“When you sit across from someone whose nervous system is already overloaded by trauma, homelessness, or withdrawal, asking them to carry the burden of our fragmented infrastructure is not neutral. It is another kind of harm.”
A System That Cannot See Itself
Los Angeles County operates some of the most sophisticated veterans justice programming in the country. Veterans Treatment Courts convene at Los Angeles Superior Court downtown, at Antelope Valley, at Van Nuys, at Compton. At each session, the coalition assembled around a single veteran is genuinely impressive: public defender, prosecutor, probation, the Department of Mental Health, the Department of Public Social Services, the Department of Rehabilitation, VA staff, community treatment and housing providers, law-enforcement crisis liaisons, and custody-based program staff from the Justice-Involved Veterans Division.
On paper, this is what collaborative justice is supposed to look like. The mandate is whole-person care. The intention, at every agency and at every court, is to catch veterans before they fall through the cracks.
The infrastructure holding this together is Microsoft Teams, email threads, and secured text messages.
When a veteran misses a treatment appointment tied to a court order, someone sends a text. When the Department of Mental Health receives a crisis report, someone updates their own system and then tries to catch the rest of the coalition on a call. When JIV staff inside the jail learn something important about a veteran’s stability or risk, that information begins its life as a message, a note, or a phone call that may or may not reach the right people before the next court date.
Each agency is doing the right thing inside its own lane. There is no lane for the veteran’s complete story.
The result is a gap that shows up in very human ways. Veterans are asked to retell traumatic histories again and again because records do not travel with them. A missed message in one system becomes a sanction in another. Families cannot identify who “has the whole picture” — because the truthful answer is that no one does. National research has already documented that justice-involved veterans are among the hardest populations to identify and track consistently, and that self-report alone misses a substantial share of those who served. Los Angeles County’s experience confirms this: every department does what it can with agency-specific case management systems, electronic health records, spreadsheets, shared drives, and a patchwork of commercial cloud services. There is no joint interdepartmental capability that pulls these pieces together around one veteran — and no shared data layer that can produce the cross-system metrics our Board of Supervisors needs to allocate resources wisely.
The Scale of the Problem
To understand why this matters, consider the full range of systems a single justice-involved veteran in Los Angeles County may touch simultaneously:
The Department of Probation and the courts. The Department of Mental Health, substance-use providers, and JCOD re-entry programs. The Department of Public Social Services and housing and shelter vendors. The Department of Rehabilitation. LAPD’s CAMP Unit and the Sheriff’s MET Units. The County Department of Military and Veterans Affairs and its Veteran Peer Access Networks. The District Attorney’s Office and the Public Defender’s Office. Custody-based programs including the Justice-Involved Veterans Division. And the VA itself.
Each of these entities operates under different legal authorities, different data-sharing rules, different intake systems, and different definitions of what constitutes a crisis. The veteran moving through all of them has no ambassador who can see all of it at once. Neither does any individual worker. The social worker in Veterans Treatment Court can see what the court sees. The DMH clinician can see what DMH sees. The probation officer can see what probation sees. The synthesis — the continuous narrative of one person’s journey across all of these systems — exists nowhere except in the veteran’s own exhausted memory.
That is not a technology problem. It is a governance problem. And it is costing us — in human terms, in avoidable crises, in duplicated intake, in recidivism that better coordination could have prevented.
The veteran moving through a dozen agencies has no ambassador who can see all of it at once. The synthesis of their journey exists nowhere except in their own exhausted memory.
The Arriving Wave of Commercial AI Makes This Worse, Not Better
Into this already-fragmented landscape, new tools are arriving. AI chatbots. Document summarizers. Automated case “copilots” wired to private commercial cloud models. These tools arrive with promises of efficiency and insight — and the vendors deploying them have no obligation to understand what a Veterans Treatment Court order is, what JCOD is trying to achieve in re-entry, or what trauma-informed practice requires when a veteran is in crisis.
They were not built to respect the different legal, ethical, and cultural standards that govern Probation, DMH, DPSS, JCOD, law enforcement, and the courts. They were not designed to operate across agency boundaries in ways that comply with the consent and data-sharing frameworks those boundaries represent. They were built to be deployed inside a single system, for a single agency, under a single set of procurement assumptions.
Deploying more commercial AI inside each department’s existing silo does not heal the underlying fragmentation. It risks multiplying it — more portals, more places a veteran’s story is stored without their understanding, more automated flags that reflect one agency’s risk model without any knowledge of what another agency already knows.
The structural critique here is the same one that applies to commercial cloud infrastructure more broadly: when critical public functions are governed by private tools optimized for private purposes, the public interest is an afterthought, not a design constraint. For justice-involved veterans — a population already carrying compounded trauma, navigating systems built for bureaucratic convenience rather than human recovery — the cost of that misalignment is not abstract.
What Governed AI Actually Means
The alternative is not to avoid AI. It is to use a different kind — governed AI — housed in a different kind of infrastructure.
