In 2019, U.S. hospitals were the model. Bed utilization in the high 80s. Just-in-time staffing. Inventory leaned out by consultants who promised to unlock margin. It worked. Until it didn't.
When COVID arrived in March 2020, the same systems that posted record margins ran out of capacity in two weeks. The problem was never that we had too few beds. We had exactly the right number of beds for a normal Tuesday, and zero slack to absorb a shock.
I keep coming back to that data point because public agencies, the ones I work with in child welfare and homelessness, are about to do the same thing to themselves. The pitch is AI. The promise is efficiency. The unstated cost is the same one the hospitals paid in 2020.
I have spent the last seven years working inside public systems. The thing that absorbs shocks is friction. Friction is the supervisor who pulls a case off the queue because something feels wrong. Friction is the family resource worker who pauses the intake script because the parent just said something the algorithm was not built to hear. Friction is the second pair of eyes on a kinship placement before the child gets moved.
Friction is the difference between a system that runs and a system that holds.
Here is what worries me about the AI moment in social services. Most of the pilots in child welfare and homeless services do not try to give workers more friction-absorbing capacity. They try to take it away. The pitch to an agency director is always some version of "free your caseworkers from administrative burden so they can focus on higher-value work." That higher-value work shrinks every quarter. The cases the algorithm gets wrong do not shrink. They get harder.
A risk-screening tool can flag scores faster than a human can read them. A vulnerability-ranking system can sort thousands of households before lunch. An eligibility chatbot can handle a meaningful share of inquiries with no caseworker time. Each one looks like progress. Each one is progress on the average case. The average case is not what destroys public agencies.
What destroys them is the case where the algorithm misses. A family scored as low risk. A tragedy that the model did not see coming. A news cycle that takes the whole agency apart for six months. The director is replaced. The funder pulls back. The reform window closes.
That case will happen. Statistically, it has to. The question is whether the agency has any friction left in the system to absorb it.
This is the question I want every funder, every agency director, every philanthropy program officer to sit with right now. Not "how do we get more efficient with AI." But: where in this system do we need friction back?
In our practice, this is what a Program Assessment sprint does. We trace where systems have quietly become brittle, where intentional slack has been audited out, where a workflow that does not see edges has replaced the worker who used to catch them. It is the least exciting consulting product I have ever sold. It is also the most necessary one for the next decade.
The hospital systems that survived March 2020 were not the most efficient ones. They were the ones that quietly preserved slack under intense pressure to optimize it away. Spare beds, cross-trained staff, redundant supply chains. The CEOs who held the line on that slack were not popular in 2019. They were heroes in 2021.
The same calculus is coming for every public agency under pressure to streamline with AI right now. The director who holds back fifteen percent of caseworker time from automation will look slow this year. They will look prescient when the next shock arrives: a policy change, a class-action lawsuit, an adverse outcome the system should have caught, a new administration.
Efficiency is a virtue when the world matches your model. Resilience is the virtue you need when it does not. The systems that serve families do not get to assume the world will match the model.
