On a Tuesday morning in a county human-services building, five people walk past each other in the same hallway and never think of themselves as having the same job. One is a child-protective-services investigator headed out to a first-contact home visit with a screener's intake form on her phone. One is an ongoing caseworker carrying twenty-six open families and a court report due Thursday. One is an eligibility worker who will process forty benefit applications before lunch, each one a household waiting on food or rent. One is a case manager in housing services trying to match a client leaving a shelter to a voucher and a landlord who will take it. One is the supervisor who signs off on all of their work and answers to a director, a county board, and, when it comes to that, a judge. They use different software, sit in different units, and would describe their work in different words. But within the next year, every one of them will be handed an AI tool that drafts their documentation or scores their cases, and every one of them will be held accountable for what that tool produces under their name. The skill this program teaches, operating AI defensibly inside a consequential public-services decision, is becoming a core competency for all five of them at once. This lesson walks through each role and shows exactly where that competency lands.

Why This Is a Role Question, Not a Tech Question

It is tempting to treat AI as a tools question: pick the software, run the training, move on. That framing misses what actually changed. The tool does not sit in a vacuum. It sits inside a specific role, attached to a specific decision, governed by a specific statute, and carrying a specific consequence when it goes wrong. A large language model (LLM, the kind of AI that generates text from prompts and documents) that drafts a clean home-visit note is doing something very different in the hands of a CPS (child protective services) investigator than the same model is doing for an eligibility worker checking a benefit threshold. The verification a worker owes, the bias they have to watch for, and the due-process obligation they carry all change with the role.

So the right question is not "should our agency use AI." That decision is largely being made above the caseworker's head, by procurement offices and state systems. The right question for the working professional is: given my role, my decisions, and the people I am accountable to, what does it take to use this tool in a way I can defend to a court, an advocate, and my own conscience? That is a competency, not a setting you toggle. And it is a competency that shows up differently in each of the five roles below, even though the underlying discipline is the same: AI informs, the human decides, and every AI-touched claim gets verified to a court-record standard before it counts.

The tool is generic. The accountability is not. It attaches to a person, a role, and a decision that can change a life.

One more framing point before the roles. The competency this program builds is not about becoming a technologist. None of these five workers needs to understand the mathematics of a neural network, write code, or evaluate a model's architecture. What they need is the practiced judgment to know what an AI draft can and cannot be trusted to do, the habit of tracing every factual claim back to a source, and the discipline to hold the line between a tool that assists and a decision that stays human. That judgment is squarely inside their existing professional expertise. They already know what a defensible case record looks like. The new skill is applying that same standard to a draft that a machine wrote first.

The CPS Investigator: First Contact Under the Highest Stakes

The CPS investigator gets a report, sometimes within hours of a screening decision, and goes out to make first contact with a family. The work is fast, emotionally heavy, and bound at every step by due process: the family has rights, the allegation has to be assessed against a legal standard, and whatever the investigator documents becomes the foundation of everything that follows. Two AI surfaces touch this role hard.

The first is the screening or risk signal. Many agencies now run an AI risk-screening model that scores an incoming report for early indicators of possible abuse or neglect, intended to help decide which reports to investigate and how fast. The investigator may see that score, or a flag derived from it, before they ever knock on a door. Here the competency is precise and non-negotiable: the score is one audited input, never a verdict. History is the teacher. The debate over the Allegheny Family Screening Tool and the collapse of automated benefits-fraud systems such as Michigan's MiDAS and the Dutch childcare-benefits scandal all showed the same thing: a screening or detection model can encode the inequities in its training data and then apply them at scale to real families, disproportionately to the families who were already over-surveilled. An investigator who treats the score as a conclusion is not doing their job; they are laundering a statistical guess into a removal decision. The skill is reading the signal as a prompt to look carefully and independently, weighing it against what they actually observe, and documenting their own reasoning so that the human decision, not the model, is on the record.

The second surface is documentation. After the visit, the investigator may use an AI tool to draft the contact note or the initial safety assessment from their field notes. The hours figure here is real: investigators commonly lose a large share of every day to documentation, often half or more, and a draft that comes back in two minutes instead of forty is a genuine relief. But the contact note from a first investigation is among the most legally weighty documents in the entire system. It can support a petition to remove a child. So the verification standard is at its absolute strictest: every observation in the draft must trace to something the investigator actually saw and wrote in their raw notes, and any invented observation, a detail of bruising, a description of the home, a characterization of a parent's behavior that the model added because that kind of sentence statistically follows, must be caught and removed before the note is filed. The investigator who files an AI draft with a fabricated observation has put a false statement into a legal record under their own name, and the consequence can be a family separated on the strength of something that never happened.

