Visual Abundance and Brand Authority

Prologue

On a fashion retailer’s website, a woman stands beside a window wearing a dark green coat. The room is bright and sparsely furnished. In the next image she turns slightly, allowing the coat to be seen from the side. Two closer views show the collar, the stitching and the texture of the fabric.

The sequence is entirely familiar. Images like these accompany almost every purchase made at a distance. They show an object clearly, place it in a plausible setting and help the customer imagine what it might look like beyond the screen. Most are viewed for a few seconds. Few are remembered.

Nothing about this particular set calls for closer attention. The light is soft, the proportions convincing. The woman appears at ease in the room. There is no obvious reason to question the simple assumption that she was there when the pictures were taken.

Yet the room may never have existed. The woman may never have worn the coat. The light may not have fallen through any window. The garment is real, as is the page on which it appears, but the scene connecting them may have been assembled from a product image and a series of choices made elsewhere.

The result still behaves like a photograph. It shows material, fit and detail. It presents a moment that seems to have happened, even when there may have been no such moment. Nothing in its appearance explains where the photographed object ends and the constructed scene begins.

For the customer, this may make little difference. The coat can still be considered, compared and bought. The image remains useful. It may even be better suited to its purpose than the one it replaced. Seen in passing, the change is easy to miss because the surface of the experience has remained much the same.

What has altered lies behind that surface. The familiar scene no longer reveals much about the circumstances in which it was made. It does not tell us whether there was a studio, a model or a camera. It does not show how many versions preceded it, who chose this one or what kind of work was required before it could appear on the page.

The image offers no guidance on how to interpret this change. It might be viewed as a practical improvement to an ordinary commercial process. It might suggest a wider change in photography, creative work or the way companies present themselves. It might simply be one more technical adjustment that customers absorb without giving it a name.

Each reading begins with the same picture.

That is what makes the picture difficult to place. Its appearance is conventional, while the conditions behind it are becoming unfamiliar. We know what we are looking at, but less and less about what had to happen for us to see it.

Before asking what images of this kind may mean for brands, customers or creative work, we need to understand what has changed in the act of producing them.

What Is Actually Becoming Abundant?

The claim that artificial intelligence is making visual content abundant has an intuitive force. An image that once required a brief, a photographer, a model, a location, equipment, post-production and several rounds of approval can now appear on a screen within seconds. The contrast is striking enough to suggest an obvious conclusion: what used to be scarce is becoming limitless.

Yet an image appearing quickly is not the same as an image being produced quickly. A generated candidate is not necessarily a product photograph, a campaign asset or an approved expression of a brand. Before abundance can be treated as an economic fact, we need to ask what, precisely, has become easier to produce, and where the effort previously involved in production has gone.

Professional visual production has never been limited by one resource alone. Its scarcity arose from several constraints acting together. Budgets determined how many concepts could be developed and how many images commissioned. Time restricted the number of iterations possible before a launch. Physical production depended on the availability of products, people, studios, locations and equipment. Specialist knowledge was spread across photographers, art directors, stylists, designers, retouchers and production teams. Once an image had been made, selection, adaptation, legal review and approval imposed further limits on the amount of content that could actually reach the market.

These constraints shaped more than the price of an image. They determined the logic by which images were produced. Because each additional execution consumed meaningful resources, many decisions had to be settled before production began. Concepts were narrowed, shots planned and variations rationed. Uncertainty was expensive. To explore another idea often meant paying, in one form or another, to make it.

Generative systems loosen the connection between exploration and execution. A textual or visual instruction can produce multiple candidates without the need to rebuild a physical scene. A setting can be altered, a composition proposed, an existing image extended. Variations in format, colour or context can be made without repeating the original production. Reference images, spatial controls and multimodal editing have also made these systems more steerable than earlier text-to-image tools. Technical advances, from diffusion in computationally efficient latent spaces to systems conditioned by poses, depth maps, edges and reference objects, have enlarged the range of work that can take place inside a digital environment.¹

The clearest evidence concerns the early, divergent stages of creative work. Controlled design studies by Fu and colleagues, as well as other research groups, find that generative tools can reduce the burden of visual search, accelerate ideation and increase the number of possible starting points.² Once the work moves from exploration towards the development of a final result, the advantages become less consistent.

In a controlled textile-design study, Shin Young Jang found that a generative workflow reduced perceived workload and improved procedural efficiency. It also produced weaker editability, less expression of personal style and lower assessments on several dimensions of the finished result. Lin and colleagues similarly found no consistent creative advantage over traditional methods in a product-design experiment.³ What emerges from this young body of research is not a general law of creative productivity. It is a more circumscribed observation: generative systems are particularly good at producing possibilities.

Evidence from platform behaviour points in much the same direction. Eric Zhou and Dokyun Lee analysed more than four million works created by over 50,000 users of an art platform. Following the adoption of text-to-image tools, users increased their output by an estimated 25 per cent. Average visual novelty, however, declined.⁴ The study concerns platform art rather than professional brand production, and the adoption of the tools was not randomly assigned. Even with those qualifications, it captures the distinction that matters here. A larger volume of images can be observed without demonstrating an equivalent increase in their quality, originality or economic value.

What is becoming abundant, in the first instance, is the candidate: the preliminary rendering, the alternative composition, the contextual variation, the first approximation of an idea. Producing one more candidate can cost very little compared with organising another physical production. A prompt can be changed without rebooking a studio. A background can be replaced without returning to a location. A new format can be rendered without reconstructing the original scene. For these bounded tasks, generative systems reduce the friction of iteration and widen the field of options.

That change is more substantial than the modest language of “assistance” sometimes implies. The ability to explore twenty versions where only two were previously affordable alters the practical boundaries of creative work. It may also change its sequence. A process that once moved from concept to execution in a broadly linear fashion can become a cycle of generation, inspection, rejection, modification and regeneration.

A qualitative study by Wenyi Chu, David Baxter and Yang Liu observed this shift in a show-production company. Generative AI accelerated creation and iteration and became a medium through which people communicated during production. But what the authors found was mainly a recombination of existing routines, not a complete transformation of the system.⁵ Parts of the process moved faster. Whether the whole process became more productive was not measured.

That distinction matters when technical capacity is translated into economics. The marginal cost of another raw generation is not the total cost of a professional asset. An image intended for use may still require suitable source material, the preparation of references, access to models and computing infrastructure, integration with existing systems, repeated prompting, selection, correction, retouching, rights review, quality assurance, storage and approval. Failed attempts count too, as does the work required to repair them. Cheap attempts do not cease to be costs merely because each one is inexpensive.

