At online travel agency loveholidays, a marketing team recently built an interactive competition microsite in a single afternoon — without waiting on engineering. A product manager prototyped a new way to search for holidays and watched it go live on the company’s website. Neither had to write code the traditional way. Both used OpenAI’s Codex.
That’s the picture OpenAI paints in a customer case study published August 26, 2026, describing how loveholidays has embedded its AI coding agent into workflows for non-engineers — product managers, designers, and commercial staff — turning them into what the company calls “builders.” The case study, told through interviews with two loveholidays executives, is a notable real-world example of what the industry has started calling the “citizen developer” model: using AI to let people without formal engineering training contribute directly to production software.
It’s also, unavoidably, a vendor’s story about its own customer. The headline numbers are striking — an elevenfold jump in AI-assisted code changes, a 73% rise in deployment frequency, tens of thousands of pounds in reported savings. None of them have been independently audited. And loveholidays’ own public engineering writing, when read alongside OpenAI’s account, tells a more layered story than the case study lets on.
Table of Contents
What Happened
loveholidays operates as an online travel agent across eight European markets. According to OpenAI’s case study, the company has been increasingly opening up its codebases to non-engineers, letting product managers, designers, and commercial stakeholders “make changes to our applications, infrastructure, and deploying code,” in the words of Dmitri Lerko, loveholidays’ Head of Engineering.
Mike Jones, the company’s CTO, frames this within a larger strategic goal: building what loveholidays calls “the general intelligence for travel” — combining its existing technology platform with the expertise of its staff, and making both more accessible through AI.
This isn’t a pilot program or a future promise. Per OpenAI’s account, the shift is already underway and producing shipped features.
Search Playground — From Idea to Live Feature
The clearest example in the case study is a tool called Search Playground, built by loveholidays engineers using the company’s existing design system and frontend technologies, with Codex layered on top.
Before Search Playground, according to the case study, anyone outside engineering with an idea for a new customer experience had to convince an engineering team to prioritize building a prototype — meaning every experiment carried a real opportunity cost in engineering time. Search Playground was designed to remove that dependency, letting people across the business turn ideas into working experiences and test whether they deliver value, without entering an engineering queue.
OpenAI reports that more than ten new search experiences have been built through the tool, most of them by non-engineers, with at least three now running live on loveholidays.com. One is “Inspire Me,” a feature that helps travelers explore different types of trips, from beach breaks to food-focused getaways. Another came out of a marketing campaign called “Crisps from Abroad” — an interactive microsite built to collect competition entries and share holiday inspiration. Previously, that kind of project would likely have gone to an external agency, adding cost and delay. Using Codex and Search Playground, the marketing team says it built the experience itself in hours.
“We wanted to decouple our ability to trial new ideas from actual engineering time,” Lerko told OpenAI.
The Numbers loveholidays Is Reporting
OpenAI’s case study includes a cluster of specific metrics attributed to loveholidays. All of them are company-reported figures — supplied by loveholidays to OpenAI for a case study OpenAI is publishing to promote its own product — and none have been independently verified by an outside analyst or auditor. They should be read with that context in mind:
| Metric | Reported Figure |
|---|---|
| AI-assisted code changes (year-over-year) | Grew from 7% to 79% |
| AI-assisted deployment frequency | Up 73%, with engineering headcount held roughly flat |
| Data Platform change success rate | Rose from 58% to 93% |
| Data Platform changes per support request | 4x increase |
| Self-service infrastructure workflow success | Rose from 63% to 90% |
| Cloud storage cost savings | Approximately £36,000 per year |
| Data-processing waste savings | Approximately £100,000 per year |
| New search experiences via Search Playground | 10+ built, 3+ live in production |
Jones frames the intent behind these numbers deliberately: “Technology is just a means to an end. It’s not about the technology itself; it’s about the impact it has. We’re intentional about not just giving people access to tools, but helping them solve business problems — and measuring the impact.”
