Developing the rules and policies for a trust layer that can govern your AI needs consensus among key stakeholders – which is not always simple to obtain. Hidden agendas, power imbalances and plain disengagement can hinder progress.
To overcome these roadblocks, it helps to set some ground-rules for the process of consensus building, while being alert to the priorities of various stakeholders. Where roadblocks prove stubborn, anonymous feedback and voting can help get a working trust layer over the line, so long as these approaches are agreed from the get-go.
Why is consensus needed for trust layer infrastructure?
Trust layer infrastructure is a system to enforce machine-readable rules, policies and guardrails on AI so that it can be deployed at scale, while reducing hallucinations and mistakes. But most organisations quickly discover that key definitions of simple words such as ‘customer’ are not applied uniformly across departments. Sales might apply it to users of a free app, for example, while finance might define it more strictly as someone who pays an invoice.
While humans are adept at finding ways around such muddled concepts, machines are not. Semantic drift has become a major problem as various industries try – and often fail – to get reliable outputs from their AI. Large language models, which work on the basis of statistical probability rather than hard-and-fast rules, are not necessarily the solution either, even though many billions of dollars are being invested in the hope that they are.
Structured knowledge is a good thing for AI accuracy and reliability. But without shared understanding of rules and concepts, an ontological engineer risks building a system that’s based on wrong assumptions, or that doesn’t perform a genuinely valuable function. Only by getting users involved can AI trust layer infrastructure be shaped for the things they need it to do, in the way they want to do it.
What steps build consensus for a trust layer?
Step one is for a moderator – preferably an independent party or even a specialised AI agent – to identify the stakeholders of the domain, and design the process for reaching agreement. Step two is to embark on joint fact-finding, as a means of finding common ground where possible on technical aspects of the domain. One way to break the ice is to ask domain experts three key questions.
Step three is to clarify everyone’s interests; in other words, to understand what each participant needs from the trust layer for it to be useful. This is also a good time to generate potential options to resolve these issues or find common ground, with a focus on new ideas.
Step four involves working on developing a single text, even a simple Google document, that everyone uses as the basis of a plan, rather than working on competing or alternative texts. This can go through multiple iterations. Step five is where decisions are formalised and hopefully implemented.
What are some red flags for consensus building?
A consensus process that goes extremely smoothly is not always a good sign. Sometimes it suggests that nobody actually cares about expressing their disagreement because they have no intention of using the system. In other words, easy consensus can spell poor levels of adoption, especially when discussion is minimal or perfunctory, akin to a ‘box-ticking’ exercise.
In the same way, when stakeholders figure out that a trust layer is going to directly affect their job, their level of commitment to the process becomes markedly higher. But this can make consensus harder to obtain, causing disagreement, lack of trust, and confusion because people are often not speaking frankly about their real concerns.
Inevitably, consensus building ends up addressing the core competencies and raison d’être of the business, enough to justify the participation of senior leadership. Designing a trust layer should not be easy; nor should it be impossible. If it’s that hard, maybe the business itself has serious issues that need to be tackled.
How do I resolve trust layer disagreement?
Consensus approaches such as the Delphi Method can be used to anonymise feedback and resolve stubborn roadblocks that may be caused by internal politics or genuine technical differences. Ontology-as-a-service platforms are emerging that provide digital tools to facilitate the convergence of expert opinion and get a structured trust layer over the line.
Some consensus building approaches are using LLMs with success in the scientific domain, according to a 2025 study authored by Janis Kampers et al. The approach is said to accelerate key tasks while retaining domain expert-in-the-loop validation.
Ultimately however, a moderator – or a digital version of a moderator – may be needed to pilot an effective consensus building process. A moderator can navigate stakeholder special interests, foster a spirit of collaboration, and focus minds on the objective of the process, which is to develop AI systems that people will not only trust but enjoy using because it makes their lives easier and more productive.