Product and UX designer Dec Roberts explains how Gousto Ireland used an engineer-built AI “harness” to turn design changes into customer-facing updates faster, without lowering the quality bar.

During an eCommerce Expo 2026 case study Roberts set out a practical case for AI in grocery ecommerce, stressing that it is not a replacement for engineers or UX designers, but as a way to move faster when the right guardrails are already in place.

Roberts shared how the recipe box company’s Ireland team used Claude coding inside a harness (a set of constraints, standards and workflows built by engineers) to redesign parts of its Irish website, including the homepage and pricing information.

Having operated in the UK for nearly 15 years, Gousto launched in Ireland in February 2025. This meant the Irish site, and “squad”, were young enough to experiment, but live enough that any change had to work for real customers.

From backlog pressure to faster releases

The starting point was a familiar challenge for digital grocery teams: customer-facing improvements were taking weeks or months to move from signed-off design to production. Build slots, engineering capacity, support issues and sprint priorities all slowed the process.

Roberts explained: “We always focus on supercharging speed, focusing on quality, maintaining it. We’re still learning. Model share isn’t complete; it’s not 100 per cent perfect. We’ll continue going back to these workflows and iterating, but it’s a step in the right direction.”

Gousto’s Ireland squad had set itself a target to complete much of its product backlog by the end of the year. As that deadline tightened, Roberts and colleagues began exploring whether AI could help reduce waiting time while keeping engineering control intact.

The harness: AI with engineering guardrails

A central theme of Roberts’ talk was that AI alone was not the story, it was the system around it. The harness was not purpose-built AI infrastructure, but “ordinary engineering hygiene” that made generated code safer and more useful.

“The harness is sort of the rails that exist around base,” he noted. “Engineers build it, and that’s usually years before anyone even mentions an AI workflow. It’s what keeps generated code safe. But it also allows non-engineers, designers like me, to run and build in a code base.”

Gousto was not simply asking a chatbot for code snippets, Roberts described agents working within the repository, reading files, making changes, testing them and retrying when something failed. Skills, meanwhile, turned repeatable instructions into team standards. This included how components should be structured; which naming conventions to follow; and what was out of scope.

Subscribe to Grocery Gazette for free

Sign up here to get the latest grocery and food news each morning

Redesigning pricing and the homepage

The team began with smaller changes before moving to more substantial work. One early project focused on Gousto Ireland’s pricing page.

Where the UK site had an interactive calculator, Ireland had a static table that Roberts described as “not very fun” and heavy on cognitive load. Reworking it produced an unexpected commercial benefit too, it helped unlock Google Shopping for the Ireland marketing team because the pricing information could be referenced more effectively.

The homepage redesign was more complex. Roberts gave the agent Figma designs, mobile, tablet and desktop breakpoints, annotations, Jira acceptance criteria and constraints. The work was divided into sections under a feature flag so changes could be switched on and off for customers.

“Giving that early on meant less prompting,” he said. “It was a lot more automated as we went through.”

The workflow also changed how teams reviewed work. Instead of designers handing over static designs and waiting for engineering time, Roberts said the team reviewed working pages earlier and more concretely.

He shared: “Reviewing a live working page felt closer to the end in our column that our customers would actually experience.”

Engineers stayed central

Roberts was clear that the process did not remove the need for engineering expertise. In fact it made that expertise more important because engineers defined the environment in which non-engineers could safely build.

“I didn’t write the code. The harness provided the guardrails, and I just gave it a prompt. The quality of the code came from what engineers had done beforehand,” he said.

That meant engineers were not simply receiving specifications to translate: “They set the constraints so that anyone can build in. That’s much higher leverage than them just translating specs one at a time.”

Lessons for grocery teams

The Ireland squad’s experiment showed strong gains, but Roberts cautioned against scaling too quickly. The team found that faster generation could create a new bottleneck: multiple AI-assisted pull requests landing with one engineer to review them.

The answer was to begin with achievable work, keep projects narrow, and ensure engineers are involved before production code is generated.

Roberts advised: “Start small and work your way up. Build something achievable. You don’t want to just end up with something that’s a prototype. Make one genuine change — end to end, reviewed and merged by an engineer.”

For Gousto Ireland, the measure of success was not just speed for its own sake. Roberts said the team had moved from a process that could take three weeks to one week, with the ambition of working in two-day cycles for some changes. But the broader lesson, she argued, was about deliberate use of AI within a disciplined product process.

He concluded: “I’ve talked a lot about speed, and that is important. But we’ve learned that you need to be deliberate and intentional as well with what our customer experience is.”