The useful measure is therefore not how many images a system can generate. It is how many survive the production process and are approved for use. Public industrial cases often report the first number while leaving the second obscure. They may state how many images were produced in a day, or how quickly a generation was completed, without disclosing acceptance rates, the number of regenerations, the time spent on human review, the extent of post-production or the proportion of assets ultimately published. Technical throughput can rise dramatically while dependable output grows more slowly.

The ABOUT YOU case makes this difference unusually easy to see. At the launch of SCAYLE STUDIOS in July 2026, the company reported that approximately 90 per cent of its e-commerce productions were running through AI-based workflows. It also reported production costs around 90 per cent lower, a reduction in time-to-market of more than 95 per cent and annual savings exceeding eight million euros. Earlier company materials had compared costs of roughly 80 euros per stock-keeping unit with three to four euros, and a seven-week campaign process with a two-day workflow.⁶

These are the company’s figures. They have not been independently audited in the material currently available. Nor does the published information offer a consistent cost boundary, a full account of the measurement method or enough operational data to reconstruct the comparison. Some of the figures also changed between the earlier TAYLA materials and the later launch communication. They are best understood as reported performance within one company’s particular production system, not as general estimates for visual production.

The workflow itself is more revealing than the headline numbers. ABOUT YOU describes a process that begins with product imagery and includes the selection of models, styling, settings and poses. The generated image sets are then retouched, reviewed and approved. This is not a commercially usable result emerging from an isolated prompt. It is a production chain that combines source assets, configurable models, automation and human control. The company also says that high-value campaign shoots continue to be used for selected purposes.⁷ What the case documents is the extensive digitisation of a standardised segment of e-commerce production, not the disappearance of physical photography as a category.

The same pattern can be found elsewhere. Industrial cases reporting large efficiency gains usually concern hybrid systems rather than a stand-alone image model. NVIDIA’s account of its work with Unilever attributes faster and cheaper product imagery to a combination of digital twins, three-dimensional assets and the Omniverse platform. AWS reports that Scenario can generate as many as 100,000 images a day, but does not say how many become final, usable game assets. Google Cloud reports a 400 per cent increase in Dresma’s production volume, again without a complete account of approval rates or total production costs.⁸

Such cases matter because they show that generation at considerable scale is operationally possible. Their limits are equally instructive. Baselines are only partly visible; the boundaries around quality assurance are often unclear. They cannot tell us how much the full cost of professional production falls across organisations, although they make it plausible that substantial savings are possible in certain workflows.

The lack of reliable total-cost comparisons is not a methodological footnote to the abundance argument. It sets the boundary of what can currently be known. Independent, longitudinal studies that measure the same professional workflow before and after the introduction of generative AI remain scarce. A useful comparison would need to include labour, infrastructure, licences, rejected output, rework, legal review and the proportion of images finally used. Without these measures, a large saving in one production step may indicate a large saving overall. It may also be partly absorbed by the system needed to make the output reliable.

Reliability becomes central because visual plausibility and production suitability are different standards. An image can look persuasive at first glance and still fail to preserve a product across several views. It may assign the wrong attributes to objects, misrepresent quantities or lose fine details. Benchmarks such as GenAI-Bench and GeckoNum continue to find weaknesses in compositional and numerical adherence to prompts. ControlNet and newer object-preservation methods improve spatial direction and fidelity to references, but their very existence tells us something important.⁹ Repeatability does not simply arrive with the base model. It has to be built through further inputs, models, data and evaluation.

The work of production changes shape around these requirements. Some manual execution gives way to the preparation of references and controls. Some physical iteration becomes the generation and selection of candidates. Certain forms of retouching may be avoided, while unfamiliar errors require diagnosis and correction. Quality checks can sometimes be automated, but professional approval still has to account for matters generic metrics cannot settle: whether the product is represented accurately, whether the image suits its context, whether a series remains consistent, whether an asset can legally be used and whether it meets the organisation’s own standard.

Scarcity, viewed from inside the production system, has not simply given way to abundance. It has been redistributed. Raw visual options, preliminary concepts and certain kinds of digital variation can be created in far greater numbers. Their marginal cost can fall sharply, especially where the alternative would have required a new physical setup. But usable, controlled and contextually appropriate output remains limited by the quality of the inputs, the reliability of the system and the effort needed to judge what it produces.

This is also why the effect differs so much by task. Replacing a background behind an approved product image poses a narrower problem than generating a complex campaign scene in which the product, people, typography and spatial relationships must all remain exact. A standardised catalogue image lends itself more readily to automation than a production whose value depends on difficult physical access, documentary credibility or precise art direction. In complex, high-value or regulated settings, human expertise and physical production may remain indispensable. Generative capability enlarges what is feasible. It does not dissolve context.

The most defensible conclusion is therefore more restrained than the language of unlimited content, but no less significant. Generative AI makes particular acts of visual generation less scarce. It increases the supply of candidates, accelerates exploration and lowers the marginal cost of certain variations. Professional visual production as a whole does not thereby become uniformly cheap, autonomous or boundless. The cost of obtaining an image is increasingly accompanied by the cost of making it dependable.

Once that distinction is made, another question begins to press. If visual candidates can be produced faster and by a much wider range of organisations, what strategic value remains in production capacity itself? That question only becomes answerable after abundance has been defined with some care.

When Production Stops Differentiating

The ability to produce more images may offer an advantage while that ability is unusual. It cannot offer the same advantage once comparable capacity is available to competitors, agencies, suppliers and small teams using many of the same systems. The images may still be useful. The production savings may still be substantial. What changes is the strategic value of the capability itself.

This is not an automatic consequence of cheaper generation. Production competence has not become uniformly abundant, as the previous chapter showed. Complex campaigns, regulated categories, difficult physical settings and exacting product representation can still depend on scarce expertise and expensive infrastructure. Even highly automated workflows require control. The strategic question is narrower: where the capacity to generate credible visual candidates becomes broadly accessible, what can an organisation still do that others cannot readily reproduce?

Producing at greater volume is the most immediate answer, and often the most seductive. More assets can cover more products, channels, formats, markets and audience segments. Campaigns can be refreshed more frequently. Ideas that once remained unmade for lack of budget can be tested. A retailer with a large assortment may gain real economic value from that reach.