Governed AI, in this context, means AI systems constrained by policy, role, consent, and law at the boundary where they deliver information or trigger actions. It means tools that share only what a specific worker is authorized to see, whether that worker is in Probation, DMH, DPSS, JCOD, VA, JIV, or a contracted housing program. It means every AI-generated summary, alert, or recommendation is logged and auditable across agencies and over time. It means data never leaves a clearly defined, publicly governed environment — one with statutory authority, civilian oversight, and an explicit mandate to serve California’s public interest rather than a shareholder’s.
In operational terms, governed AI applied to justice-involved veterans would do several things that the current patchwork cannot.
It would merge minimum-necessary information from court and probation systems, DMH and JCOD records, DPSS and housing providers, Rehabilitation, the VA and County Military and Veterans Affairs, and JIV program data into one coherent, role-appropriate view — so a veteran is not asked to carry their own case in their head, and so a worker is not reconstructing that case from memory and email before every court appearance.
It would track critical timelines automatically — court reviews, probation conditions, treatment milestones, housing recertifications, benefit applications — so a missed letter does not silently become a violation before anyone on the team realizes what happened.
It would provide trauma-sensitive decision support for law-enforcement crisis units operating in the field, surfacing relevant context in ways that de-escalate rather than inflame.
And it would refuse to answer out-of-scope questions or disclose information outside a worker’s legal authority or a veteran’s documented consent — building the boundary enforcement into the tool itself rather than relying on individual workers to remember the rules under pressure.
This is not AI replacing human relationship. It is AI holding the administrative burden so that humans can do what only humans can do: sit with someone in pain and help them find a path forward.
CalCompute Is the Infrastructure This Requires
California has already started building the environment this approach requires.
SB-53 — the Transparency in Frontier Artificial Intelligence Act, signed into law in 2025 — did more than establish transparency requirements for large AI developers. It also authorized CalCompute: a publicly owned AI computing infrastructure, governed under statute, housed within the University of California system, and explicitly designed to serve California’s public institutions rather than private commercial interests.
CalCompute is not merely a cloud alternative. It is a governance architecture. Data processed on CalCompute stays in California, under California law, subject to civilian oversight and legislative accountability. Decisions about what can be run, what can be accessed, and who can see what are made through public deliberation — not boardroom procurement.
For the coalition of agencies already assembled around justice-involved veterans in Los Angeles County, CalCompute provides something that no commercial cloud arrangement can: a shared, sovereign backbone that all of them can trust.
Imagine a CalCompute-hosted Veterans Pathways Workspace — built under existing legal authorities and appropriate consents — that pulls minimum-necessary data from every system in the coalition into a single governed environment. Access is strictly role-based. A judge, a probation officer, a DMH clinician, a JCOD case manager, a DPSS worker, a peer specialist, and a JIV staff member each see a view tailored to what they are authorized and need to see. Every access is logged. Every AI-generated output is auditable. Nothing leaves the governed environment without authorization.
On top of that backbone, governed AI services running on CalCompute generate concise, role-appropriate situation summaries in seconds rather than forcing each worker to reconstruct context from memory and inbox. They track timelines and surface risks across agency boundaries before those risks become crises. They provide field support for law-enforcement crisis responders without exposing clinical records those responders have no legal basis to access. They produce, from a unified data layer, the cross-system metrics — recidivism, program completion, housing stability, treatment engagement — that our Board of Supervisors has never truly had: a shared, evidence-based picture of how justice-involved veterans are faring across the whole system.
A judge, a probation officer, a DMH clinician, a peer specialist, and a JIV staff member each see a view tailored to what they are authorized and need to see. Every access is logged. Every AI output is auditable. Nothing leaves the governed environment without authorization.
The Case for Los Angeles County as an Early CalCompute Use Case
Justice-involved veterans occupy a particular position in California’s public systems: they sit at the intersection of public safety, health, housing, and justice simultaneously. The agencies that serve them are already in relationship with each other. The will to coordinate exists. What is missing is the shared infrastructure that lets that coordination happen at the speed and completeness that veterans’ lives actually require.
This makes justice-involved veteran services not just a worthy use case for CalCompute — but an instructive one. The fragmentation documented here is not unique to this population. It is the default condition of California’s public-service landscape wherever multiple agencies share jurisdiction over a person’s life. The governed AI architecture that works for a veteran navigating Probation, DMH, DPSS, JCOD, and the VA is the same architecture that works for a foster youth, a person returning from incarceration, a senior citizen managing complex chronic conditions across county and state systems.
If CalCompute is built as a true public utility — and SB-53’s mandate gives California both the authority and the obligation to build it that way — justice-involved veterans and the workers who serve them should be among its first beneficiaries. Their case is complex enough to test the governance architecture seriously. Their coalition is strong enough to implement it with real partners at the table. And the human cost of continued fragmentation is documented, ongoing, and entirely avoidable.
SB-53 made a clear statement: California will not leave AI infrastructure entirely to private providers whose priorities are not aligned with the public interest. The next step is to put that principle to work where fragmentation is costing us most dearly — in human terms, in avoidable crises, in tax dollars spent on duplicated paperwork instead of housing and treatment and peer support.
The infrastructure is now within reach. The coalition is already assembled. What remains is the decision to build the backbone they are all waiting for.