The judgment layer

At the 2025 Indie Game Awards, Clair Obscur: Expedition 33 won Game of the Year and Debut Game. Two days later, both awards were withdrawn.

The organizers had a firm rule excluding games developed using generative AI. They later learned that the studio had used AI-generated placeholder assets during production. The assets had been removed from the finished game, but their use still meant the title did not qualify.

The decision divided opinion, but it exposed a question that reaches far beyond gaming. Organizations need to determine which uses of AI are acceptable, what must be disclosed, when human review is required, and who has authority to make those decisions.

That boundary is the judgment layer. It is the point where automated execution gives way to accountable human decision-making.

The automation default

Pressure to reduce costs and accelerate delivery has made automation the default across content production, customer communication, market expansion, and internal operations. Many companies are automating first and designing the operating model around it later.

Far less executive attention is being given to where automation should deliberately stop.

Deloitte’s 2026 Global Human Capital Trends research found that organizations which deliberately design human-AI interactions are nearly 2.5 times more likely to report better financial results than those that automate without that intentionality. The emphasis falls on “deliberately.” Sixty-six percent of leaders consider the intentional design of human-AI interactions important to organizational success. Only 7 percent say their organizations are leading in this area.

The advantage goes to organizations that think carefully about where the boundary sits between automated execution and human judgment. It comes from deciding how that technology interacts with people, where decision rights remain, and how responsibility is assigned when the output reaches a customer.

Gaming’s early warning

Gaming encountered this tension earlier than many industries because the commercial and reputational consequences of getting it wrong are immediate and public.

Tamara Tirják, spent a decade leading global content at Frontier Developments and now advises organizations on content strategy, described the objective on a recent episode of the In Other Words podcast.

She draws an important distinction between the velocity challenges that mobile gaming demands and the creative scale of premium titles. Mobile games operate on two-week release cycles, using machine translation and AI for first-pass content that is verified through quality evaluation and contextual checks. Premium titles face a different challenge, where the volume of content can run into millions of words and the creative expectations are considerably higher.

In both environments, automation earns trust when it operates reliably in the background and gives people a clear route to intervene when it does not.

Gaming had to establish these boundaries quickly because its accountability mechanisms are unusually visible. Reviews, refund requests, community discussions, and player behavior reveal failures within hours. Other sectors often have longer feedback loops, but the consequences are no less significant.

Correctness is the threshold

As AI increases the volume of content organizations can produce, the definition of quality must become more demanding.

Content can be grammatically flawless and still fail to connect with its audience. It can be accurate, fluent, and delivered on schedule while missing the cultural cues or emotional register that influence whether someone engages with it.

Correctness is the threshold, but fitness for purpose is the standard that determines whether content performs.

When a campaign launches across several regions with linguistically accurate content and engagement falls in key markets, the underlying problem is rarely grammar. The content may have failed to reflect local expectations or recognize how the intended message changes when placed in a different context.

Riccardo Cocco, Director of Localization at Tripadvisor, illustrated the risk at SlatorCon London. He described an AI system that generated multilingual travel tags containing culturally problematic stereotypes in Spanish and Catalan. The prompts had been designed around American travelers, giving the system no framework for identifying how those assumptions would be received elsewhere.

The system was performing the task it had been given. The failure sat in the judgment surrounding the task. Riccardo warns,

“The cracks are going to show up sooner or later.”

Automation can apply the context it has been given. It cannot be accountable for deciding whether that context is complete, whether the resulting trade-offs are acceptable, or whether an experience is ready for an audience.

Quiet automation

Semih Altinay, Vice President of AI Solutions at Phrase, calls the operating model that addresses this challenge “quiet automation.”

Its value comes from working within trusted systems rather than demanding constant attention as a separate activity. Automation handles repeatable volume, while people remain responsible for intent, risk, and the outcome the organization is trying to achieve. Semih articulated the principle in a recent editorial for MultiLingual Magazine.

The concept draws on a broader principle. As Dr. Meeta Yadav Vouk, former  Vice President of AI and Analytics at Teradata, observed in a conversation on In Other Words,  cited in Altinay’s piece, “AI will be successful when we stop talking about AI.” The measure of maturity is how dependably AI performs without requiring constant oversight, not how many processes it can accelerate.

Pega experienced this transition as it expanded its global content operations. Marketing materials that once followed English-language releases by as much as four weeks now launch in as little as one. Its regional teams processed more than 14 million words over a year, more than double the previous volume, while turnaround times fell by up to 75 percent without increasing the size of the team.

The larger change concerned the role of the people operating the system. Once automated workflows were performing consistently within defined guardrails, the team could spend less time managing individual pieces of content and more time assessing whether the work supported the company’s broader objectives.

The function became more strategically valuable because automation had a defined role rather than unlimited authority.

Designing the judgment layer

Every organization already has a judgment layer, even when it has not been formally identified. It appears whenever someone determines that an output may be technically correct but is not yet suitable for a customer.

Leadership teams can make that layer explicit through four decisions.

Consequence

Identify where an error can be corrected easily and where it could damage customer trust, reduce revenue, create legal exposure, or influence an important decision. The level of human involvement should rise with the consequence of failure.

Context

Define the information a system needs before it can perform reliably. This can include brand standards, market expectations, cultural knowledge, audience intent, approved terminology, and examples of unacceptable output. Context should be treated as operating infrastructure rather than informal knowledge held by individuals.

Authority

Specify which outputs can be released automatically, which require review, and who has the power to pause publication. Escalation should be designed into the workflow rather than dependent on someone noticing a problem at the final stage.

Accountability

Assign ownership for the final outcome. Automated execution does not remove responsibility. It changes where responsibility sits and how early it must be exercised.

These decisions create an operating model in which automation can scale without leaving teams uncertain about when they should intervene.

Judgment has to be built into the system

Technology can outperform individuals on clearly defined tasks. It cannot reproduce the combined knowledge created by people with different backgrounds who challenge one another’s assumptions and build on each other’s ideas.

Tamara described this as the difference between individual efficiency and collective intelligence.

“Technology is more efficient than humans, individuals, for certain tasks, but not more efficient than a group of humans with different backgrounds and approaches, feeding off each other’s ideas and creativity and knowledge.”

That accumulated understanding often sits informally within teams. It includes knowledge of the brand, its customers, previous failures, regional sensitivities, and the compromises the organization is willing to make.

Leaders need to make that knowledge visible. Style guidance, quality criteria, market rules, examples of acceptable output, decision rights, and escalation routes allow human judgment to influence the system before content is produced.

Human involvement delivers greater value when it shapes the operating environment from the beginning instead of serving only as a final inspection.

The leadership decision

The Clair Obscur controversy made the judgment layer visible because gaming’s feedback loops are short and its audiences are vocal. Other industries will face similar scrutiny as AI becomes embedded in everyday content operations.

Customers may never know which parts of an experience were automated. They will still experience the consequences of the decisions made around it.

The strongest operating models will know where automation ends, what evidence is required at the boundary, who can intervene, and who owns the final outcome.

As automation becomes easier to deploy, that judgment layer will become harder to build and more valuable to retain.