Yet capacity is not the same as distinction. It may improve coverage without changing why a customer should notice or remember one brand rather than another. If several organisations use similar models to generate variations from similar briefs, increased output can enlarge the category’s visual supply without creating a durable difference between its participants. Each organisation becomes more productive. The field becomes more crowded.

That distinction is familiar in strategy. A resource contributes to sustained advantage not simply because it is useful, but because access to it is limited and imitation is difficult. Generative tools can be highly useful while becoming progressively less rare. Model access spreads. Interfaces simplify. Techniques circulate between organisations. Specialist vendors turn once exceptional capabilities into services. Early operational skill may provide a meaningful lead, particularly when competitors have not yet integrated the technology. But a lead based mainly on access or basic proficiency is exposed to diffusion.

The point is not that every company reaches the same level. Adoption quality will differ, sometimes greatly. Data, infrastructure, talent and accumulated learning can remain hard to copy. Production systems may contain proprietary assets and years of operational knowledge. What becomes less defensible is the belief that the mere ability to make synthetic imagery, or to make a great deal of it, will remain distinctive on its own.

Visual abundance also changes the competitive environment in less comfortable ways. It lowers the cost of exploring an idea, but also the cost of approximating an idea already visible in the market. A composition, setting or mood can be described, referenced and varied with little delay. Categories whose visual conventions were already narrow may fill with polished work that is competent at the level of execution and familiar at the level of imagination.

The early research gives some reason for caution. In their analysis of more than four million works on an art platform, Eric Zhou and Dokyun Lee found that users produced more after adopting text-to-image tools, while average visual novelty declined. The setting is not brand communication, and the design does not establish that the tools alone caused every observed change. Even so, the combination is revealing. Higher output and lower average novelty can coexist.¹⁰

Controlled design research points to a related tension. Generative systems can broaden exploration and help people reach possible starting points more quickly. Other studies have found weaker expression of personal style, lower editability or no consistent creative advantage over traditional methods.¹¹ None of this proves that AI-generated brand imagery must converge. It shows why greater technical variety should not be confused with greater strategic originality. A system can produce thousands of visibly different images inside a surprisingly narrow field of ideas.

The danger is not uniformity in a literal sense. The images may vary in colour, location, pose and composition. They may be difficult to classify as copies. What they can share is a more elusive sameness: the same visual fluency, the same borrowed cues, the same tendency to satisfy a prompt without advancing a point of view. Difference between files is easy to generate. Difference between brands is harder.

Volume can deepen this problem within a single organisation. When creating one more option is cheap, fewer ideas are excluded by the cost of making them. That sounds liberating, and often is. It also transfers the burden of exclusion to a later stage. Someone must inspect the alternatives, decide which are relevant, recognise repetition, detect departures from the brand and determine when further variation has ceased to add value.

More possibilities therefore create a peculiar form of scarcity. The organisation is no longer constrained only by how many images it can produce. It is constrained by how many choices it can make well.

Selection is sometimes treated as the lesser creative act, as if the real work lies in producing and the final choice merely identifies the winner. Under conditions of abundance, that hierarchy becomes difficult to sustain. A choice expresses an understanding of purpose, context and consequence. It distinguishes an image that is merely attractive from one that belongs to this brand, for this product, in this moment. It recognises when a technically imperfect direction contains a stronger idea than a polished but generic alternative. It also knows what to reject.

The value of judgement rises because the generator does not bear the strategic cost of its own output. It does not suffer when a campaign dilutes a brand, when a product representation makes the wrong promise or when a succession of individually acceptable images produces a confused whole. Those costs appear elsewhere and later. The person choosing among candidates must consider relationships the model was not asked, or was not able, to see.

This makes the quality of the brief more consequential, not less. If execution becomes easier, vague intentions can be rendered at scale. A poorly framed objective no longer produces only a disappointing image. It can produce hundreds of plausible interpretations, each inviting further attention. Speed magnifies the quality of the question placed before the system.

The same applies to brand knowledge. Many organisations possess guidelines, mood boards and approved assets. Fewer can state clearly which aspects of their visual identity are essential, which may change with context and which conventions they are prepared to break. Traditional production could conceal some of this ambiguity. A small number of experienced people carried tacit knowledge from one commission to the next. They understood what the brand would or would not do, even when the rule had never been written down.

Scaled generation exposes the limits of that arrangement. Tacit judgement does not disappear, but it cannot guide every output personally when the number of outputs multiplies. If a brand wishes to expand visual production without dissolving its identity, it has to make more of its judgement transmissible.

This is where the asset begins to give way to the visual system as the relevant unit of thought. An asset is a finished item: a product image, a social post, a campaign execution. A visual system is the set of principles that allows many such items to vary while remaining recognisably related. It does not prescribe a single look for every purpose. It defines the field within which variation can remain meaningful.

Such a system may include familiar elements such as colour, typography, composition and image treatment. Those are only its visible surface. It also contains decisions about what deserves emphasis, how products relate to people and settings, how realism is used, what forms of imperfection are acceptable, how cultural references are handled and where consistency should yield to context. Its strongest principles are not merely descriptive. They help resolve new situations.

The distinction matters because generative models are adept at reproducing explicit surface instructions. They can apply a palette, simulate a photographic genre or maintain a recurring setting. These controls can support consistency, but consistency alone is not identity. A brand can repeat its colours across thousands of images and still say very little. The more difficult achievement is to encode a point of view without reducing it to a template.

System design is therefore not an exercise in eliminating variation. Excessive control would squander much of the exploratory value the technology offers. Nor is it a licence for unrestricted generation followed by a final brand check. The strategic work lies in deciding where variation is productive and where it becomes erosion. Some properties can remain fixed. Others can change by product, market or channel. A few may be deliberately unsettled to keep the system alive.

That balance cannot be derived from the model. It depends on choices about the brand’s identity and ambitions. The technology expands the possible field; it does not establish which region of that field is worth occupying.

Seen in this light, creativity has not been removed from visual production. Some forms of execution may become less scarce or less valuable economically. Certain tasks may be automated, and some specialist contributions may face genuine pressure. It would be evasive to describe every displacement as a benign elevation towards more strategic work. The distribution of value can shift unevenly, and capabilities that took years to develop may be purchased more cheaply once approximations become acceptable.

