that you shouldn’t automate the upgrade. Sabir keeps dev, staging, and production separate and tests every new model in dev first. He tried Meta’s Muse as an auditor and it burned through a $50 plan before round one finished. Gemini 4 Argon hallucinated problems that his other auditors confirmed didn’t exist. DeepSeek earned a spot as a third auditor for some workflows in dev, and he said it won’t reach his production server until he fully understands the privacy picture for his data.
If your AI rollout has stalled, look at the inputs before you shop for another tool. Is the process written down? Is the attribution data right? Is anything checking the output before it moves to the next step? Sabir would ask you those questions before letting you into the program.
