All About AI: Matt Rouif of Photoroom

We’re observing a big disconnect (in both our own fashion-only data and in wider benchmarks from other sources) between the increasing capability and reliability of AI, in both a general sense and in discrete applications, and the lack of trust that end users place in its outputs. What does trust in your specific applications of AI look like, and how do you achieve and measure it?

For e-commerce, trust is not really about whether an AI image looks impressive, it is about whether the product that arrives matches the expectations that image creates. Visual content shapes purchasing decisions long before a customer experiences a product firsthand, which means imagery effectively becomes the bridge between intent and transaction. When there is a significant gap between what a customer sees and what ultimately arrives, businesses feel the consequences through returns, customer dissatisfaction and, perhaps most importantly, a loss of trust that is considerably harder to recover than a single sale.

The emphasis we place on product fidelity reflects the fact that, in e-commerce, the commercial success of an image is ultimately determined by its accuracy rather than its aesthetics. The details that influence purchasing decisions are often highly nuanced, spanning colour accuracy, texture, branding, shape and consistency across an entire product catalogue, which means that while generative AI has become increasingly effective at producing visually compelling content, the more consequential challenge lies in ensuring that those images continue to serve as faithful representations of the products they depict across thousands, and in many cases millions, of customer interactions.

Current returns data reinforces why this matters commercially. US retail returns are projected to remain close to $850 billion in 2026, with misleading product images cited by 34% of shoppers as a reason for sending products back. For enterprise retailers running catalogues at scale, that translates directly into financial exposure, which means the question of whether AI-generated imagery accurately represents a product is not an aesthetic one, it is a commercial one.

Ultimately, trust is measured through outcomes. We look at whether imagery supports conversion without creating friction further down the customer journey, whether catalogue quality remains consistent at scale, whether return rates improve, and whether teams feel sufficiently confident in the reliability of the system to deploy it across entire product catalogues rather than isolated use cases.

Image generation has reached what we might refer to as “functional maturity” very quickly. Four years on from the launch of Midjourney and Stable Diffusion, which users needed to either self-host or access through Discord, it’s now trivial for anyone to generate a believable-looking image of a product, either in lay-flat, or as part of a campaign conceived and executed end-to-end with AI.

In that same timeframe, image generation and editing models have become much better at adhering to prompts, and trying to avoid editing critical details about products. But democratising access to those core capabilities has proven to be easier than giving users the tools to control visual accuracy, and now that generated or enhanced images are becoming so ubiquitous in eCommerce and advertising, that accuracy is becoming essential. What do you see happening next here?

In e-commerce, purchasing decisions are often made before a customer reads a single line of copy, which is why visual accuracy has become such an important commercial consideration. Over the past two years, fashion and retail teams have been testing whether generative AI can produce believable imagery and, as the technology has matured, the conversation has gradually shifted away from capability and towards questions of consistency, control and whether these systems can be trusted across entire catalogues rather than individual campaigns or isolated test cases.

That distinction is important because, while the broader market has understandably focused on what AI can generate, the challenge for retailers increasingly centres on whether those outputs can be deployed repeatedly, reliably and at scale without compromising the integrity of the product being represented.

What we are seeing among the organisations moving fastest is that they are not using AI simply to create more imagery. They are building controlled and repeatable production systems that improve the economics of visual content while maintaining the consistency and accuracy required for commercial deployment. Decathlon is a good example of that shift, having reduced image production costs by 99% across 35,000 images in three months, reduced workload by four times and achieved six-figure cost savings, all while maintaining the product fidelity required to support buyer trust. GoodBuy Gear, operating in recommerce, reported a 23% uplift in conversion after deploying Photoroom’s visual production tools, a result that reflects what happens when imagery accurately sets expectations rather than creating them.

A lot of money has flowed into AI. Obviously that’s true for the companies building the technology, but it’s also true for the organisations that are making use of it; licenses and tokens are expensive, and in a lot of cases they’re becoming more costly over time.

There’s been a prevailing sentiment that judging the return on that investment is a new project – that AI operates according to different commercial logic than other technology, or should be benchmarked differently to other production overheads – and that project is becoming much more pressing today as AI budgets undergo more scrutiny. What’s your take on this?

If AI is being used in production, I do not think it should be evaluated according to a fundamentally different commercial logic from any other operational technology. Ultimately, it has to be measured against business outcomes. In visual commerce, the metrics that matter are time-to-market, cost per image, return rate, conversion rate and brand consistency rather than generation speed, model size or aesthetic novelty. Speed only creates value when the output is accurate enough to use commercially because, if imagery misrepresents a product, the downstream impact is felt through returns, customer dissatisfaction and erosion of trust, which means efficiency and fidelity have to be measured together rather than in isolation.

We increasingly think about visual production in the same way brands think about logistics or supply chains: as operational infrastructure with measurable commercial outcomes. The value is not simply that AI can create images more quickly, but that it enables teams to produce accurate, consistent visual content across large catalogues and multiple channels with outcomes that can be measured, scaled and improved over time.

Tied to the question of return on investment analysis is the question of how AI outputs should be billed in the first place. Today, any generative workflow involves a lot of iteration, refinement, and waste. We’ve seen it referred to as “top down” work, rather than the usual “bottom up,” in the sense that anyone can prompt an output that’s 90% of the intended result in a single shot, but the real work – and the real token or credit usage – is the progressive steps that make up the remaining 10%.

