The deputy director had eleven minutes on the county board's agenda. She had a slide deck the vendor had given her, full of words like transformation and efficiency, and she could feel in advance how it would land. The board had read the same news stories everyone had read: the benefits algorithm that wrongly accused thousands of fraud, the screening tool accused of reproducing bias, the chatbot that gave a person in crisis the wrong number. If she walked in and said the agency was bringing artificial intelligence into child-welfare casework, the first question would not be about efficiency. It would be sharp, and it would be from the board member whose district had the highest removal rate, and it would be some version of: are you going to let a computer decide whether to take a child. She deleted the vendor deck. She had eleven minutes to make a case that an agency could defend to a court, to an advocate, and to a parent who had every reason to distrust the government, and overpromising even once would cost her all of it. This lesson is about how she made that case, and how you make yours.

Why the Case Is Hard, and Must Be Made Anyway

Making the case for AI in human services is harder than making it almost anywhere else, and the difficulty is not a problem to route around. It is a signal that you are working in the right register. In a commercial setting, the case for AI is a productivity story and the audience wants to hear it. In human services, the audience, leadership and the public, has every reason to be skeptical, because the failures in this field have been severe and the people harmed were the least able to absorb the harm. A board member, a journalist, an advocate, or a parent who hears AI and child welfare in the same sentence is not being unreasonable when they brace. They are remembering the history.

And yet the case must be made, because the alternative is also a harm. The documentation burden is the field's defining pain. Caseworkers spend a large share of every day, often half or more, documenting instead of being with families, and that burden is a top driver of burnout and turnover, which raises caseloads for those who remain, which lets more harm slip through. AI transcription and summarization tools that draft a clean case note from a home visit, the pattern now among the most widely adopted in the field, can return real hours to direct work. To refuse that benefit entirely is to leave caseworkers drowning in paperwork and families getting less of their worker's time. So the leader's job is not to choose between the benefit and the risk. It is to make a case honest enough that both are visible, and disciplined enough that the audience believes the agency can hold the line.

The cost of getting this wrong is concrete and political. An agency that oversells, that promises an efficiency revolution and then has an incident, loses not just the program but its credibility, and the next responsible proposal dies in the backwash. An agency that undersells, that hides the AI use or frames it as a minor tooling change, gets caught being less than transparent, and in this field a transparency failure is a due-process failure. The narrow path between overselling and hiding is the only one that holds.

In human services, the audience's skepticism is earned by history. A case for AI that does not honor that skepticism will not be believed, and should not be.

Lead With the Mission, Not the Technology

The deputy director's first decision was to delete the word transformation. The vendor deck led with the technology because the vendor sells the technology. The agency does not serve the technology; it serves children and families, and the case has to be built from that mission outward. The opening of a defensible case is not what the AI does. It is the problem the agency is trying to solve for the people it serves.

Framed that way, the case starts on ground the audience already shares. Every board member, every advocate, every parent wants caseworkers to spend more time with families and less time typing. Every one of them wants lower turnover, because they know that a family who gets a new worker every few months gets worse outcomes. Every one of them wants documentation that is accurate, because they know the record drives the court. The mission frame turns the AI from an imposition into a means: we are trying to give caseworkers their time back so they can be present with families, reduce the burnout that churns our workforce, and produce records that are more accurate, and here is one carefully bounded tool that can help us do that, with these protections.

This is not a rhetorical trick. It is a discipline that keeps the program honest, because a case built on the mission forces the agency to measure the mission. If the claim is time back for families, the agency must measure hours returned and where they went. If the claim is lower burnout, the agency must track it. A case built on transformation can hide behind a vague word. A case built on the mission has to show the mission moving, which is exactly the accountability the audience wants and the program needs.

The Worked Numbers, Honestly

Concreteness earns trust; vague grandeur destroys it. Suppose a unit of 15 caseworkers each spends roughly half of a working day on documentation, and a verified AI-assisted note-drafting workflow returns, after the time spent verifying every draft, a conservative net of 30 to 45 minutes per worker per day. Stated honestly, that is the better part of an extra home visit per worker per week, returned to direct contact. The leader presents it that way, as hours measured and redirected to families, not as a percentage efficiency gain that sounds like staff cuts. And the leader states the cost in the same breath: those returned hours assume workers actually have time to verify every draft, and the program is designed so the saved time funds verification and direct work, not a heavier caseload. A number presented with its own caveat is a number the audience can trust. A number presented as a guarantee is a liability waiting for the first quarter that misses it.

