All About AI: Chad Malley of Hyland

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 do you think trust in AI looks like, and what’s the right rubric for measuring it?

Trust can be broken into different levels. The first level is the trust in the source content and data accuracy, as well as relevancy and freshness before feeding into AI services. Large language models (LLMs) are public services that may or may not have granular enough insights to have meaningful predictions for discrete applications. 

Second, the source data and content are often not readily usable as-is by AI models without dedicated enrichment and processing to deliver curated context that’s interpretable by models, otherwise they hallucinate and become indecisive. Teams must trust the data and content has been properly enriched and curated for AI.

Regarding the degree to which humans can trust AI results and how oversight is achieved; work processes must include a level of human in the loop to check AI analysis and predictions for accuracy and error handling. Few AI services are mature enough to be trusted autonomously.

A final trust validation is with the knowledge worker – are they trustworthy in their use of AI? Are they outsourcing key decisions in isolation of the context of the entire workflow? Are they using approved, governed and secure AI services?

Determining the right rubric for measuring trust comes down to experience and outcome verification. Until the results of the AI service and how teams are using it have been validated, stakeholders should remain hesitant to simply relinquish control of their responsibilities to AI.

Beyond that trust, what level of maturity do you see industry-wide in sectors like fashion and beauty that are extremely content-heavy compared to other industries? How far has the typical company come along the journey from experimenting with AI to having operationalised, scaled applications and use cases? And what separates the companies that are further down the road from the ones that are stuck in pilot phases? Are there patterns you see developing that run deeper than trust?

In terms of maturity level comparisons, fashion & beauty are in the middle of the pack compared to other industries AI adoption. There are a few reasons for that. 

Fashion and beauty rely on industry and company-specific vocabularies and ontologies that are critical context inputs for AI to function with high accuracy. However, the source data is fragmented among multiple systems, often with misaligned data models (PLMs, PIMs, MDMs, CRM, loyalty, etc.) On the content side, even more fragmentation exists from shared drives, to SharePoint, to email, to downstream DAMs, Content Management Systems, etc. This content and data model division creates barriers to delivering curated and context-consistent AI interpretation.

Work processes are also often decentralized across business units, product categories, or because of brand acquisitions making implementation more involved.

Fashion and beauty are content heavy, highly visual operations which present challenges for AI. Processing 3D, rich-media and time-based media through AI pipelines requires both dedicated services and a higher degree of enrichment and context curation. Some use cases require a layer of pre-processing and custom training to deliver the most sought-after capabilities – like visual style identification. This upfront investment can be a barrier, so many firms seek more easily achievable wins.

Another inhibitor of adoption is fashion organizations tend to have very creative user bases, and those personas are often reluctant to feel comfortable outsourcing their judgement and creativity. You combine that with pre-defined seasonal calendar pressures, and workflows get reverted back to manual, non-AI augmented tasks.

So, in terms of how far fashion companies have come, I would say many are using AI for general content classification and tagging. Some are creating digital fingerprints to protect their product content and ensure image integrity. However, most have not transitioned to a point where they have developed AI agent pipelines to orchestrate their mission critical processes. AI agents maintain their presence in the call center and in assisting with inventory and demand planning, but they have not evolved into a ubiquitous mesh where heterogenous agents that interoperate across the end-to-end supply chain. This will require significant data integrations and a robust context layer that remain unavailable for many organizations.

We just alluded to the volume of content that fashion and beauty need to create and manage, across different channels, lifecycle stages, and formats. Managing content isn’t just a case of dealing with fixed form finished still photography or video for a couple of channels, and a “digital asset” as we define it today is something more multi-faceted, and can refer to 3D files, design elements, and compound assets. How do you see AI changing the way companies think about their content? And what barriers, if any, are standing in the way of it helping them to scale increasingly complicated content and technology ecosystems?