How It Works — Guardrails and Architecture
loveholidays’ Data Platform and infrastructure, according to the case study, were originally built for technical users — making changes required knowledge of specialist tools, source control, and internal processes that most employees didn’t have. When someone got stuck, an engineer had to step in.
Codex is positioned as a way to scale that specialist knowledge without scaling headcount. Engineering teams, Lerko says, “codify all the best practices” and validations into workflows that Codex then guides other employees through — meaning a non-engineer doesn’t need to understand every underlying system, just what they’re trying to accomplish, while Codex proposes a change, runs checks, and helps move it through the release process.
“The expertise of our data and infrastructure engineers is available through Codex — so anybody self-serving their infrastructure or data needs gets that expertise on tap, 24/7,” Lerko said.
That said, the case study describes this guardrail approach only at a general level. It doesn’t detail exactly what validation, testing, or human review steps stand between a non-engineer’s Codex-assisted change and production — an important gap for anyone evaluating how much oversight actually exists.
The Part OpenAI’s Case Study Doesn’t Mention
loveholidays maintains an unusually detailed public engineering blog, including a running “AI diary” series that predates this case study by nearly a year. Reading it alongside OpenAI’s account adds useful context that the case study leaves out.
The diary shows loveholidays testing multiple AI coding tools starting in early 2025 — GitHub Copilot, a brief trial of Cursor, Google’s Gemini CLI, and OpenAI Codex among them. But it was Anthropic’s Claude Code, introduced internally in April 2025, that the team describes becoming its primary coding tool, citing its terminal interface and output quality.
Later entries in the diary describe concrete guardrail engineering loveholidays built for AI-assisted coding generally — including automated pre-commit checks that run linting, tests, and code-health scoring, plus a safeguard specifically designed to block an AI agent from bypassing those checks. The company also reports that a code-health metric it tracks internally rose from 6.0 to more than 9.75 across a twelve-month period, with over 80% of its teams now above what it considers a healthy threshold. Notably, these specific guardrail mechanisms are described in the context of Claude Code, not Codex.
None of this contradicts OpenAI’s case study. Codex clearly is in active use at loveholidays — the company’s own diary describes a same-afternoon production deployment built with Codex to fix a cloud cost problem that had previously been rejected as too low-priority to fix. But the broader picture is of a company running a multi-tool AI strategy, where Codex is one significant contributor among others rather than the sole engine behind loveholidays’ reported gains. Readers should be cautious about attributing all of loveholidays’ AI-coding progress to Codex specifically.
What Is Codex, Really, in 2026?
Some readers may recall that OpenAI’s original Codex — a code-generation API launched in 2021 — was deprecated in 2023. The Codex referenced in this case study is a different, currently active product.
OpenAI relaunched Codex in April 2025 as a cloud-based coding agent, distinct from the earlier API. It has since evolved into a broader product family — available through ChatGPT plans, a command-line tool, IDE extensions, and a desktop app — running on OpenAI’s current GPT-5.x-Codex model family, including versions like GPT-5.2-Codex. In OpenAI’s own materials, that model is described as its most advanced agentic coding model at the time of its release, with improved handling of large, long-running coding tasks. So while the name is the same, today’s Codex is an actively maintained, frontier-model-backed product — not a revival of the older, discontinued API.
Where Codex Fits Among AI Coding Tools
Codex competes in a crowded field of AI coding assistants. GitHub Copilot, Google’s Gemini Code Assist, and Amazon’s Q Developer (formerly CodeWhisperer) all offer comparable code-generation and agentic coding capabilities aimed at professional engineers. Microsoft Power Apps represents a different, longer-standing approach to non-technical “citizen development,” built around low-code visual tools rather than AI code generation.