Name the Risks Before They Do

The single most powerful move in the deputy director's eleven minutes was to raise the hardest objection herself, before the board member could. She put up one line: this tool will never decide whether to remove a child, substantiate a report, or approve or deny a benefit, and here is how we guarantee that. By naming the fear out loud and answering it, she did three things at once. She showed she understood the stakes, which earned the room's attention. She took the most damaging question off the table by answering it on her own terms. And she signaled that the agency had thought about the risk before deploying, not after a harm.

The risks to name are the program's own non-negotiables, stated as commitments. The cardinal rule: AI informs, humans decide, and a named caseworker, supervisor, and court own every consequential call. Documentation discipline: the model drafts but never invents an observation, and every AI-touched claim is verified to a court-record standard before it enters the record. Equity: every risk signal is one audited input under mandatory human review, and the agency runs equity audits on a schedule because history (the Dutch childcare-benefits scandal, Michigan's MiDAS, the Allegheny Family Screening Tool debate) proves these tools can encode the inequities in their training data. Due process and privacy: notice, the right to a fair hearing and to challenge a determination, and the protection of the most sensitive data, all preserved. Transparency: the agency discloses its AI use to clients, courts, and advocates, because disclosure is what keeps the work defensible.

Naming these is not a concession that weakens the case. It is the case. The audience does not need to be told the technology is impressive; they have read that. They need to be shown that the agency is the kind of agency that will not let the technology cross the line, and the only way to show that is to name the line and the mechanism that holds it. A leader who lists the risks and the controls is far more credible than one who insists everything will be fine.

Raise the hardest objection yourself, answer it on your own terms, and you have shown the room you understand the stakes better than the objection assumed.

The Discipline of Not Overpromising

Every vendor figure is a benchmark to verify, never a guarantee, and a leader who repeats a vendor's accuracy claim as a promise has borrowed a liability. The defensible posture is to present vendor performance numbers as claims the agency will test in its own pilot, with its own data, against its own equity gate, before any scaling. This protects the leader twice: it keeps the public case truthful, and it builds in the pilot-before-scale discipline that good governance requires anyway. When a board member asks whether the tool is as accurate as the vendor says, the honest answer, we will measure that ourselves before we trust it, is more reassuring than any number, because it shows the agency does not take the vendor's word, which is precisely what the public fears the agency might do.

Tailor the Case to Each Audience

The same honest case has to be told differently to different audiences, because each one carries a different fear and a different authority. The substance never changes; the emphasis does.

To leadership (the board, the director, the budget authority): the emphasis is on the workforce and fiscal story bounded by risk control. Burnout and turnover are expensive and harmful, the documentation burden drives them, the AI program returns measured hours and the program includes the governance, equity auditing, and audit-readiness posture that keeps the agency out of a headline. Leadership needs to hear that the upside is real and the downside is managed, with named owners and a governance board, because their accountability is institutional.

To the public, advocates, and the people served: the emphasis is on protection and rights. The message leads with what the AI will never do and what stays human, and it leads with disclosure: the agency will tell you when AI was used in your case, you retain every right to notice, a fair hearing, and to challenge any determination, and the agency audits for equity so the tool does not fall harder on some families than others. This audience does not need the efficiency story; they need to know their rights are intact and the agency is being honest with them.

To frontline caseworkers and their unions: the emphasis is on the time-back-for-the-mission story and the assurance that the tool serves them, not replaces them, and that the returned hours go to families and verification, not to a heavier caseload. Caseworkers are both the beneficiaries and the verifiers, and a case that treats them as the people the program is built to help, rather than the people it is done to, earns the buy-in without which no deployment survives contact with the floor.

