What DAM benchmarks miss about how AI interprets brand content

AI readiness starts before a company adopts AI

A company can reject every AI product on offer and still have AI systems explain its products to buyers. Those systems can use public material the company considers outdated, combine it with third-party descriptions, and present the result as fact. For people responsible for brand content, AI readiness therefore starts with whether that content can explain the brand without help from someone who works there.

Digital asset management (DAM), the discipline of organizing, governing, and making content usable, already addresses much of this responsibility. But assessing how employees manage assets answers a different question from how an unfamiliar machine will interpret them. AI readiness adds machine interpretation to content stewardship, regardless of decisions about buying AI tools.

The practitioner behind What Did the DAM Industry’s Benchmarks Miss? identifies this gap in both an established industry framework and their own Brand Management Maturity Model (BMMM). As the BMMM’s creator, they are assessing a model they developed and continue to revise. They propose examining whether content can represent the brand without insider knowledge, while making clear what teams can observe.

Content that works for colleagues can still fail machines

The practitioner encountered this problem among people already preparing content for AI use. At a go-to-market summit dinner in Miami this spring, they joined seven other attendees discussing AI assistants. Participants compared skills they had written, instructions that tell assistants how to perform particular tasks, and debated which material belonged in a knowledge base, a collection an assistant can consult. They wanted to supply enough context for useful work while preventing an assistant from embarrassing them in front of a customer.

Supplying that context took substantial manual work. One participant described spending “the whole weekend” rewriting internal documents that colleagues understood but an AI model could not use effectively. The documents had worked because human readers filled in missing meaning from their own knowledge. Preparing them for a machine required someone to identify what was missing and write it down.

Despite this work, the discussion lasted two hours without anyone using the term “digital asset management.” Participants were assembling small private libraries for machine use while an existing organizational brand library containing 50,000 assets remained untouched. The established library was failing to serve their immediate task, so they created separate collections. Each required decisions about what an assistant should know and which material was suitable.

Those decisions were familiar territory for DAM. The practitioner points to the field’s literature, conferences, and two decades of advocacy for controlled vocabularies: agreed terms used consistently to describe and retrieve content. Through their work with assistants, the dinner participants had taken on the same responsibilities for selecting and governing content. Yet their work remained disconnected from the infrastructure intended to support it.

The disconnect exposes a difference in what readers bring to a document. A colleague can interpret an incomplete description using product knowledge, organizational history, and internal shorthand; a machine lacks that shared background. Content considered “good” and metadata considered “clean” for employee use can therefore leave essential meaning unstated. An asset count reveals little about whether the library supplies the context an unfamiliar reader needs.

Established benchmarks leave machine interpretation unassessed

Because the missing context concerns content quality and governance, a DAM assessment might seem likely to expose it. The practitioner examined the DAM Capability Model, published by the Center for Advanced Studies in Digital Asset Management, and describes it as the industry’s most detailed benchmarking framework. Its 15 dimensions and five maturity levels cover metadata, findability, rights, governance, integration, and reuse. They continue to recommend it for those substantial operational responsibilities.

Yet their inspection found no occurrences of either “artificial intelligence” or “AI.” They report that “Machine” appeared “exactly once,” in a metadata context. The framework predates mainstream use of AI and machine learning, which explains the omission. It remains useful for established DAM work, but leaves machine interpretation unassessed.

The practitioner identifies the same omission in their own BMMM, originally published as a five-part series in 2019/2020. Its five connected dimensions examined whether strategy was documented, whether people owned the responsibility, whether processes were repeatable, how extensively technology was used, and whether impact could be demonstrated. These dimensions assess work employees perform with organizational knowledge behind them. Neither the original BMMM nor the DAM Capability Model asks whether an unfamiliar machine can interpret the content accurately.

Outside AI can represent a brand without its permission

The consequences extend beyond assistants that employees configure. Consider an illustrative case in which an outside answer engine, a system that assembles responses from accessible information, answers a buyer’s product question. It retrieves the brand’s website and PDFs alongside an outdated distributor specification sheet, an old review, and a three-year-old press release naming a discontinued product. The brand cannot configure or audit the system.

Once those materials enter the response, current and stale descriptions can be combined without anyone checking whether each document still reflects the product. The buyer receives a polished answer and may form an inaccurate impression of the brand. This can happen even if the company has never adopted an AI tool. The company can improve its published information and investigate descriptions maintained by others, but those actions cannot determine what the outside system retrieves or how it resolves conflicts.

Internal shadow AI, employees’ use of unapproved AI tools, creates a different governance problem because the organization can intervene in the work. In an earlier discussion in April, the practitioner described employees turning to such tools when the authorized workflow takes three days and the deadline is tomorrow. The need to finish drives them outside the established process. The organization can investigate that process and change how it operates.

For those internal cases, the practitioner recommends identifying the unapproved activity, bringing it into governed DAM work, and adding review where it is missing. Teams can then examine who uses the content and how an output reaches a customer. An inaccessible outside answer engine offers no equivalent point of intervention. The difference is the organization’s ability to inspect and change the workflow producing the answer.

That limit expands a responsibility the practitioner says DAM has spent the previous 20 years addressing: controlling who could use a file, under which terms, and for how long. Access and rights controls remain necessary, but they cannot ensure accurate representation by systems reading information already available to them. The practitioner therefore revises a position expressed in July, when they described a closing opportunity for organizations to choose their response. They now argue that exposure was never optional because outside systems were already reading published material.

