Decisions · Technology
Should loyalty AI decisions be explainable to members
Yes. Explain any AI decision that changes what a member earns, keeps, or redeems. The explanation should state the rule, the member data used, and a plain language reason. Marketing offers that only rank rewards need no explanation. For consequential decisions, opacity costs more than the engineering required.
The decision
The decision is whether to give members a readable account of why an AI system changed their points, status, or redemption outcome. The mechanism is contestability. An explanation lets a member see the operational rule, the data used, and the result. It does not mean showing model internals. A programme that withholds this creates a one-way balance where members must trust a black box. We recommend explanation by default for consequential decisions. Consequential means a decision that changes what a member earns, keeps, redeems, or the status they hold. Decisions that only rank offers or personalise a display are not consequential.
What it costs you
The recommended path costs engineering time, support capacity, and a small amount of operating headroom. You do not give away model weights or fraud rules. You give the reason for the outcome. That protects margin while reducing member distrust and complaint volume.
Say a programme makes 100,000 consequential AI decisions a month. An explanation service that retrieves the rule, the member data used, and the outcome costs 0.25 dollars per decision to generate, store, and serve. Monthly variable cost is 25,000 dollars. Annual variable cost is 300,000 dollars. Fixed integration and design cost is 60,000 dollars one time. First year cost is 360,000 dollars.
If one in five unexplained consequential decisions produces a support contact at 5 dollars each, support cost is 100,000 decisions times 0.20 times 5 dollars, which is 100,000 dollars a month. Explanation at 25,000 dollars a month is cheaper than that support load. The figures are illustrative; your volumes and contact cost set the real comparison.
How to decide
Apply two questions to each AI decision type before launch. First, does it change an amount the member already holds or expects to hold, such as points, status, or a pending redemption? Second, can the member act on the explanation to correct a wrong outcome? If both are yes, explain the decision. The explanation must name the rule, the member data used, and the remedy for an error. If the decision only ranks an offer or personalises a message, it fails the test and needs no explanation. Run this test for points adjustments, redemption declines, status downgrades, expiry notices, and any automated reversal of member value.
What would change our mind
We would reverse this recommendation if a programme measures unexplained decision cost against explained decision cost over two consecutive quarters and finds unexplained cheaper. That comparison must include support contacts, member complaints, dispute handling time, and any regulatory penalty exposure, not just engineering cost. It must also show that explanations do not cause a measurable increase in successful gaming that pushes a programme's margin negative. Until that measurement exists, the default is to explain consequential AI decisions. The threshold is concrete: two quarters of member-level cost data, not a belief.