The Same Question, Three Fears

It helps to see how a single likely question lands differently in each room, because the question that follows the pitch is rarely about the technology in the abstract. Suppose the question is: what happens when the tool gets something wrong. To leadership, the fear behind that question is institutional and fiscal: a wrong output becomes a lawsuit, a headline, a finding, a budget hole. The honest answer for that room is the governance answer: every AI-touched claim is verified before it enters the record, the program logs who verified and who decided, and the agency runs an audit-readiness posture so that if a reviewer ever asks, the trail already exists. To the public and the people served, the same question carries a different fear: a wrong output becomes my child removed on a false detail, or my benefits denied on a rule that did not apply. The honest answer for that room leads with the protection: the tool never makes the decision, a person checks every fact before it is filed, you will be told AI was used in your case, and you keep every right to challenge what the agency does. To the caseworker, the fear is professional and personal: a wrong output becomes my name on a false record, my license in question. The honest answer for that room is the verification answer: the workflow is built so you can check every draft, the saved time funds that check rather than a bigger caseload, and the agency stands behind the discipline rather than blaming the worker when the tool errs. One question, one set of facts, three fears, three emphases. A leader who hears only the words of the question and misses the fear behind it will give the right facts to the wrong fear and lose the room.

This is why scripting the case as a single deck read identically to every audience is a mistake even when the deck is accurate. The board does not need the rights lecture they will read as boilerplate, and the parent does not need the fiscal model they will read as the agency caring more about money than children. Each room needs to hear the agency speaking to the thing they are actually afraid of, in the agency's own honest words, with the same underlying commitments holding it all together. The discipline is not to change the truth per audience. It is to lead with the part of the truth that answers the fear in front of you.

When the Case Meets a Real Incident

The truest test of a case built honestly is what happens when something goes wrong, because in a program at scale something eventually will: a hallucinated observation caught in review, an equity audit that surfaces a disparity, a determination challenged at a fair hearing. The agency that oversold has no room to absorb this; the gap between the promise and the reality becomes the story. The agency that built its case honestly, that said from the start the tool can err and that is exactly why we verify everything and audit continuously, has already told the public that errors are expected and controlled. When an error is caught, it confirms the agency's account rather than refuting it.

This is why the honest case and the audit-readiness posture are the same discipline pointed in two directions. The leader who told the board the agency runs quarterly equity audits and corrects what they find can, when an audit finds a disparity, point to the correction as proof the system works as promised. The catch is the proof, not the failure. A program presented as flawless has nowhere to put a flaw. A program presented as carefully governed has a place for every problem, which is the governance that catches it. The deputy director closed her eleven minutes on exactly this point: we are not promising you a tool that never errs, because no such tool exists. We are promising you an agency that catches the errors before they reach a family, proves it caught them, and tells you the truth either way. That is a promise an agency can keep, and the only kind worth making.

Key Takeaways

• The case for AI in human services is hard because the audience's skepticism is earned by real history (the Dutch childcare-benefits scandal, Michigan's MiDAS, the Allegheny Family Screening Tool debate); a case that does not honor that skepticism will not be believed and should not be.
• The case must still be made, because refusing the documentation benefit leaves caseworkers buried in paperwork (often half or more of the day) and families getting less of their worker's time, which drives the burnout and turnover that harm the work.
• Lead with the mission, not the technology: frame AI as a bounded means to give caseworkers time back with families, reduce burnout, and produce more accurate records, which also forces the agency to measure the mission rather than hide behind a vague word like transformation.
• Present numbers honestly and concretely (for example, a conservative net 30 to 45 minutes per worker per day after verification time, redirected to families) with their caveats stated in the same breath; a number with its own caveat earns trust, a guaranteed number is a future liability.
• Name the hardest objection yourself before the audience does: the cardinal rule that AI never decides a removal, a substantiation, or a benefit, plus the commitments on verification, equity auditing, due process, privacy, and disclosure, stated as the controls that hold the line.
• Never repeat a vendor performance figure as a guarantee; present every vendor claim as a benchmark the agency will test in its own pilot against its own equity gate before scaling, which is both honest and good governance.
• Tailor emphasis to each audience: workforce-and-fiscal-bounded-by-control for leadership, protection-and-rights-and-disclosure for the public and people served, and time-back-without-a-heavier-caseload for frontline caseworkers and unions, with the substance unchanged.
• A case built honestly survives a real incident: an agency that promised careful governance rather than perfection can point to a caught error or a corrected equity disparity as proof the system works, because the catch is the proof, not the failure.