---
title: "What AI-assisted development looks like when the launch date is real"
url: https://www.axelerant.com/blog/ai-assisted-development-commerce-launch
published: 2026-07-22T10:00:00Z
author: "Panshul Khurana, Staff Software Engineer"
source: Axelerant Thinking
---

# What AI-assisted development looks like when the launch date is real

> How AI-assisted development worked on a real commerce launch: skills in the repo, engineers crossing stacks, and a sprint of rework shipped in two days.

Most AI-in-delivery content is written by people who have never had a launch date. This is not that. This is what happened on [a composable commerce build for a UK kitchen manufacturer](/blog/commerce-platform-manufacturer-d2c-leap), with a real client, real dependencies slipping, and a July go-live that held.

The setup, first, because it is the part most teams get wrong. AI assistance on this project was not individual developers with chat tabs open. The project repositories, frontend and CMS both, carry committed Claude Code skills: codified knowledge about this codebase, its conventions, its integration patterns. Alongside them, project-specific AI guidelines that every engineer onboarding to the project reads, the same way they read the deployment runbooks. New engineers joining the team had "review the Claude Code skills" as a checklist item next to "get repo access."

That framing matters. Prompting skill is personal and leaves with the person. Skills in the repo are infrastructure, and they compound.

### What it changed

Two things happened during this project that would not have happened without it.

First, engineers crossed stack boundaries under pressure. Late in implementation, frontend issues were piling up faster than the frontend team could clear them. A backend engineer whose home turf was Magento picked up frontend tickets, with the committed skills and AI assistance carrying the codebase-specific knowledge he did not have, and helped clear the backlog. In a traditional staffing model, that is a resourcing escalation and a delay. Here it was a Tuesday.

Second, rework got cheap enough to do right. Mid-project we discovered [the 3D planner integration, the core of the buying journey, was built on hardcoded behavior that would not survive the remaining requirements](/blog/magento-3d-planner-integration). The correct call was a rebuild, estimated at a full sprint we did not obviously have. One frontend engineer redesigned it and shipped it in two days. The judgment about what to rebuild and how was entirely human. The mechanical volume of the rework was where AI carried weight. Without that, the honest counterfactual is that we patch the hardcoded version and carry the fragility into production.

There was a third effect, softer but real: frontend feature velocity in the middle sprints ran high enough that the client commented on it. Not on the volume alone, but on how quickly features moved from request to working software.

### What it did not change

Every architecture decision on this project was made by people, argued in review, and owned by a named engineer. AI did not decide to [split content from commerce](/blog/strapi-magento-composable-commerce), did not decide to rebuild the planner integration, did not decide [which data compromises were acceptable for MVP](/blog/ecommerce-launch-with-incomplete-product-data). Quality gates did not move either: the same review discipline, and a QA function that caught multiple showstopper issues before production regardless of how the code was authored.

And it did not make the hard week easy. The final stretch before launch was still integration issues, production readiness, and daily blockers. AI-shouldered means the machine carries load. It does not mean nobody sweats.

### What you can take from this

If you want AI assistance to matter on delivery work, put it in the repository, not in the individual. Commit the skills, write the guidelines, make them onboarding artifacts. Measure the result in the moments that used to be escalations: cross-stack flexibility under load, rework you can afford to do properly, feature velocity a client notices. And keep the decision-making human and named, because that is not the part you want faster.

We are building all our delivery this way now. If you are working out how to do the same, we will show you ours.

---

Read on the web: https://www.axelerant.com/blog/ai-assisted-development-commerce-launch