For content teams, that exposure creates work on the information those systems can encounter. The practitioner recommends auditing legacy and third-party product descriptions and improving what remains available. The outdated distributor sheet and discontinued-product press release show why the audit extends beyond the current website. Updating accessible information reduces the outdated material available for reuse, while leaving the outside system’s decisions beyond the team’s control.

Assess the content and workflows teams can improve

To assess the work within a team’s reach, the practitioner added AI readiness to the BMMM this year as a sixth dimension. The revision retains the original five maturity levels and introduces criteria for content, policy, review, and governance. Teams can use these criteria to examine how their operations support machine use. Purchasing an AI product, by itself, establishes none of those conditions.

The progression makes content suitability an operational question. Early experiments can produce useful results while leaving input standards and responsibilities undefined. More coordinated work requires dependable content and a policy that guides daily practice, followed by review and governance within the workflow. Teams can add this dimension to their existing maturity review to identify which conditions they have established.

The practitioner also recommends choosing a target that fits the team’s needs. They explicitly reject Level 5 as the appropriate ambition for most teams and recommend Level 3 where it fits. A team with findable content, a written policy governing AI interaction with brand material, and a clear understanding of its risks is in a sound position. Its assessment should also record risks that remain beyond observation.

Within that target, the recommendation to maintain “Clean” metadata requires more than consistent labels. Teams first identify the context employees supply implicitly, then remove dependencies on internal shorthand. They rewrite descriptions so each carries its own meaning and improve the underlying documents that AI systems use. Labels have limited value when the file itself still relies on unwritten knowledge.

Count the interventions that reveal problems

Once employees use that content in AI-assisted work, their interventions provide another way to assess it. The practitioner proposes an AI override rate: how often people correct an AI output or abandon it and complete the task themselves. Counting both captures two different failures. Corrections reveal output that needs repair, while abandonment reveals work the assistant failed to support.

An illustrative image-search example shows why abandonment belongs in the count. A user asks an assistant for an approved product image, receives no satisfactory result, and retrieves the image manually. The user never edits an AI-generated answer, but the assistant has still failed to deliver the content needed. Excluding that intervention would hide a problem with the assisted workflow.

A separate illustrative messaging example shows what correction reveals. An AI-generated product description conflicts with current messaging, and a reviewer catches the discrepancy before it reaches a customer. The review prevents external exposure, while the correction records a failure that still needs attention. A workflow can contain an error and also reveal that its inputs or outputs need improvement.

To make those failures useful, the practitioner recommends recording interventions and tracking the override rate over time. They propose the measure as a signal of underlying content quality and employees’ trust in the AI-assisted process, arguing that decreases depend on improvements to the content beneath it. Teams can use the recorded cases to investigate where retrieval, documents, or messaging break down. The rate gives them an ongoing measure within work employees can observe.

Maturity scores should make uncertainty visible

The practitioner’s benchmarking session this summer illustrates how a score can open that investigation. Scoring took about two minutes, followed by an hour of participants comparing their actual positions, experiments, successes, and failures. The useful detail emerged as people explained what they had tried and where it had broken down. The score gave them a shared starting point for examining the work.

The practitioner compared that exchange with the Miami dinner two months earlier, retrospectively describing its participants as eight strangers. The dinner had produced a similarly useful discussion without scoring. At both gatherings, people learned by comparing concrete experiences and explaining their difficulties. A maturity assessment can make room for that exchange, while participants’ accounts supply the detail needed to act.

For the BMMM’s creator, this purpose also explains the criticism of maturity models designed to make every team feel inadequate until it reaches an offering sold at the top. Such a model gives its seller a commercial incentive to encourage higher scores. The practitioner presents the BMMM’s intended value as an honest account of current practice. A score should help a team choose appropriate work without turning every higher-level criterion into an obligation.

Making that account useful requires separating conditions a team can check from behavior it cannot fully inspect. A team can establish whether it has a taxonomy, an agreed structure for classifying content, as well as documented workflows and enforced governance. It can inspect records and observe the work to answer those questions. External model behavior is less visible.

The practitioner applies this limit to their own framework as well as to others. Internal overrides show what happens where employees can observe and intervene, so a declining rate cannot establish that an inaccessible answer engine represents the brand accurately. A claim of perfect readiness would require knowledge of systems the team cannot fully see or control. Even the highest maturity score cannot reveal what an outside system is telling a buyer.

Final thoughts

AI readiness is a content responsibility before it is a technology investment. Outside systems can already shape how buyers understand a company using information the company no longer considers current. A decision not to adopt AI does not remove that exposure.

For business leaders, the practical task is to establish which conditions the organization can improve and who owns that work. That includes correcting accessible product information, making content understandable without internal knowledge, and setting clear policies for AI-assisted use. Existing DAM capabilities provide a foundation, but operational maturity alone does not establish whether machines can interpret the content accurately.

Investment should follow observed gaps rather than a desire to reach the highest maturity level. Documented policies, dependable content, and review before customer exposure can be appropriate targets. Tracking corrections and abandoned AI tasks can then show where further work is needed.

The assessment must also preserve a clear boundary between evidence and uncertainty. Teams can verify their content controls and inspect internal workflows. They cannot fully verify what an outside system tells every buyer. A useful maturity score makes that distinction visible, helping leaders direct resources without mistaking stronger governance for complete control.