AI can summarise a meeting. That does not mean it understands governance.
A summary can be useful. It can reduce the effort required to review a long discussion, identify themes or retrieve relevant passages. But governance depends on distinctions that are easy to flatten if a system is optimised only for fluency.
Was a formal decision actually made? Was an issue deferred? Was approval conditional? Which authority applied? Was a statement part of the approved record or generated interpretation? What evidence supports the answer? Who is permitted to see it?
In governance, an answer that sounds confident can still be unsafe if its source, authority or uncertainty is unclear.
Useful intelligence needs visible provenance
Chaird’s position is that material AI-assisted answers should remain connected to accessible source information.
If a user asks why a decision was made, the system should not produce an elegant narrative that cannot be traced back to the relevant paper, minute, resolution or later evidence. The user should be able to see the basis for the answer and distinguish source material from interpretation.
That principle matters for ordinary retrieval as much as it does for complex analysis. Governance information can carry consequence. A generated statement should not quietly acquire the status of an approved record simply because it is well written.
Permissions apply to generated outputs too
Governance information is rarely uniform. Different forums, entities, matters and documents can have different confidentiality and access boundaries.
An AI layer should not become a way around those boundaries.
If a person cannot access a source document, a generated answer should not reveal the restricted information contained within it. Permissions need to apply consistently to both source material and the outputs derived from that material.
This becomes especially important when an organisation connects information across multiple governance systems. Better retrieval should not mean broader access by default.
Uncertainty should be a feature, not a failure
Generative systems are often rewarded for providing an answer. Governance sometimes requires the opposite behaviour.
A trustworthy system should be able to say that it cannot determine whether something was a formal decision, that relevant sources conflict, that evidence is incomplete or that human confirmation is required.
That is not a weakness. It is an important control.
The objective should be appropriate confidence, not maximum confidence.
Approved record and generated interpretation are different things
Minutes, resolutions, delegations and approved papers have a status that generated content does not.
An AI-generated summary may be useful for orientation. A generated comparison may reveal an important pattern. A suggested answer may accelerate research. None of those outputs should silently overwrite or become indistinguishable from the underlying governance record.
A governed system should make the boundary visible.
That creates a clearer operating model for people using AI in consequential settings: source records retain their authority, generated interpretation remains identifiable and human users decide what should be relied upon or formally adopted.
Human authority remains central
Chaird is not pursuing an idea of autonomous governance.
Boards, committees, executives and governance professionals remain responsible for the judgements they make. AI may assist with retrieval, comparison, context and interpretation, but it should not quietly assume authority that belongs to people or formal governance bodies.
This is particularly important because governance decisions often involve trade-offs rather than objectively correct answers. Evidence can inform judgement without eliminating the need for it.
Traceability creates a better user experience too
Trust controls are sometimes presented as friction. In practice, visible provenance can make an AI experience more useful.
A director preparing for a meeting may want to move quickly from an answer to the relevant source. A Company Secretary may need to verify whether a statement reflects an approved resolution. An executive may want to understand how a current recommendation differs from a prior decision.
When citations, chronology and permissions are part of the experience, the user can investigate rather than simply accept.
That is a better fit for governance than a black-box answer.
What accountable intelligence could look like
For governance use, Chaird is testing a set of practical principles:
- Sources remain visible and accessible where the user has permission.
- Permissions are respected across both records and generated outputs.
- Approved records and generated interpretation remain distinguishable.
- Uncertainty and conflicting evidence are communicated honestly.
- Material answers can be traced to evidence and chronology.
- People remain accountable for governance decisions.
These are design positions, not claims that every implementation question has already been solved. The exact controls, workflows and integration boundaries still need to be tested with practitioners and organisations.
The question is not whether governance will use AI
The more useful question is what conditions should apply when it does.
Three questions are worth putting to any governance AI experience:
- Can a user understand where a material answer came from?
- Would the same permission boundaries apply if the answer were generated from several restricted sources?
- Can the system communicate that it does not know, rather than converting ambiguity into certainty?
AI can make governance information easier to use. The standard should be higher than convenience alone.
The goal is better-informed human governance, supported by intelligence that is traceable, permission-aware and reviewable.
Chaird is testing these principles with people who understand governance in practice and welcomes perspectives on what should be non-negotiable.