AI in Homeless Services: My Honest Take After a Year of Testing

The Hype

A vendor promised us AI that would predict which clients were most likely to succeed in housing. "Prioritize your limited resources," the pitch went. "The algorithm knows who to help."

I was skeptical but curious. We ran a pilot.

What Happened

The algorithm worked, technically. It predicted housing success with reasonable accuracy. But when we dug into the predictions, patterns emerged that bothered us.

The model favored people who looked good on paper: steady income history, no criminal record, shorter homeless episodes. These are exactly the people who are already easiest to house. We didn't need an algorithm to tell us that.

What we needed was help with the hard cases—people the model flagged as "unlikely to succeed." Those people didn't become less deserving just because they were harder to help.

What Actually Helps

After a year of experimenting, here's what I think AI can usefully do in homeless services:

• Administrative stuff: Transcribing notes, scheduling, routing referrals. Boring but time-saving.
• Identifying gaps: Which clients haven't been contacted in 30 days? Who fell through the cracks?
• Pattern recognition: What do successful housing placements have in common? What are warning signs of returning to homelessness?

What it can't do: replace human judgment about who deserves help.

My Advice

If a vendor promises AI will solve your problems, ask them: what happens to the people the algorithm deprioritizes? If they don't have a good answer, walk away.