Notably, loveholidays’ own case is a reminder that Codex isn’t operating in isolation even within this one company: Anthropic’s Claude Code has, per loveholidays’ own account, played a significant and at times leading role in its AI-coding transformation. Other tools sometimes mentioned in this space, such as Replit’s Ghostwriter, have largely been superseded within their own product lines by newer agent-based tooling, so they’re best understood as legacy references rather than current competitors.
Risks and Open Questions
The most significant limitation of this story is sourcing: every specific metric traces back to loveholidays and OpenAI, with no independent analyst, auditor, or journalist confirmation yet available. That’s a normal feature of vendor case studies, but it means the numbers should be treated as claims, not established facts.
There’s also a real gap in what’s disclosed about oversight. loveholidays describes codifying “best practices” into Codex-guided workflows, but doesn’t specify what review process — if any — a non-engineer’s code goes through before reaching production, nor how that compares to the more detailed guardrail systems (automated testing gates, code-health scoring) the company has documented for its engineer-driven AI coding work.
More broadly, letting employees without formal engineering backgrounds ship production code raises familiar governance questions across the industry: code security, consistency, and the risk of changes that pass basic tests but miss subtler issues. loveholidays’ public materials don’t describe any specific incidents tied to this program, but the general risk is one worth naming for readers evaluating whether “everyone is a builder” is a sustainable model or an early-stage experiment still finding its guardrails.
Conclusion
loveholidays offers one of the more detailed, named examples yet of a company trying to genuinely extend software-building capability beyond its engineering team using an AI coding agent. The reported results — more shipped features, fewer engineering bottlenecks, measurable cost savings — are encouraging on their face. But they come from a single vendor-customer pairing with no outside verification, and loveholidays’ own engineering blog suggests the full story includes tools and guardrail work the case study doesn’t mention.
What would move this from an interesting anecdote to a clearer industry signal is independent reporting or third-party data on outcomes, more detail on how non-engineer-authored code is reviewed before deployment, and a clearer public account of how much of loveholidays’ progress is specifically attributable to Codex versus its broader AI-coding practice. Until then, it’s reasonable to treat this as a genuine, live example of the citizen-developer trend — just not yet a fully independently confirmed one.
FAQ
What is loveholidays using OpenAI Codex for?
loveholidays uses Codex to let non-engineers — including product managers, designers, and commercial and marketing staff — build and modify software, including customer-facing search experiences and self-service infrastructure and data-platform changes, according to OpenAI’s case study.
What is “Search Playground” at loveholidays?
It’s an internal tool, built by loveholidays engineers using the company’s design system and frontend stack plus Codex, that lets people across the business prototype customer experiences and test them without going through an engineering queue.
Is the old OpenAI Codex the same as the current Codex product?
No. OpenAI’s original Codex API launched in 2021 and was deprecated in 2023. The Codex referenced here is a separate product relaunched in April 2025 and now built on OpenAI’s current GPT-5.x-Codex model family.
What metrics has loveholidays reported since adopting Codex?
Company-reported figures include AI-assisted code changes rising from 7% to 79% over a year, a 73% increase in deployment frequency with flat headcount, Data Platform change success rising from 58% to 93%, and roughly £136,000 combined in annual storage and data-processing savings. These are company claims, not independently audited.
Does loveholidays only use OpenAI Codex, or other AI coding tools too?
loveholidays’ own public engineering blog indicates a multi-tool strategy, with Anthropic’s Claude Code described as a heavily used, at times primary, coding tool alongside Codex, Gemini CLI, and GitHub Copilot.
What guardrails does loveholidays use for AI-generated code?
OpenAI’s case study describes loveholidays codifying best practices and validations into Codex-guided workflows at a general level. loveholidays’ own blog separately describes more specific mechanisms — including automated pre-commit checks and code-health scoring — though those are documented mainly in the context of Claude Code rather than Codex.
Are loveholidays’ reported results independently verified?
No independent audit or third-party confirmation of the headline metrics was found. They originate solely from loveholidays and OpenAI.