Most companies have a Responsible AI policy. Almost none have a Responsible AI system.

There is a version of the Responsible AI conversation that everyone is tired of. It lives in ethics committees, in principles documents and keynote slides. It may be a genuine intent, but an intent destined to fail if it remains disconnected from what actually happens when an AI system makes a wrong decision at 3am and nobody knows who owns it.

That disconnection is the real problem, and it is more widespread than most organizations want to admit.

Only 1% of companies implementing Responsible AI initiatives have fully operationalized them. That figure comes from the World Economic Forum, published in 2025. It means the remaining 99% still face the hard work of turning those initiatives into business value, not to mention the vast majority of organizations increasing token consumption without scaling impact. They have policies. They do not have systems.

The gap hiding in plain sight

For two years, most organizations have been either paralyzed or in deployment mode: rolling out chatbots, exploring generative models, automating isolated tasks. That is a necessary first step, but it's not a long-term strategy.

While that was happening, employees relied on personal AI tools, even if IT departments never approved them. As a result, organizations fed proprietary data, tacit enterprise knowledge, and sensitive client information into external models.

The "why" of AI is broadly understood: the efficiency gains, the pattern recognition at scale, the availability. What remains genuinely open, in most boardrooms and team meetings, in most technology functions, in most risk frameworks, is the "how." How do you implement AI in a way that is actually governable? That protects what makes your organization distinct? That doesn't disempower human agency? That can answer for itself when autonomous operations go wrong?

The question no one wants to answer

When an AI system makes a wrong decision, who is responsible?

The answer most organizations would give today, if they were being honest, is: we're not sure. The data scientist built the model. Procurement approved the vendor. Operations deployed the workflow. Legal reviewed the contract. When something fails, everyone points in a different direction.

This is not just a theoretical matter. Air Canada was held legally liable for decisions made by its chatbot. UnitedHealth Group faces ongoing regulatory scrutiny for algorithmic bias affecting hundreds of millions of patients. Documented AI incidents increase year after year, and regulators across the EU, US, LATAM and MENA are moving from guidance to enforcement.

The path to operationalize responsibility

Responsible AI that works in practice, not just on paper, needs to be built across four dimensions that are less a checklist than a chain. Each one makes the next one possible.

It starts with governance, but not the kind that lives in a shared drive. Real governance is the decision architecture that answers: what AI is actually running in this organization right now, including the tools nobody approved? Who decided how each system behaves in situations the original design didn't anticipate? Who has the authority to shut something down? Without a live inventory and a clear ownership model, every other effort is built on a foundation that doesn't exist.

Once you know what you have and who owns it, the next question becomes what you're giving away. Sovereignty is about protecting the organizational knowledge that makes you distinct. As Satya Nadella, CEO of Microsoft, put it at Davos this year, companies that fail to codify their tacit knowledge into their own models will transfer their competitive advantage to third parties. Besides, organizations rarely define formal exit conditions from their AI vendors.

Knowing what you have and protecting what's yours still leaves the hardest problem unsolved: accountability. AI cannot be legally or morally responsible for its decisions. Humans must be. That means designing, before deployment, which decisions require mandatory human review, who holds final authority, and what the escalation path looks like when a model produces something it shouldn't. Accountability is not a retrospective audit function. It is a design principle, and organizations that treat it as the former will keep discovering failures through customer complaints and regulatory findings rather than through their own monitoring.

Which brings the chain to its final link: workforce, or the people who keep the entire system running. Most organizations have not redesigned their roles around AI, adding AI to workflows that didn't change. The result is a technology that delivers far less than it could, because the teams using it are still structured to execute tasks that AI could handle, rather than to govern, validate, and improve the systems doing the executing.

In sum, the operational shift is complete when an organization creates mutual value between human agency and agentic capabilities. Only then does the return on AI investment actually compound over time.

Final words, inspired by Donna Haraway

Ten years ago, before the agentic breakthrough, and even before the publication of the paper on transformer architectures that marked an inflection point in AI research, Donna Haraway was exploring ideas that speak directly to what we now call Responsible AI. That is perhaps unsurprising coming from a professor emerita in History of Consciousness and Feminist Studies at the University of California, Santa Cruz, and a leading voice in science & technology studies.

In her book Staying with the Trouble (2016), Haraway reflects on sustainability and interdependence, inviting readers to cultivate what she calls "response-ability": the capacity to remain present and answerable to one another amid uncertainty, rather than retreating into detachment. For her, this capacity extends beyond the human, entangling our perspective with that of other species and systems we're bound up with, whether we acknowledge it or not.

Less interested in debating whether an LLM is truly conscious, and a bit skeptical about it, to be honest, we would like to borrow her concept to think about the questions that remain unanswered throughout this article.

For organizations navigating AI transformation in 2026, those staying with the trouble of building robust Responsible AI systems, a response-ability mindset may offer a way forward to scale autonomous operations with human agency.

illuminem Voices is a democratic space presenting the opinions of leading Sustainability Thought Leaders, their views do not necessarily represent those of illuminem.

The world needs sustainability knowledge. At illuminem, no interest group or shareholder can influence our work. Thank you for supporting our mission to make high-quality and independent sustainability information free for all. Every contribution helps. Thank you for donating today.