The Ongoing Caseworker: The Court Report and the Caseload

The ongoing caseworker carries the case after a child is in the system: the regular home visits, the service plan, the relationship with the family, and the recurring court reports that tell the judge how things are going. This is the role where caseload and documentation collide most painfully. A caseworker carrying twenty to thirty families is generating contact notes after every visit, updating service plans, and writing court reports on a schedule, and the paperwork is the single biggest reason good caseworkers burn out and leave. When they leave, the families redistribute onto the workers who remain, caseloads climb, and the documentation pressure that drove the first worker out gets worse for everyone. The AI documentation tool lands directly on this wound, which is exactly why it is so widely adopted and exactly why the competency matters.

The court report is the document to anchor on. It is read by a judge as the professional judgment of a licensed worker, and it usually includes a history section summarizing months of case activity. Three AI failure modes converge here. An invented observation can slip into the current-status section. A misapplied policy reference can misstate a statutory standard. And fabricated history, a prior service that was never delivered, a prior incident that never happened, a prior CPS report cited without its later unsubstantiated finding, can quietly distort how the judge reads the whole family. Fabricated history is the most dangerous because it is the easiest to skim past: the history section looks familiar, so a tired worker reviews the recent observations carefully and lets the background ride. The competency is the opposite habit: open the case-management system, often a state CCWIS (Comprehensive Child Welfare Information System, the case record platform agencies use), beside the draft, and trace each historical reference to a specific record entry. A claim that cannot be traced is removed.

For the ongoing caseworker, there is a second, subtler skill: protecting the time dividend. The genuine promise of these tools is hours returned to home visits and to being present with families instead of typing about them. If a unit deploys AI drafting and then simply raises everyone's caseload to absorb the saved time, the verification step is the first thing that gets skipped under pressure, and the agency has not reduced its documentation risk, it has moved it into a harder-to-see place. The caseworker who understands this can advocate for the time to verify as a condition of using the tool at all.

The Eligibility Worker: Where a Wrong Rule Denies Food and Shelter

The eligibility worker determines whether a household qualifies for benefits: SNAP (the Supplemental Nutrition Assistance Program, federal food assistance often called food stamps), TANF (Temporary Assistance for Needy Families, cash assistance), Medicaid (public health coverage), and housing or energy assistance. The volume is high, the rules are dense and nested, and the consequence of a wrong determination is direct: a family that should get food does not, or a person who should get coverage is denied. This role sits squarely on due process, because a denial is a government action the household has the right to be notified of, to understand, and to challenge through a fair hearing.

The AI surface here is policy application. An eligibility worker may use an AI tool to help check whether a household meets a threshold, and the tool will return something that looks exactly like a researched determination: a clean conclusion, a cited regulation, the right program vocabulary. The trap is that the model generates a plausible-looking policy citation without reliably navigating the actual structure of the rule. It may apply a gross-income test to a household that is categorically eligible through a member receiving SSI (Supplemental Security Income), where that test does not apply at all. It may apply a federal standard where a more generous state option governs. It may cite a threshold that was correct eighteen months ago, before the model's training data, but has since been updated through a legislative session or an administrative action. Each of these produces a confident, wrong denial.

The competency for the eligibility worker is verification against an independent, current source, never against the AI itself. Asking the model to confirm the rule it just cited only produces a confident restatement of the same error. The check has to go to the actual current policy manual, the regulation, or the state option in force today. And because a wrong denial leaves someone without food or shelter and may not be caught until a fair hearing, this verification is not bureaucratic caution; it is the due-process safeguard built into the worker's daily practice. The decision to approve or deny stays the worker's, made on a verified reading of the policy, with the AI relegated to a research assistant whose work is always checked.

The Case Manager: Client-Facing Accuracy in Resource Navigation

Case managers work across housing, aging, disability, re-entry, and behavioral-health services, connecting clients to resources and coordinating care. Their work is less court-bound than CPS but no less consequential to the person in front of them: a wrong referral, an outdated eligibility instruction, or a hallucinated program detail can send a client in crisis to a door that does not open. Two AI uses touch this role.