Still, the broader strategic conclusion is not that creative contribution ceases to matter. Its centre of gravity moves. When producing a candidate is expensive, making it is a substantial part of the achievement. When candidates are plentiful, more value attaches to defining the problem, establishing the criteria, reading the alternatives and forming a coherent body of work from them. Craft remains important, especially where precision and physical credibility are hard to reproduce. It is joined by a more visible demand for direction.

There is a useful asymmetry here. A weak production process can prevent a strong idea from reaching the market. But once adequate production becomes widely attainable, further improvements in production capacity may contribute less to differentiation than improvements in what the organisation chooses to produce. The threshold is contextual. For a company still unable to create reliable product imagery, production competence may be the immediate strategic constraint. For one already able to generate and approve thousands of assets, the constraint may be whether those assets amount to anything more than coverage.

Human and physical production may also acquire a different kind of value. In some settings, provenance, effort and craft can become part of what an image communicates. A location shoot may matter because the place is real. A photographer’s authorship may matter because a particular perspective cannot be separated from the work. A labour-intensive process may carry meaning in luxury, culture or documentary contexts. As synthetic execution becomes common, such qualities may become more conspicuous.

The present evidence does not justify treating human authorship as a universal premium signal. Consumer responses differ by category, disclosure, perceived quality, existing attitudes and the role assigned to people in the process. Nor should physical production be romanticised merely because it is expensive. Scarcity can support distinction only when someone values what is scarce. Yet these counterexamples are important. They show why the shift away from production is conditional rather than absolute. Production can still differentiate when its method, difficulty or provenance is integral to the meaning of the result.¹²

For most visual work, however, provenance will not rescue an undifferentiated idea. Nor will a proprietary model compensate for a brand that cannot decide what it wants to express. Technical advantage may widen the available space, but the organisation still has to occupy that space deliberately.

This is why selection, judgement and system design belong together. Selection without principles becomes preference. Principles without judgement become rigid rules. Judgement without a system remains trapped in the minds of a few people and cannot travel reliably across a growing volume of work. Strategic differentiation requires the three to reinforce one another: principles make choices more coherent, choices refine the principles, and judgement mediates between them when context resists a rule.

The resulting advantage is more difficult to observe than a production metric. Images per day, cost per asset and time to market can be counted. The quality of a decision is usually visible only through comparison with alternatives and through its consequences over time. This creates a management temptation to optimise what the production system can measure most easily. A rising output curve offers reassurance. A disciplined act of refusal does not.

Yet refusal may be one of the most valuable capabilities in an abundant system. Not every available format requires an asset. Not every segment needs its own visual variation. Not every generated possibility deserves testing. An organisation that cannot stop producing may confuse responsiveness with relevance and personalisation with endless difference. The capacity to impose limits becomes part of creative direction.

Those limits need not be conservative. A coherent system can make experimentation safer because it distinguishes the enduring identity of the brand from the conventions that happen to surround it. Teams can explore more widely when they know what must survive the exploration. Constraints, in this sense, are not remnants of an era of expensive production. They are instruments for navigating an era of inexpensive possibility.

This chapter’s claim remains a strategic synthesis, not a measured law of the market. The evidence shows that generative tools can increase output, that higher output need not increase average novelty and that human filtering remains a consequential part of generative work. Resource logic explains why a useful capability may lose differentiating power as it spreads. Brand and design theory make a coherent case for identity, selection and consistency. What has not yet been established is that organisations with better visual systems will invariably earn stronger customer preference or more durable brand value.

That boundary is essential. A system can be internally coherent and externally irrelevant. It can express the brand faithfully and still fail to interest anyone. It can produce recognisable work while weakening trust in a particular category or context. Strategic discipline improves the odds that abundance will become purposeful, but it does not prove how people will respond.

When another image is no longer difficult to make, the scarce contribution is increasingly the decision that gives the image a reason to exist. Production capacity creates the field of possibilities. Selection, judgement and system design determine whether those possibilities form a distinctive direction or merely add to the visual supply.

The next question is therefore not how coherent such a system can become. It is whether that coherence creates relevance, trust or value in the eyes of the customer.

More Content Is Not More Brand

The distinction between a coherent visual system and an externally valuable one is easy to overlook. Coherence can be judged inside an organisation. It is visible in recurring choices, recognisable codes and the disciplined treatment of variation. Brand value arises elsewhere. It depends on what those choices come to mean to people who encounter them, often briefly and with little interest in how they were made.

An organisation may therefore solve the problem described in the previous chapter and still fail in the market. It may generate a large body of work that is consistent, technically accomplished and unmistakably its own, yet irrelevant to customers. It may even become more efficient at repeating a visual language that no longer carries much meaning.

This is where the argument has to leave the production system and meet the customer.

The first temptation is to look for a general verdict on AI-generated images. Do people accept them or reject them? Can they tell the difference? Does disclosure build trust or damage it? The available research does not provide one answer, partly because these are not one question. It separates into a series of narrower effects, each occurring under particular conditions.

Some experiments find that technically convincing AI-generated advertising images perform similarly to human-produced images when their origin is not disclosed. In a set of studies across coffee, medical aesthetics and public-service advertising, Lei Zhang and Chung Hur found no significant differences in several immediate evaluations under blind conditions. When participants were told how the images had been produced, the evaluations changed. Images described as human-made received higher ratings for trust and purchase intention in that condition.¹³

The image had not changed. Its interpretation had.

That distinction matters because brand communication is never received as pixels alone. People infer intention, effort, credibility and motive from what they see and from what they know about its origin. A photograph can be valued as evidence that a person was present, as a record of craft or as an expression of somebody’s judgement. A synthetic image can be read as inventive and useful, or as an attempt to simulate those qualities without undertaking the work they imply. The production method becomes part of the message when people are made aware of it, or when the image gives them reason to suspect it.

Several recent experiments report a penalty after AI disclosure, especially in measures of authenticity, credibility, attitudes and declared intentions. The pattern is notable, but it should not be mistaken for a universal aversion to synthetic imagery. Most of these studies examine a small number of advertisements in a single encounter. They reveal how a label can influence an immediate judgement.¹⁴ They do not show that all customers reject AI-generated content, that the effect will persist as the technology becomes familiar or that a lower stated purchase intention produces fewer purchases.

