Explicit invocation: analysis begins only after the user opens a specific Gmail thread and invokes Minerva.
Useful decision support makes responsibility clearer—not easier to evade.
Minerva is designed to show its evidence boundary, uncertainty, dissent and reversal condition while reserving the final judgment and every external action for the executive user.
The control points that keep assistance from becoming delegation
Source trace: the decision record distinguishes material present in the thread from inference and missing facts.
Visible dissent: alternative interpretations and stakeholder consequences remain inspectable before commitment.
Reversal condition: the recommendation states what evidence or event should cause the decision owner to reconsider.
Human confirmation: the executive can select, edit or reject the recommendation and next action.
Controlled action: Minerva does not search the mailbox or automatically send the reply; the user remains responsible for external action.
See the mechanism tested, case by case
The evidence library contains 50 independent teaching simulations run through the actual Minerva Decision Room. The named consulting firms are public-source references, not Minerva customers or endorsers.
McKinsey & CompanyChemical Pricing: Unified Algorithm or Frontline Judgment?First run a pilot combining analytical recommendations with frontline review on '1–2 product lines × a small group of non-critical customers,' rather than repricing across the board. The core is to design it as an experiment that can be stopped, attributed, and reversed — not a contest to validate the algorithm.KPMGMachine Learning Governance at a Global Payments Company: Full Rollout, or a Risk-Tiered Pilot First?Prioritize a scoped, reversible risk-tiered governance pilot: given the lack of evidence on current coverage and control costs, focusing first on validation practices for potentially higher-impact models is more controllable and reversible than a simultaneous full rolloutghSMARTPharma CEO Succession: Commit to a single candidate five years out, or keep a candidate slate and converge gradually using strategy, board, and readiness gates?Recommendation: retain the candidate slate and converge gradually through strategy, readiness, and governance gates, as this preserves adjustability over the five-year uncertain horizon while balancing fairness and retention (F1, F2, F3, F6)
Questions executives and evaluators ask
Does Minerva make the executive decision?
No. It reconstructs context, separates evidence from inference, challenges the recommendation and records reversal conditions. The executive selects, edits or rejects the recommendation and owns the action.
What information does Minerva use?
For the Gmail add-on flow described here, Minerva uses the current thread the user explicitly opens and invokes. It does not search the wider mailbox.
Are the 50 evidence cases customer results?
No. They are independent teaching simulations based on public sources and actual product runs. They do not claim affiliation, endorsement, customer outcomes or ROI.
Where it fits：Minerva reads only the Gmail thread the user deliberately opens and invokes. It does not search the mailbox, decide autonomously, or send an external action without the human user.
Keep the speed of AI without transferring accountability to AI.
Evaluate Minerva by whether it makes uncertainty, disagreement and ownership more visible—not by whether it sounds confident.