The pilot had gone well, which was exactly why it was about to fail. For eight months a single deputy director had carried the agency's AI work as a passion project on top of her real job: she chose the documentation tool, wrote the first verification checklist, ran the equity review of the screening pilot, answered the advocates' questions, and trained the two units that volunteered. Then she was promoted and moved to another county. Within six weeks the verification checklist had stopped being used in one unit, the screening equity review was three months overdue with no one assigned to it, and a caseworker had quietly started pasting case histories into a consumer chatbot because nobody was left to tell her not to. The agency had not lost an AI tool. It had lost the only person who made the AI tool safe, and it had never written that person's work down as a job. A responsible AI program that lives in one enthusiast's spare time is not a program. It is a single point of failure with a child's safety and a family's benefits riding on it. The work has to become someone's defined role, on the org chart, with a title, a mandate, and a successor, or it does not survive the first promotion.

Why Roles, Not Volunteers

The single most common way a human-services AI program collapses is the one the opening describes: it was never anyone's actual job. It lived in the margins of a deputy director's week, or it was split informally among a few interested supervisors, or it was assumed to be "IT's problem," which meant it was nobody's. The work got done as long as the enthusiast stayed and had spare capacity, and it stopped the moment either ran out. This is not a personnel problem. It is a design problem. The responsible use of AI in a consequential field requires specific, continuous work (verification standards, equity auditing, incident response, training, vendor evaluation, policy maintenance) and continuous work that is not assigned to a defined role does not reliably happen.

Defining roles does something a policy alone cannot. An enterprise AI policy says what must be done. Roles say who does it, which means roles are what make the policy true rather than aspirational. A policy that requires a scheduled equity audit is a wish until a named person owns the audit, has the time and authority to run it, and answers for it if it slips. A policy that requires verification to a court-record standard is enforced only if a defined role audits that verification actually happens across units. The org chart is where governance stops being a document and becomes a set of people who can be held accountable, who can be backfilled when they leave, and whose work survives a promotion.

There is a defensibility dimension as well. When an advocate or an oversight reviewer asks who is responsible for ensuring the agency's AI does not encode inequity, "we all sort of watch for it" is not an answer a court accepts. "Our equity auditor, a defined role with this mandate, ran this audit on this schedule and paused this tool when it found disparity" is. Naming the roles is part of what makes the program defensible to the court, the advocate, and the community the agency serves.

Consider the arithmetic of the opening failure, because it shows why this is structural rather than a matter of finding more dedicated people. The deputy director was covering, on top of a full-time job, perhaps eight to ten distinct streams of continuous work: tool selection, the verification standard, the equity review schedule, advocate correspondence, two units of training, incident triage, vendor questions, and policy drift. Each stream is small until it is no one's job. The moment she left, every one of those streams had a deadline that nobody owned. The verification standard did not fail because workers stopped believing in it; it failed because the person who walked the units reminding people to use it was gone and no successor had been named. An agency that wants the work to continue cannot solve this by hoping the next enthusiast appears. It has to write the streams into roles before the enthusiast leaves, so that when the inevitable promotion comes, the agency replaces a position rather than discovering a vacuum.

A policy says what must be done. A role says who does it. Without the second, the first is only a wish, and a wish cannot survive the day its author gets promoted.

The Agency AI Lead

The agency AI lead is the role that owns the program. This is not a technologist who manages servers and not a vendor liaison who manages contracts, though the role works closely with both. The agency AI lead is the person accountable for the whole responsible-AI operating model: the enterprise policy and its maintenance, the roadmap of where AI is and is not used, the coordination of equity auditing and verification standards across programs, the incident-response path when a tool causes harm, and the agency's answer when a court or an advocate asks how the work is governed. If the program were a system, the AI lead is the named owner of the system.

The mandate matters more than the title. An AI lead without authority is a coordinator who can be overruled by any unit director who finds the verification standard inconvenient, which means the standard is optional, which means it is not a standard. The role needs the authority to pause a tool, to require a control before deployment, and to escalate when a unit drifts from the policy. It needs a seat where decisions are made, not a seat where decisions are reported after the fact. In a small agency the AI lead may be a senior practice leader who carries the role alongside other duties; in a large agency it is a dedicated position. Either way, the work is defined and assigned, so that when the person in the role is promoted or leaves, the agency replaces the role rather than discovering, six weeks later, that the program left with the person.