Nor is disclosure a single, neutral intervention. A bare label reading “AI-generated” does more than supply provenance. It can suggest that human effort has been removed, that the brand is concealing something artificial or that the viewer should inspect the image for manipulation. A fuller account may convey something different: that AI was used to adapt a background, that a human team directed and approved the work, or that a synthetic person was used to protect a real person’s privacy. In studies of charitable advertising, an ethical reason for using synthetic imagery softened negative reactions.¹⁵ Information about origin cannot be reduced to a box marked transparent or opaque. Its meaning depends on what is explained and why.

The context in which an image appears matters even more.

Consider the difference between an artificial room behind a real dentist and an artificial dentist. Both may involve the same technology and reach a similar technical standard. Yet the person in the second image appears to stand for the human being who will provide the service. The visual is no longer merely atmospheric. It is being asked to establish trust in a real relationship. Experiments in service advertising have found this distinction between synthetic surroundings and synthetic human elements to be consequential.¹⁶

A related sensitivity appears in charitable communication, where an image of suffering is often received as testimony that someone exists and needs help. If the person is fictional, the image may preserve the general truth of a cause while weakening the specific claim on empathy. Luxury poses a different version of the same problem. When effort, authorship and craft form part of the offer, a disclosed shortcut in the production of the communication can sit awkwardly with the meaning the brand is trying to sustain.

Rita To and her colleagues found such a response in studies comparing luxury and mainstream advertising. AI disclosure reduced evaluation, and in one experimental setting clicks, for luxury brands, while the mainstream comparison did not show the same significant penalty. Perceived effort and authenticity helped explain the difference. High perceived creativity also weakened the negative effect.¹⁷ The finding is not that luxury brands must avoid generative AI. It is that production choices become risky when they contradict the source of value a brand asks customers to recognise.

This principle extends beyond category labels. An established heritage brand may carry different expectations from a technology-led entrant. An emotional appeal places a different burden on an image from a rational product explanation. A synthetic model used to show how a garment hangs is not interpreted in quite the same way as a synthetic founder recounting the company’s history. Even within one campaign, some images operate as illustration while others operate as evidence.

Questions of authenticity are often treated as if they concerned realism. They concern congruence more deeply. A flawless image may feel inauthentic because the attributed process conflicts with the brand’s character. A visibly constructed image may feel entirely appropriate when experimentation is part of what the brand represents. Technical quality can remove one source of rejection. It cannot decide whether the method belongs.

Human involvement changes that judgement too. Research on social-media content by Jasper David Brüns and Martin Meißner found weaker perceptions of brand authenticity when generative AI was presented as replacing human content creation. The effect was less severe when AI assisted people. Other studies in fashion and advertising point in a similar direction: participation, direction or a credible human contribution can improve how the work is received.¹⁸

This should not be converted into the comforting claim that adding a person to the workflow protects the brand. Customers rarely see the workflow clearly, and a token claim of oversight may carry little weight. The more useful insight is that people can distinguish between different accounts of agency. Who had the intention? Who made the consequential choices? Who is answerable for what the image represents? In contexts where those questions matter, complete automation and human-directed use are not equivalent propositions.

The same caution applies to commercial performance. ABOUT YOU reports that AI-based image sets increased gross merchandise value by 9.2 per cent and add-to-basket by 5.1 per cent in its own tests. If the tests were well designed, the figures would be commercially important. They remain company evidence. The test protocols, sample sizes, duration and distribution across products and markets have not been published, nor have the effects on returns or repeat purchase.¹⁹ The case demonstrates that scalable production can be connected to behavioural testing in a live retail environment. It does not establish a general conversion advantage for AI-generated imagery.

There are many plausible reasons why an additional image set might improve an immediate retail outcome. It may show a product more clearly, present a more relevant setting, make the assortment easier to navigate or simply replace a weak existing image. Greater production capacity can help a retailer discover and deploy such improvements across a large catalogue. The value lies in the additional relevance created for a particular decision, not in the fact that more content exists.

Volume is an input to that search. It is not the outcome.

This becomes clearer when the possible outcomes are placed in order. An image may attract attention without being believed. It may earn a click without improving product understanding. It may improve add-to-basket while setting an expectation the physical product cannot meet. It may contribute to a sale while making the brand less distinctive. A campaign can perform well during a short test and still teach customers to regard the brand as generic, opportunistic or difficult to trust.

Each step asks a different question. Attention concerns whether the image is noticed. Interaction concerns an immediate response. Conversion concerns a transaction. Trust concerns confidence in the representation and its source. Brand value concerns memory, meaning, preference and the ability to sustain those qualities over time. Movement at one level does not guarantee movement at the next.

Most current studies remain near the beginning of this sequence. They measure attitudes towards an advertisement, perceptions of authenticity, stated purchase intentions or, less often, clicks and engagement. Such measures are useful. Experimental designs can isolate short-term effects that ordinary market data would obscure. But they cannot tell us whether repeated exposure to AI-generated communication strengthens recognition, supports a price premium or changes loyalty over several years. Longitudinal evidence on these questions is largely absent.²⁰

The absence is strategically uncomfortable because production metrics arrive much sooner. Cost per asset, time to market and number of variants can be reported within weeks. Clicks and baskets follow. The slow effects are harder to observe and easier to ignore. A system may receive continual reinforcement for producing short-term gains while gradually weakening the associations that make those gains cheaper to obtain in the future.

This is not an argument against performance measurement. It is an argument against allowing the nearest measure to define the value of the system. Conversion deserves to be measured as conversion. It should not be renamed relevance, trust or brand strength merely because those outcomes are more difficult to establish.

Nor does the evidence support the opposite romantic conclusion. Human production is not inherently more meaningful. A conventional shoot can be generic, misleading or careless. Expense is not proof of thought, and craft has no strategic value simply because it is scarce. Many customers may be indifferent to provenance when an image helps them understand a functional product accurately. In those circumstances, efficient synthetic production may create value without creating a meaningful authenticity problem.

There are, however, contexts in which provenance can become more visible as synthetic production spreads. A real place may matter because the communication makes a claim about being there. A particular photographer may matter because authorship is part of the work. Physical making may support a premium when customers value the skill, time or material encounter involved. The available studies suggest the possibility of such a counter-movement, especially in luxury, design and emotionally charged communication. They do not yet demonstrate a general premium for human-made imagery.

Brands will be tempted to turn provenance into another label. “Made by humans” could become a badge much as other production attributes have become badges. Its force will depend on whether the claim is relevant and credible. Used indiscriminately, it will become one more convention. Used where human presence changes the meaning of the image, it may draw a valuable boundary around what the brand refuses to simulate.