As we’ve already said, AI budgets are under the microscope today, so the volume of generations it takes to get to 100% (or close enough to it, depending on the use case) can easily be seen as excess overhead, and neither the model provider nor the workspace host have been keen to rein them in, because extra usage benefits them both.

You’re taking a different approach to accountability and guaranteed outputs, so explain how that works, both for the user and for you, commercially.

One of the challenges that has emerged across generative AI is that many commercial models continue to reward activity rather than outcomes. If customers are paying for every generation, including failed attempts, inaccurate outputs and assets that still require substantial manual correction, then there is an argument that incentives are not fully aligned with the outcome those organisations are actually trying to achieve.

The core distinction in how we build is worth stating plainly: in commerce, the goal is not a beautiful image. The goal is an image that sells. That means product fidelity is not a downstream quality-control problem, it is a design objective built into the model from the start.

Our research among 322 enterprise buyers in March 2026 found that 37% identified inaccurate AI outputs as their primary barrier to adoption, which suggests that the challenge is rarely about model capability alone. Organisations need confidence that the broader production system, including quality control, product fidelity and governance, can consistently deliver outputs that are suitable for commercial deployment. That thinking led directly to the Enterprise Guarantee, launched in June 2026 as the first AI visual guarantee backed by a proprietary fidelity evaluation model. The distinction matters: while others have offered human-reviewed quality checks, the Photoroom guarantee is underpinned by a proprietary fidelity evaluation model, trained, owned and controlled by Photoroom, that evaluates and assesses outputs against agreed product-fidelity criteria, with the End-to-End tier also providing managed correction and redelivery until those standards are met. Photoroom owns its full model stack, from generation through to fidelity evaluation, which means it can stand behind outputs contractually in a way competitors cannot. The programme operates in two tiers: a credit-back guarantee included within qualifying enterprise contracts, and a fully managed end-to-end workflow in which Photoroom regenerates or corrects outputs until they meet agreed standards, in that tier, customers pay only for accepted results.

More broadly, it reflects what we believe will become an increasingly important commercial principle across AI, namely that value should be measured according to usable outcomes rather than usage volume, particularly as enterprise buyers place greater emphasis on accountability, predictability and return on investment.

The immediate appeal of generative photography is its low cost compared to traditional studio or location shooting. On a pure unit economics basis, it’s faster, cheaper, and more flexible to create vertical or lateral image / video coverage through AI than it is to book time and talent. Is this how fashion brands should actually be thinking about AI, though, or does the short-term cost saving risk clouding the potential for something deeper?

While cost is often the catalyst for experimentation with generative photography, the organisations creating the greatest value from AI tend to move beyond the cost-saving conversation relatively quickly because the larger opportunity lies in rethinking how visual content is produced, distributed and maintained across increasingly complex commerce environments. Traditional studio production remains expensive, time-consuming and difficult to scale across every SKU, colourway, market and channel, but the most sophisticated organisations are not treating AI as a cheaper version of the photo studio. Instead, they are treating product visuals as infrastructure: a scalable, brand-controlled production layer capable of supporting every channel, every season and every catalogue while maintaining product integrity and brand consistency.

That shift fundamentally changes the role of imagery from a campaign asset into an operational capability, enabling teams to create coverage for marketplaces, product detail pages, localisation, paid media and emerging formats without rebuilding production processes each time. Depop is a good example of that evolution, having enhanced 1M+ listings within weeks while delivering measurable improvements in sell-through performance. Completeful, a commerce operations business, reported 100% year-on-year sales growth after integrating Photoroom’s API.

What ultimately differentiates the organisations creating the most value is not the volume of AI-generated content they produce, but their ability to build repeatable visual production systems that improve catalogue quality, operational efficiency and commercial performance simultaneously.

Last year, we asked technology executives whether they predicted AI would become more obvious, as a primary interface paradigm, or whether it would disappear into the background in the way that vital but invisible platforms like AWS and Azure have done. This year’s data suggests that the answer is both: fashion professionals that use AI interact with it as both the engine and the steering wheel. But we also see that, for a lot of companies, both kinds of applications are still largely in the scoping or refined-pilot stage. What do you see being the trigger for deeper adoption and wider roll-out?

What ultimately drives adoption of any technology is not novelty but integration, and we are beginning to see a shift from AI being evaluated as a standalone creative tool towards becoming embedded within the production systems that organisations rely on every day. In practice, both outcomes are likely to emerge simultaneously. AI will become more visible where teams require direction, creative judgement and control, while becoming increasingly invisible where it is handling the operational work of adapting, scaling, standardising and optimising visual content. In many respects, the most successful AI systems are often the ones users notice least because they are quietly improving efficiency, consistency and quality in the background.

Most organisations can now see a path towards incorporating AI into visual production. The remaining challenge is less about technical capability and more about confidence in the consistency, governance and reliability required for large-scale deployment. Fashion brands already understand the economics. Traditional photography remains expensive, slow and difficult to scale across every SKU, colourway and channel. The organisations moving fastest are already using AI to improve catalogue quality, increase conversion and expand visual coverage while maintaining control over how products are represented. Wider adoption will be driven by the development of visual production systems that are accurate, repeatable and scalable enough to operate as core commerce infrastructure.