The volume of content is a pertinent consideration in that it continues to explode, but the more challenging vector is the diversity of content that fashion organizations must contend with. Over 80% of any organizations content is unstructured (video, images, audio, social media, etc). Most AI services are focused on structured data sources, which leaves many organizations to start their AI journeys there. Nonetheless, multi-modal content management is imperative to scale AI adoption across channels and technology ecosystems. 

We see companies testing different models for different content types, but the tendency is to test those different models in isolation and as part of independent use case exploration. What’s missing is a sufficient content layer that provides enrichment, and a context layer that can inform AI sufficiently while maintaining governance and security, while also performing in an environment that allows iterative refinement in the context of actual user flows – not just isolated testing. 

The barriers to scaling AI operations for complex content types are the continued proliferation of content stores and related data sources (PLM, PIM, etc.), inconsistent data models (vocabularies and ontologies), flexibility to leverage a wide range of dedicated AI services (like generating embeddings and enrichment for 3D) in an assorted environment, and the resource justification for an effort that benefits a subset of the process participants. 

Unfortunately, visually rich and complex content types are still a small portion of most companies’ overall unstructured content volume. AI services for these content types are not as highly developed. This means exploration is lagging the AI for text-rich content types. 

We talk a lot about generative AI for content creation, at the beginning and the end of the product funnel. Outside of those use cases, where do you see the biggest impacts of AI in core and adjacent content disciplines like operations, governance, and decision-making? Are we getting better at making pretty pictures and not much else, or is AI already influencing how fashion actually uses the content it’s making? 

The maturity of Agentic processing is having the biggest impact in adjacent content disciplines. Agentic processing can potentially automate much of the manual mundane knowledge work that creatives and content authors typically loathe. 

Agents can also leverage analytic data and surface insights to aid decision handling; exposing opportunities humans miss or simply never had bandwidth to discover.  

Beyond basic content tagging, classification and design-level content generation: agents can recommend popular content, alert authors that content is outdated, validate if the copy is representing the company’s brand voice, apply digital rights enforcement, and ensure the appropriate access controls are maintained.

While agents may exist in vendor supplied platforms, implementing an agentic network of diverse components is likely necessary to amplify the content creation and adjacent content disciplines across the product funnel.

An evolution from there is the potential for dynamic UI presentation based on the AI results and operational findings. This has the potential of fundamentally changing not only the user experience design, but also how content is created, represented, and consumed.

Data comes up a lot where AI is concerned, because it’s increasingly being seen as something that has to be robust, consistent, and centralised. Is that assessment of data, as the primary lever for “AI readiness” correct, or is it too simplistic? And what about the additional scope of compliance information that’s needed in fashion and beauty where brand integrity, rights management and licensing for modelling talent, and so on are concerned? 

Yes, data is a critical component of AI readiness.  Without high-quality accurate data, AI models will produce ineffective or untrustworthy results. The data doesn’t necessarily have to be centralized, but it must be governed and authenticated for compliance & risk avoidance. Security preservation & regulatory safeguards have become equally as crucial as whether a company’s data is clean or not.

In the fashion & beauty industries, strict governance controls are essential. Infringement risks are just one consideration. Data privacy, workforce protection, and consumer trust are equally as significant. Litigation costs and privacy violations have material consequences and create a growing level of exposure & liability for brands of all sizes and market presence.

Beyond data quality and governance, creating additional content can be required to be deemed AI ready. For example, a design or style guideline PPT may be perfectly clear to a human reviewer when evaluating product images – visually depicting what’s ok and not ok. But, to an AI model, that PPT is almost impossible to interpret accurately. Curated tables are required to specify for things such as for this product category, this shot type and this orientation. AI is used to pre-process what’s depicted in the visual (human-facing) guidelines in order to generate AI-consumable artifacts.

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?

We aren’t seeing AI adoption being heavily influenced by whether the services are exposed on the surface or behind the scenes. The more prominent driver is the perceived outcome of the AI service. Investments and resource designations are being allocated to the areas where the biggest business impact is identified. 

AI accessibility and usability are often vendor promoted drivers, but we are seeing organizational drivers like executive sponsorship being the most significant roll-out catalyst.