The strategic task is therefore more exacting than choosing between human and artificial production. It is to understand what each image is being asked to do, what customers are likely to infer from it and which form of provenance supports that role. A functional illustration, a product claim, an emotional story and a record of a real event do not carry the same obligations. Neither should they be governed by one assumption about acceptance.

This also changes how a visual system should be judged. Internal consistency remains important, but it is only one condition of value. The system must preserve product truth where customers rely on images for information. It must protect credibility where an image stands as evidence. It must leave room for human intention where intention belongs to the promise. It must vary enough to remain relevant without varying so freely that the brand loses its accumulated meaning.

No model can settle those tensions in the abstract. The research cannot supply a universal rule either. Its strongest contribution is to remove several false ones. Customers do not simply hate AI images. They do not simply accept them when the pixels are good enough. Disclosure does not automatically create trust. A click does not establish brand value. Human provenance is neither irrelevant nor an automatic premium.

What emerges is a conditional account of impact. Scalable visual production can strengthen a brand when it makes communication more useful, more relevant or more responsive while preserving the qualities on which the brand’s credibility depends. It can weaken a brand when additional output becomes noise, when synthetic representation conflicts with the promise being made or when short-term response is purchased at the expense of trust and distinction. Production method and volume shape the possibilities. They do not determine the effect.

That conclusion leaves an awkward burden. Context must be read before images are generated, not discovered only after they have circulated. Quality has to include truthfulness and fit, not just visual finish. Consistency must be maintained across more outputs, while the measures used to assess those outputs must distinguish immediate response from slower brand consequences.

These are no longer tasks that can be resolved by the quality of a single creative choice. Once visual production scales, they recur across products, markets, channels and thousands of decisions. The unresolved question is how an organisation can make that degree of contextual judgement reliable without sacrificing the speed and range it set out to gain.

The Organisation Behind the Image

The organisational problem begins before an image is generated. It begins when someone decides what the image is for, which product truth it must preserve, what the brand can credibly say and which risks the organisation is prepared to accept. Once production accelerates, those decisions do not disappear. They recur more often, across more channels, in more markets and with less time available for each one.

It is tempting to treat this as a tooling problem. A company selects a model, gives creative teams access and adds a review step before publication. That may be enough for isolated experiments. It is unlikely to be enough when generated content becomes part of ordinary commercial operations. At that point, the quality of the output depends on information and decisions distributed across the organisation: product data, brand rules, campaign objectives, customer context, usage rights, local knowledge, legal constraints and performance evidence. The image generator touches only part of that system.

The more useful unit of analysis is therefore the production system behind the image.

Such a system extends from briefing and planning through generation, editing, review, asset management, publication and measurement. In many organisations, these activities already sit in different functions and software environments. Brand teams define identity. Creative teams interpret it. Commerce teams maintain product information. Legal teams assess claims and rights. Technology teams manage access and integration. Local markets adapt work to their own contexts. Agencies and platforms add further layers. Generative production enters this arrangement not as a self-contained replacement, but as a new source of speed and variability inside an already divided chain of responsibility.

That division was easier to tolerate when production itself imposed a slower rhythm. A limited number of shoots, campaigns and adaptations created natural points at which people could align. Scarcity rationed not only images but decisions. When hundreds of variations become technically possible, the old cadence no longer provides the same control. Ambiguities that once appeared occasionally can be reproduced at scale. A weak product description can travel into every generated scene. An unresolved exception in the brand guidelines can become a recurring inconsistency. A rights error can pass through several markets before anyone encounters it as an individual asset.

Speed changes the consequence of organisational vagueness.

Current industrial development reflects this shift. Software providers increasingly connect planning, work management, digital asset management, generation, review, activation and performance data into what they describe as content supply chains. The terminology comes largely from vendors, and claims about the resulting gains should be treated accordingly. Yet the underlying direction is observable. The market is moving from individual creative applications towards linked systems in which assets, rules, permissions and feedback can pass from one stage to another.²¹

What is commercially available today is more modest than the most ambitious product visions. Organisations can automate format changes, background replacements, simple variations, localisation, tagging, routing and some rule-based checks. They can use templates and permissions to let more people produce within defined limits. They can connect approved assets with product data and channel requirements. These capabilities are real, though their end-to-end effect varies with implementation and context.²²

The idea of an autonomous agent interpreting a vague commercial objective, devising a strong creative strategy, resolving rights questions, judging cultural fit and optimising for long-term brand value belongs to a different evidential category. Elements of that future are being announced and piloted.²³ Their reliable organisational performance has not been established. The distinction matters because a roadmap can reveal where an industry is investing without showing that the destination works.

For now, the clearest pattern is not replacement but redistribution. Work moves among people, models and conventional software. Generative systems may produce initial options. Other tools may rank them against formal criteria. A person selects a direction, edits a result or returns it for another cycle. Automated checks can verify dimensions, required elements, metadata or known exclusions. Human reviewers assess product truth, brand fit, novelty, taste and the implications of the image in its intended setting. Approved assets then enter systems that manage publication, reuse and measurement.

This is not simply the old workflow with a faster first step. The boundaries between activities begin to move. Generation and editing can merge into repeated cycles. Review may occur throughout production rather than at its end. Asset management becomes an input to creation because existing images, rights and metadata provide the context for new work. Measurement can influence the next round of variants. The workflow becomes less linear and more recursive.

Human work also changes position within it. Research on generative production consistently finds human contribution in prompting, reference selection, candidate filtering, correction and quality control.²⁴ These findings concern tasks and particular settings, not whole occupations. They offer no sound basis for declaring the disappearance of photographers, designers or art directors. They do show why counting the share of pixels produced by a model reveals little about where value is being created.

A person who once spent much of the day executing a small number of adaptations may supervise a larger field of machine-produced candidates. Another may spend less time creating a first draft and more time defining the conditions under which drafts can be trusted. Creative direction may include translating a visual identity into examples, exclusions and tolerances that a system can use. Production expertise may shift towards diagnosing where a model or workflow fails. Rights, provenance and product accuracy can become continuous operational concerns rather than checks reserved for final approval.

