Generative systems can turn a fragment of team friction into a theme, a proposed action, a research question, a draft ticket, or even a code change. That is the new leverage: interpretation can become executable work with almost no mechanical distance between them.
The same leverage creates an authority problem. If a model’s proposal can silently become a task, a policy, or a permanent record, then “human in the loop” may describe the interface while the system has already made the consequential choice.
Helpful is not authoritative
AI is valuable here because it can perform transformations people do not have time to perform continuously. It can connect a signal to prior history, identify a recurring pattern, or render an ambiguous complaint as a falsifiable experiment. None of those abilities make the output true.
The distinction is easy to say and hard to enforce. Language models produce fluent, complete-looking responses. Product interfaces reward momentum. A neatly written action with an owner and due date already feels more decided than the uncertain signal that produced it.
The more executable the proposal becomes, the more visible the authority boundary must become.
Explanation can increase acceptance
Explanations are often offered as the answer to over-trust: show people why the system made a recommendation, and they will use better judgment. A CHI 2021 study by Bansal and colleagues complicates that assumption. In their experimental tasks, explanations increased the chance that participants accepted an AI recommendation regardless of whether the recommendation was correct. The explanation did not improve complementary team performance beyond the recommendation alone.
The result is not a universal law for every AI interface. It is a warning against confusing legibility with reliability. A plausible rationale can function as persuasion.
Microsoft’s broader human-AI interaction guidelines therefore span more than explanation. They include making clear what a system can do, supporting efficient correction, scoping behavior when wrong, and learning from user behavior over time. Human-centered operation requires recoverability, not just an account of the model’s reasoning.
A proposal-only architecture
Solv expresses the authority boundary as event law. The model can write a proposal. It cannot write the event that makes the proposal team truth.
A person records an observation. It remains attributable and unpromoted.
An AI processor adds a typed suggestion with model and prompt provenance.
A human accepts, rejects, or parks it. This is a separate event with separate authorship.
Only resolved work can enter the chain that later asks for an outcome.
This is stronger than a confirmation dialog. If another interface, export, or automation attempts to route around the resolution door, the kernel rejects the transition. The rejected proposal remains visible as history rather than disappearing into a revised model narrative.
Meaningful human control
NIST’s AI Risk Management Framework treats governance as continuous and cross-cutting. Its core calls for differentiated roles and responsibilities for human-AI configurations and for processes of human oversight to be defined and documented. The Generative AI Profile also recommends evaluating output against known ground truth using human oversight, automated evaluation, and provenance techniques.
For an engineering improvement system, meaningful control has at least five properties:
- Authority: the reviewer can reject without finding a workaround.
- Context: the original signal and relevant evidence remain available.
- Time: the system does not create urgency that makes approval ceremonial.
- Traceability: proposal, decision, and downstream action remain distinguishable.
- Recourse: the team can revisit a decision when later evidence contradicts it.
Human governance is not a claim about human infallibility. People can be biased, inattentive, or politically constrained. The purpose of the boundary is to keep consequential judgment contestable and accountable rather than hiding it inside probabilistic output.
Failure tests, not trust slogans
A governance system should be tested at the points where convenience creates pressure to bypass it:
- Can a processor-authored event ever appear as a resolved commitment?
- Can a high-confidence proposal skip review?
- Can an integration export rejected or parked work as active?
- Can an explanation hide the source signal or contradictory evidence?
- Can a model update rewrite the historical basis of an earlier decision?
Solv’s answer is to treat proposals as powerful but inert. AI can accelerate the journey from signal to possible action. A person still decides whether that journey should continue.