The first is documentation, the same grounded-summarization discipline the other roles need: drafting case notes and service summaries from real contact, never letting the model add a service that was not delivered or a contact that did not happen. The second, more distinctive to this role, is client-facing content and resource navigation. A case manager may use AI to help draft a referral letter, explain a program to a client, or assemble a list of local resources. Here the failure mode is the AI inventing or garbling a fact that reaches a vulnerable person directly: a program that does not exist, an eligibility rule stated wrong, an address or phone number confabulated, a deadline that is incorrect. When the recipient is someone leaving a shelter or managing a disability with limited margin for a wasted trip, an invented detail is not a minor error; it is a real harm to a person who trusted the agency.

The competency, therefore, is verifying client-facing content before it reaches the client, with the same rigor a court report gets, because the client is relying on it just as a judge relies on a report. Every program named, every rule stated, every contact detail has to be confirmed against a real, current source. The case manager also carries the relationship, which means a second obligation: transparency where it matters, so that the human connection that is the heart of the work is not quietly replaced by machine-generated text the client assumes came from a person who knows their situation.

The Supervisor: Seeing the Pattern Across a Unit

The supervisor reviews and signs off on the work of a unit and answers up the chain to a director and out to courts and advocates. When AI enters the unit, the supervisor's role changes in a way the individual contributor's does not: they are no longer accountable for one worker's verification, but for whether verification is happening across everyone, all the time, under caseload pressure. Individual discipline is necessary and not sufficient. A worker who is tired, whose caseload just spiked, and whose AI draft is almost entirely correct is exactly the worker who will skim the history section and file. The supervisor's job is to make sure the system catches what the individual will eventually miss.

That turns into several concrete competencies. The first is reviewing AI-assisted documentation the way all documentation should be reviewed: as a draft to be checked against the source, not a finished product to be rubber-stamped because it reads well. The fluency of an AI draft is a hazard here; it can make a fabricated observation look more authoritative than a worker's own rougher prose. The second is watching the equity pattern. A single AI risk score is one input, but a supervisor sees the aggregate, and the aggregate is where bias becomes visible: if the screening tool is flagging certain neighborhoods or certain families disproportionately, the supervisor is positioned to notice and to escalate, turning equity auditing from a one-time procurement promise into a continuous practice. The third is protecting the conditions for verification: pushing back when leadership wants to absorb the AI time dividend as higher caseloads, and insisting that workers have the time the verification step actually requires.

Finally, the supervisor owns the audit trail. When a court or an advocate asks how AI was used in a case, the answer has to be documentable: what the tool drafted, what was verified, and who made each consequential decision. The supervisor who has built that trail can defend the unit's practice. The one who cannot is exposed, because "the AI wrote it" is not a defense in a court, a licensing review, or an agency investigation, and accountability does not transfer to the vendor. It stays with the people who signed the record.

Key Takeaways

• Five roles, the CPS (child protective services) investigator, the ongoing caseworker, the eligibility worker, the case manager, and the supervisor, all now need the same core competency: operating AI defensibly inside a consequential public-services decision, even though they use different tools and sit in different units.
• AI is a role question, not a tech question. The same generic LLM (large language model) does very different work, and carries very different stakes, depending on the role, the decision, and the statute attached to it. Workers do not need to understand the technology's internals; they need the judgment to know what a draft can be trusted to do.
• The CPS investigator faces the highest-stakes version: a risk score that must be treated as one audited input and never a verdict, and a first-contact note whose every observation must trace to what was actually seen, because it can support removing a child.
• The ongoing caseworker lives in the court-report failure modes, invented observations, misapplied policy, and especially fabricated history, and must protect the time AI returns by spending it on verification and families rather than letting it be absorbed into a higher caseload.
• The eligibility worker must verify every AI policy citation against an independent, current source, never against the AI itself, because a confidently wrong rule can deny SNAP, TANF, Medicaid, or housing to a family that qualified, with the harm sometimes uncaught until a fair hearing.
• The case manager must verify client-facing content, referrals, program details, eligibility instructions, and contact information, with court-report rigor, because a hallucinated detail sent to a vulnerable person is a direct harm.
• The supervisor's competency operates at the system level: reviewing AI drafts as drafts, watching the aggregate for equity patterns a single case cannot show, protecting the conditions for verification, and owning the audit trail a court or advocate may demand.
• Across all five roles the spine is identical: AI informs and the human decides, every AI-touched claim is verified to a court-record standard, equity comes first, and accountability stays with the person who signs the record, never the tool.