Some activities will lose economic value. Routine adaptations and simple background changes are more exposed than work that depends on contextual interpretation, physical access or distinctive authorship. Certain entry-level tasks may contract, and the path by which people acquire broader creative judgement may consequently become harder to sustain. That possibility deserves attention. An organisation cannot assume that senior judgement will remain available if it automates the work through which future practitioners once learned to exercise it.

New labels for roles are already easy to invent: creative systems designer, brand rule engineer, AI content operator, provenance specialist. Whether such titles endure is less important than the capabilities they assemble. Someone must understand how brand intent becomes operational context. Someone must evaluate the behaviour of models and workflows. Someone must connect creative choices with product, legal and commercial consequences. Someone must recognise an exception that a formal rule cannot capture. These responsibilities may sit in new jobs, in revised versions of existing ones or in teams that work across conventional boundaries.

The organisational challenge lies in the interfaces. Brand knowledge is often tacit, expressed through precedent and the judgement of a few experienced people. Technology teams need specifications. Legal teams need traceability. Local teams need room to respond to context. Commercial teams need speed. A scalable production system forces these groups to make their dependencies explicit.

That work is less glamorous than generation, but more consequential. Product, campaign, channel and rights information needs a shared structure. Approved assets need reliable metadata. Brand principles need to be expressed through rules, references and examples without being reduced to a rigid style filter. Model versions, source material and edits need to be recoverable when questions arise. The organisation needs to know not only which image was published, but how it came to exist and under whose authority.

Provenance standards can help document origin and modification. They cannot establish that an image is true, appropriate or good. A signed record may show which system produced an asset and which changes followed. It cannot decide whether a synthetic person is acceptable in a particular service advertisement, whether a product depiction creates a misleading expectation or whether an otherwise compliant image weakens the brand. Technical traceability supports accountability. It does not replace judgement.²⁵

The same is true of automated evaluation. Machines can check dimensions, logos, colours, text placement and other formal requirements with growing efficiency. They can identify known objects, compare an output with references and filter obvious failures. Professional approval, however, combines several kinds of quality that do not collapse into one score. Technical finish, factual accuracy, brand fit, legal acceptability, cultural sensitivity, accessibility, commercial usefulness and portfolio distinctiveness can point in different directions.

A visually polished image may misrepresent the product. A highly converting variant may erode the consistency of the wider system. A locally resonant execution may violate a global rule that was written for good reason, or expose that the rule no longer serves the market. Quality control is therefore not just error detection. It is the organised resolution of competing criteria.

Human review is often offered as the answer, but the phrase can conceal more than it explains. Placing a person at the end of an automated chain does not guarantee meaningful control. If that reviewer faces thousands of assets, lacks access to their provenance, has no clear standard or cannot stop publication, the human presence is largely ceremonial. High output can turn oversight into rapid confirmation, especially when people assume that a system is reliable because most of its previous outputs were acceptable.

Effective human involvement has to be designed into the workflow.²⁶ Reviewers need to know what they are being asked to judge. They need relevant references, visible changes and enough authority to reject or escalate. The location and intensity of review should reflect the consequence of error. Resizing an approved banner does not require the same scrutiny as depicting a product feature, constructing a realistic person or making an emotional claim on behalf of the brand.

This introduces a second constraint on scale. Human judgement is valuable partly because it deals with ambiguity, but human attention does not expand with machine output. If every generated asset requires the same manual inspection, the review function becomes the bottleneck and some of the expected efficiency disappears. If review is removed merely to preserve speed, the organisation scales its exposure along with its production.

The answer is not necessarily more central control. A central team can maintain platforms, approved models, shared definitions and rights standards. It can also become remote from customers and slow every local decision. Decentralised teams can respond more quickly and interpret context more accurately, yet may create incompatible practices, duplicate work or use unapproved systems. Both arrangements solve one part of the problem by aggravating another.

A federated structure is one plausible response: common infrastructure, risk categories and measurement standards combined with delegated creative decisions inside defined boundaries. Existing enterprise systems already make versions of this possible through templates, permissions and approval paths.²⁷ There is not yet sufficient independent evidence to call it the generally superior model. Organisational history, regulatory exposure, brand architecture and the variety of markets will all shape what works. Some companies may need substantial structural change. Others may adapt established production and governance arrangements without creating a new function.

What cannot remain implicit is responsibility. Automated systems can perform tasks and initiate actions, but they do not assume corporate accountability. When several models and agents contribute to an asset, responsibility can become harder to see just as the organisation becomes more capable of acting. Who set the objective? Who approved the data and model? Who defined the acceptable range? Who owns an error after publication? These are organisational questions even when the immediate failure is technical.

Agentic systems make them sharper. A constrained agent can search approved assets, assemble a briefing, route work, execute a standard chain of adaptations and collect comments. These are plausible extensions of current workflow automation. Giving the same system broad permissions to choose tools, modify content and publish across channels creates a larger surface for error.²⁸ A flawed instruction or compromised document can propagate through the chain. Long sequences make deviations harder to diagnose. Several agents may interact in ways that no single operator fully observes.

More autonomy therefore increases the importance of boundaries, logs, permissions and escalation. It may also deepen dependence on proprietary platforms whose models, interfaces and commercial terms can change. The operational question is not whether an agent appears capable during a demonstration. It is whether the organisation can understand its actions, limit their reach, recover from failure and assign responsibility when the sequence does not behave as intended.

These controls have costs. Integration, data preparation, rights management, evaluation, monitoring and review all consume resources. So does the work of converting brand knowledge into a form that machines and dispersed teams can use. In some repeatable production environments, those investments may be repaid through speed, reuse and lower execution costs. In open-ended or high-risk work, coordination and control may absorb much of the gain. Reliable comparisons remain scarce, especially across complete production systems.

That uncertainty should discourage neither adoption nor scrutiny. It changes what an organisation can reasonably claim to know. A successful trial of background variation shows that background variation can work in that setting. It does not validate the wider operating model. An increase in output shows increased capacity. It does not show that the company has increased its ability to govern that capacity. The appropriate scale of inference is as important as the scale of production.

Organisational learning becomes part of the production system for this reason. Every rejection, exception, correction and performance result can provide information about where the rules, data or workflow are inadequate. Yet feedback is not self-interpreting. A click can reward novelty without revealing whether trust has weakened. Frequent rejection may indicate poor model performance, an unclear brief or an excessively rigid standard. Local deviations may be noise, or they may show that a global principle has been misunderstood.