Concretely, the AI lead is the one who, when the advocate's letter from the policy lesson arrives, owns the response: they hold the policy, they know which tools are deployed under which control sets, they can produce the audit trail, and they coordinate the incident path if the letter reveals a real problem. The role is the difference between a program with a defined owner and a program that depends on whoever happens to care that week.

Picture the AI lead's authority tested in a real moment. A unit director carrying a backlog of 30 families per worker wants to deploy a new generative drafting tool next week to relieve the documentation pressure, and proposes skipping the equity and verification mapping to move faster. An AI lead with only a coordinating title can object, be thanked, and be ignored, and the tool goes live ungoverned. An AI lead with real authority can require that the tool be mapped to the policy's control sets before deployment, pause it if it is rushed in anyway, and escalate to the governance board if the unit director persists. The difference is not personality; it is mandate. The lesson the field keeps relearning is that a standard a unit director can override at will under caseload pressure is not a standard, it is a suggestion, and suggestions do not protect children or families when the pressure is highest. The AI lead's authority exists precisely for the moments when doing the safe thing is slower than doing the fast thing.

The Equity Auditor

The equity auditor is the role that makes the field's first non-negotiable real. Equity is not an afterthought and not a one-time procurement check; it is a continuous practice, and a continuous practice requires a continuous owner. The equity auditor's defined job is to test the agency's AI tools for bias on a schedule, before harm and not after, and to have the authority to act on what the testing finds.

The need for this role is written in the field's history. The long public debate over the Allegheny Family Screening Tool turned on questions of whether a risk model flagged families by proxy for poverty and race. Michigan's MiDAS system wrongly accused tens of thousands of people of unemployment fraud. The Dutch childcare-benefits scandal saw an algorithmic fraud-detection system devastate tens of thousands of families, disproportionately families with immigrant backgrounds, and ultimately forced a national government to resign. The lesson the field took from these is not that screening tools must never be used; it is that a tool which can encode and amplify inequity must be audited continuously by someone whose job is to look for exactly that. An equity auditor is the agency's defense against repeating that history inside its own walls.

The role's work is concrete: define the disparity metrics the agency will test against, run the audit on the schedule the policy mandates, examine whether a screening tool produces disparate outcomes across the groups the agency serves, document the findings in a form a court and an advocate can read, and trigger the policy's pause-on-disparity rule when a tool fails. The authority to pause is what separates an equity auditor from an equity report-writer. A role that can find disparity but cannot stop the tool that produces it is a role that documents harm rather than preventing it. The equity auditor must be able to take a tool out of service, and the agency must back that authority, or the role is decoration.

Independence matters too. An equity auditor who reports to the same person whose pilot they are auditing faces a conflict that undermines the audit. The role should have a reporting line that lets it deliver an uncomfortable finding (that a flagship tool a director championed is producing disparate outcomes) without that finding being buried. Equity auditing that cannot survive an inconvenient result is not equity auditing.

The Practice-AI Specialist

The third role is the one that keeps the program connected to the actual work. The practice-AI specialist is a caseworker or supervisor (someone who has carried a caseload, sat in a home, and written a court report) whose defined job is to be the bridge between the AI program and the people who use AI on real cases. The AI lead owns the system and the equity auditor guards against bias, but neither of them is in the home visit. The practice-AI specialist is the role that makes sure the tools and the policy work for the worker in the field, not just on paper.

This role exists because tools designed without practitioners fail in predictable ways. A verification checklist written by someone who has never been under caseload pressure at 9 p.m. will be too long to use and will be skipped. A prompt template that does not match how caseworkers actually take field notes will produce drafts that are harder to verify than they are worth. A documentation tool that produces output the case-management system rejects wastes the time it was supposed to save. The practice-AI specialist catches these failures because they live where the work happens. They train caseworkers in plain practice language rather than technical jargon, they collect the friction reports from the field and feed them back to the AI lead, and they are the trusted face who can tell a skeptical veteran caseworker that this tool is worth learning because they have used it on their own cases.

The practice-AI specialist also carries the culture of the cardinal rule into the units. It is one thing for a policy to state that AI informs and humans decide; it is another for a respected practitioner to model that discipline daily, to show a new worker how to verify a draft against field notes, and to push back when efficiency pressure starts to crowd out judgment. This role is where the decision-aid culture is taught by example. An agency that has an AI lead and an equity auditor but no practice-AI specialist tends to build a technically governed program that the workforce quietly resists or works around, because nobody made it work for the people actually doing the cases.

