Why Auxilium does not use AI for scoring

Auxilium scores members and ministries continuously. None of it uses a model. Every score is the sum of the weights of the rules that matched, and the reasons are always shown beside the number.

This is a constraint we accept costs for, so it is worth explaining.

Three reasons

A score has to be arguable. When Auxilium says a member needs attention or a ministry has an integrity problem, somebody has to be able to disagree specifically — to point at a rule and say "that weight is wrong for us". A model can be interrogated but not argued with, and pastoral and financial decisions should be made on things people can contest.

A score has to be reproducible years later. If a denial is challenged in 2029, you need to show what the system said in 2026 and why. Rules and a stored ledger reproduce exactly. A model that has been retrained since does not.

And the reasons are the product. Knowing a member scores 82 is nearly useless. Knowing they score 82 because a hospitalization request is open, a follow-up is nine days overdue, and four outreach attempts went unanswered tells you what to do this morning.

Where AI would be appropriate

Drafting the note a staff member sends after reading the reasons. Summarizing a long case history for somebody picking it up cold. Suggesting which alias a strange spreadsheet column probably maps to.

The dividing line is simple: AI can help a human read faster or write better. It should not move a number that determines whether a family gets called.

What "explainable" has to mean to be worth anything

The word gets used loosely, and most of what it is applied to would not survive a serious question from a board member. Three things have to be true before an explanation is doing real work.

The explanation has to be the cause, not a story about the cause. A great deal of what is marketed as explainable AI produces a plausible narrative alongside a decision the narrative did not actually drive. If the reasons shown are reconstructed after the fact, they can be convincing and wrong at the same time, which is worse than showing nothing.

The arithmetic has to be checkable by hand. If a staff member cannot add the stated contributions together and arrive at the number on screen, they are being asked to trust rather than to verify, and trust is exactly what erodes the first time the system is wrong about someone they know well.

And the rule set has to be visible in full, not sampled. Showing the top three reasons for one score tells you nothing about what the system systematically ignores. Auxilium publishes every rule to administrators, including ones that did not fire.

Why this matters more in this category than most

Health care sharing ministries operate without the external checks that regulated insurers have. There is no state insurance commissioner reviewing decisions, no statutory appeals process, and no solvency examination.

That absence puts unusual weight on internal accountability. When the only thing standing between a member and an arbitrary decision is the ministry's own process, that process has to be legible to the people running it — and to the member, if they ask.

A scoring system nobody inside the organization can explain does not add accountability. It relocates the arbitrariness and gives it a number.