The way I'm (currently!) approaching this is by implementing the decision engine in 2 steps separated by an evaluation period. I refer to these to the stakehold

The way I'm (currently!) approaching this is by implementing the decision engine in 2 steps separated by an evaluation period. I refer to these to the stakeholders as "user approves" and "user reviews". Essentially, I first implement the engine with the output being an automated list of 1-click actions (typically a set of links on an email, but could easily be a dashboard button or anything else), so the system default is "detect, but don't do". After an evaluation period of the system actions, we move onto the second part, in which the system performs the actions and produces a log of activity, plus a list of 1-click UNDO actions to the user. The idea is that a) the system earns trust from the stakeholders due to their direct involvement, b) the system can enjoy some supervised learning from someone other than the dev team and c) worst case, if it never earns enough trust, I've still saved a stakeholder dozens of hours of work sifting through dashboards as opposed to taking a look at a pre-filtered list. A few systems have turned out well enough that they never move from Phase1 while still being considered a massive win.