A concrete scene shows the role's value. A new caseworker, three months into the job and already behind on documentation for a caseload of 28 families, is handed an AI documentation tool and a six-page verification checklist written by a governance analyst. At 9 p.m., facing a stack of overdue notes, she reads the checklist once, decides she does not have the forty minutes per note it seems to demand, and starts filing AI drafts after a quick skim. That is the precise moment a verification standard dies in the field, and no policy document will revive it. A practice-AI specialist who has carried a caseload sees this coming. They would have tested the checklist on their own cases first, found that the forty-minute version is unusable, and rebuilt it into a focused claim-by-claim check that fits the workflow: trace each observation to the field notes, check each policy citation against the manual, confirm each history reference in the case-management system, sign. The specialist then sits with the new worker, shows her the verification taking eight minutes rather than forty, and turns a standard she would have abandoned into one she can actually keep. The tool and the policy were identical in both versions of the story. The presence of a practitioner who made them work for the field is the entire difference between a returned hour and a fabricated observation in a court record.

How the Roles Fit the Org Chart

These three roles are not three new hires every agency must make on day one. They are three distinct accountabilities that must exist somewhere, named and assigned, and how they map to actual positions depends on the size of the agency.

In a small county agency, the three accountabilities may live in two or even one person, but they must still be written as distinct duties so that none of them quietly disappears. A senior practice leader might hold the AI lead and practice-AI specialist roles together, with the equity-auditor accountability assigned to someone with enough independence to deliver an uncomfortable finding (which is precisely why even in a small shop the equity-audit role should not collapse into the AI lead who champions the tools). The principle is that the work is named even where the headcount is small. A single person wearing two hats is governable; an unwritten habit is not.

In a large state agency or a major nonprofit, these become dedicated positions, often with teams: an AI lead with a small governance staff, one or more equity auditors with the independence and the metrics to test tools across many programs, and practice-AI specialists embedded in each major program area so the bridge to the field exists everywhere AI is used. The roles connect to the existing org chart through the governance board: legal, equity, practice, and community voices at the table, with the AI lead accountable to that board, the equity auditor reporting in a way that protects independence, and the practice-AI specialists feeding the field's reality up into the same structure.

The test of whether the roles are real is the same test as for the policy, applied to people. If the deputy director from the opening were promoted today, would the verification standard still be enforced next month, would the equity audit still run on schedule, and would a caseworker who started pasting case histories into a chatbot be corrected by someone whose job that is? If the answer is yes, the agency has built roles. If the answer is no, it still has a passion project, and the next promotion will end it. The org chart is where a responsible AI program becomes durable: someone's defined job, with a successor, rather than a side project that protects children and families only as long as one enthusiast stays in the building.

Key Takeaways

• The most common way a human-services AI program collapses is that it was never anyone's defined job. A program that lives in one enthusiast's spare time is a single point of failure that ends with the next promotion. The work must become defined roles on the org chart.
• An enterprise AI policy says what must be done; roles say who does it. Roles are what make a policy true rather than aspirational, and they are what let an agency answer an oversight reviewer with a named, accountable owner rather than "we all sort of watch for it."
• The agency AI lead owns the whole operating model: the policy and its maintenance, the roadmap, cross-program coordination of verification and equity standards, incident response, and the agency's governed answer to courts and advocates. The role needs real authority (to pause a tool, require a control, escalate drift), not just a coordinating title.
• The equity auditor makes the equity-first non-negotiable real by testing tools for bias on a schedule, before harm. The role must have the authority to pause a tool that shows disparity and enough independence to deliver an uncomfortable finding about a tool a director championed.
• The field's history (the Allegheny Family Screening Tool debate, Michigan's MiDAS wrongful fraud accusations, the Dutch childcare-benefits scandal that forced a government to resign) is why continuous, independent equity auditing must be a defined role, not an occasional check.
• The practice-AI specialist is a current or former caseworker who bridges the AI program and the field, ensuring tools and checklists actually work under caseload pressure, training workers in plain practice language, feeding friction reports back, and modeling the decision-aid culture by example.
• The three roles are accountabilities that must exist somewhere, named and assigned. In a small agency they may be carried by one or two people (while keeping equity auditing independent); in a large agency they become dedicated positions and teams connected through the governance board.
• The test of whether the roles are real: if the program's founder were promoted today, would the verification standard still be enforced, the equity audit still run, and a worker pasting case histories into a chatbot still be corrected by someone whose job that is? If yes, the agency has roles; if no, it still has a side project.