Learning requires the organisation to retain the context of its decisions. Which hypothesis was being tested? What was compared? Which outcomes mattered? What harm would stop the trial? When should the decision be revisited? Without that discipline, performance data can simply accelerate whatever the system already knows how to measure.

The emerging production system is therefore neither a creative department equipped with a new tool nor a machine pipeline with people left at a few checkpoints. It is a changing arrangement of objectives, information, rules, software, expertise and accountability. Its effectiveness depends on how well those elements work together, especially when no single function possesses all the context required for a sound decision.

That is the deeper organisational consequence of visual abundance. The scarce capability is not supervision in the narrow sense, nor is it the ability to install more sophisticated technology. It is the capacity to distribute judgement without dissolving responsibility, to formalise knowledge without making it inert and to increase speed without allowing the weakest assumption in the workflow to scale unchecked.

Once visual production is understood in those terms, process design reaches its limit. An organisation can connect systems, assign tasks and build review gates, yet those mechanisms still need a direction. They need a basis for deciding which objectives deserve automation, where human judgement must remain decisive, which evidence is sufficient for expansion and which risks the brand should refuse.

The remaining question belongs to leadership: by what principles should this human-system organisation set goals, choose, control, assign responsibility and learn?

Conclusion: Leading the Visual System

The question posed by generative AI is easy to misstate. It can appear to be a question of production: how much more can be made, how quickly and at what cost? Those questions matter, but they do not reach the centre of the change. Once the capacity to generate visual material expands, the harder question is no longer whether an organisation can produce another image. It is whether that image should exist, what purpose it serves and whether it strengthens the body of meaning for which the brand is responsible.

That distinction corrects the most persistent misconception surrounding visual abundance. A greater supply of images is not a greater supply of relevance. Nor is it a greater supply of trust, originality or commercial value. It is a greater supply of possibilities. Possibility is useful, sometimes enormously so, but it is not an outcome.

For much of the history of professional visual production, material constraints performed part of the work of management. Budgets, schedules, studios and specialist capacity limited the number of ideas that could be pursued. Those limits could be frustrating and wasteful. They also forced choices early. When the cost of another candidate falls, the constraint does not vanish. It moves into the decisions that follow generation.

An organisation can now encounter a surplus of plausible answers before it has properly defined the question. It can produce variations faster than it can understand their differences. It can multiply an unresolved ambiguity in its brand, a weakness in its source data or a narrow performance objective across markets and channels. The technology makes execution more responsive. It also makes vague intent more productive.

This is why the centre of competitive advantage shifts without leaving creativity behind. The scarce contribution becomes the ability to give production a coherent direction. It lies in framing the problem well, recognising which possibilities are relevant, judging quality in context and preserving meaning across a growing body of work. These are creative acts, even when they do not leave a visible trace in the final image. Their value is easily missed because an organisation can count assets more readily than it can count sound decisions.

The shift has implications for how a brand itself is understood. A brand cannot be protected at scale as a collection of approved surfaces. Colours, typefaces, compositions and reference images can help a system remain recognisable, but recognition alone is a thin achievement. The deeper task is to make the brand’s point of view usable in new situations: what it notices, what it values, what it promises, what it will not simulate and where it allows context to alter expression.

Some of this knowledge can be translated into rules. Some can only be carried through examples, informed interpretation and the authority to make an exception. Too little definition leaves a generative system without direction. Too much turns identity into a template and removes the variation that made the technology valuable. Leading a visual system requires a living boundary between coherence and freedom, not a final specification of how every image should look.

The organisation behind that boundary matters as much as the model within it. Creative, commercial, technological and legal judgements converge in every asset, whether the structure of the company acknowledges this or not. Greater automation cannot make accountability disappear. It can only make it harder to locate. A system may generate, adapt, test and route content, but the organisation still chooses the objective, sets the acceptable range and bears the consequences of what reaches the market.

Human oversight is therefore not secured by placing a person near the end of a workflow. It depends on whether someone has the context, time and authority to exercise judgement where it can still change the outcome. A nominal approval step offers little protection if the criteria are unclear or the volume of work makes refusal impractical. Responsibility has to be designed into the system with the same care as speed.

None of this points to one ideal organisational form. The appropriate balance of common rules and local discretion will differ across brands, categories and risks. An image that illustrates a functional product does not carry the same burden as an image offered as evidence of a place, a person or an act of craft. The conditions under which automation creates value will vary accordingly. Uniform governance would ignore the contextual character of the very judgement it is meant to preserve.

The same restraint is needed when organisations learn from performance. Faster production brings faster feedback, but not all feedback answers the same question. An image can attract attention, support a transaction and still contribute little to the meaning of the brand. The nearest measurable response may be commercially useful without being a verdict on long-term value. Leadership has to keep these levels distinct, especially when the production system is naturally drawn towards the outcomes it can observe soonest.

Here the methodological caution of the essay becomes part of its executive conclusion. Emerging technology invites confidence before the evidence can support it. A successful task, pilot or company case may justify a decision within its setting. It does not establish a universal principle of customer behaviour, organisational design or competitive advantage. A responsible organisation does not wait for complete certainty, which may never arrive. It keeps the scope of each conclusion proportionate to what has actually been learned.

That discipline is more than intellectual hygiene. It shapes the character of the visual system. Decisions made under uncertainty need room for revision. Exceptions need to remain visible. Measures need to retain the question they were meant to answer. Failures should reveal where an assumption, rule or workflow was inadequate, not disappear inside a rising volume of acceptable output. Learning is not an activity that follows production. It is one of the conditions under which production remains worth scaling.

The executive question can now be answered with some precision. When creating visual content is no longer the central scarce resource, organisations should lead their brands by governing the system that turns visual possibility into public meaning. Their task is to hold production to a clear purpose, make brand principles usable without making them inert, connect judgement with accountability and learn without claiming more certainty than the evidence allows.

This is not a retreat from technological ambition. It is a more demanding account of it. The value of generative AI does not lie in relieving leaders of choices. It lies in enlarging the field within which choices can be made. The stronger the technology becomes, the less credible it is to treat the outputs as its responsibility.

There will continue to be settings in which production skill, physical presence and human authorship remain scarce and valuable. There will also be many settings in which reliable execution becomes ordinary. The durable principle is not tied to where that boundary happens to sit today. Whenever a capability becomes abundant, advantage moves towards the quality of the purposes, judgements and institutions that govern its use.

Visual abundance changes what an organisation can make. Leadership determines what all that making is for.