The pilot was going well, which was the problem. A county child-welfare unit had spent three months using an AI tool to draft case notes, court reports, and intake summaries, and the hours it returned were real: workers reported getting back something close to a full day a week. Then a quality reviewer noticed something in a stack of substantiation decisions. The AI-drafted assessments included a line, generated by the tool, that read "Risk level: elevated" or "Risk level: moderate" near the top of each document. Nobody had told the tool to make a risk determination. It had learned the pattern from the assessment templates it was trained on and produced the label as part of the draft. And the workers, busy and trusting a tool that was right about almost everything else, had begun leaving the label in and, without quite deciding to, treating it as a starting point for their own judgment. The handoff between what the AI did and what the human decided had never been designed. It had just happened, in the white space of a fast workflow, and it had quietly migrated a consequential decision from a caseworker to a text-prediction system. No one chose that. That is exactly why it happened.

Why the Handoff Must Be Designed, Not Assumed

The cardinal rule of this field is that AI informs and humans decide. Every practitioner who reaches this lesson can recite it. The trouble is that the rule is a principle, and a principle does not enforce itself. In a real workflow, with a real caseload, the boundary between informing and deciding is not a wall; it is a moment, and if that moment is not built into the process as an explicit, visible, unavoidable step, it dissolves. The decision does not get made by a human who carefully weighed the AI's input. It gets made by default, by inertia, by the worker accepting what the draft already implied because there was no point in the workflow that forced a stop.

This is the central insight of handoff design: the line between AI and human is not protected by good intentions or by training that reminds workers of the cardinal rule. It is protected by structure. A well-designed handoff makes the human decision a discrete, observable event that has to happen before the work moves forward. A poorly designed handoff, or an undesigned one, lets the AI's output flow straight into the record and into the decision, with the human reduced to a rubber stamp who clicks "accept" on a draft they did not truly evaluate. The opening story is what an undesigned handoff produces: not a dramatic failure, but a quiet drift in which the model's suggestion becomes the human's starting position, and the human's independent judgment is never actually exercised.

Consider the math of inertia. A worker carrying 28 families, drafting perhaps six to ten documents a week with an AI tool that is genuinely accurate 90-plus percent of the time, is being trained by experience to trust the draft. Every accepted draft that turns out fine reinforces the habit of accepting. By the fortieth correct draft, the worker is not evaluating; they are approving. The handoff has to be designed against that gradient, because the gradient runs toward automation by default, and the consequential decisions in this field, removal, substantiation, benefit denial, are precisely the ones that must never be made by default.

The boundary between AI informing and a human deciding is not protected by intentions. It is protected by structure, or it is not protected at all.

Naming the Human-Only Decisions First

Designing the handoff begins not with the AI but with the decisions. Before a single step is automated, the practitioner names the decisions in the case process that are human-only, the ones an algorithm must never make, draw the inference toward, or pre-fill. In child welfare and benefits, these are the consequential, due-process-bound determinations: whether to remove a child, whether to substantiate a report, whether a safety threat exists, whether to deny or terminate a benefit, what the risk level is, what the recommendation to the court should be. These are not steps where AI provides a starting point that a human adjusts. They are steps where the human reasons from the evidence and the AI's role is bounded well short of the conclusion.

Naming these decisions explicitly does real work, because it converts the vague rule "humans decide" into a concrete list of moments the workflow must protect. The list also reveals the danger in the opening story. "Risk level: elevated" is a human-only determination. The moment the tool generated that label, it crossed from drafting into deciding, even though it was only predicting text. The fix is not to ask the tool to be more careful. The fix is to design the workflow so the tool never produces that label at all, because a label the human merely confirms is a decision the human did not make.

The practitioner should be able to point at any step in the mapped case process and say one of two things: this step the AI may draft, summarize, or organize, and a human verifies it; or this step is a human decision the AI does not touch, though it may have helped assemble the evidence the human reasons from. There is no third category in which the AI "recommends" a consequential decision and the human "approves" it, because that category is precisely where the handoff fails. A recommendation on a removal or a substantiation, presented to a tired worker as a default, is a decision wearing the costume of a suggestion.

The Anatomy of a Clean Handoff

A clean handoff has parts that can be specified and built. The first is a stop: a point in the workflow where the AI's contribution ends and the human's begins, and where the work cannot proceed until a human acts. The stop is the opposite of the seamless flow that produced the opening story. It is friction placed deliberately, at the one place friction protects people.

The second part is a changed question. At the stop, the worker is not asked "is this draft acceptable?" because that question invites a yes. The worker is asked to perform the decision from the evidence: given the verified facts, what is the safety determination, and on what basis? The handoff is designed so that the human produces the consequential conclusion rather than ratifying one the AI produced. In the substantiation example, a clean handoff hands the worker the organized, verified evidence and a blank where the determination goes, not a pre-filled risk label the worker edits.

The third part is separation of materials. What the AI produced (the draft note, the organized summary, the assembled evidence) is kept visibly distinct from what the human decides. The worker sees the AI's contribution as input, labeled as input, and records their own determination as a separate act. This separation is what lets a later reviewer, a supervisor, a court, an advocate, see that the human decision was actually made and on what it rested, rather than finding an AI draft with a human signature appended.

Make the contrast concrete with two versions of the same screen. In the broken version, the worker opens the assessment and sees, near the top, a field already populated: "Risk level: elevated," followed by an editable narrative box. The worker reads the rest, finds nothing alarming, and leaves the label. Total time at the decision: 40 seconds. The record now shows an elevated-risk substantiation that the worker technically signed but never independently reached. In the clean version, the worker opens the assessment and sees the organized, verified evidence on the left and, where the determination belongs, an empty field with a prompt: "State your safety determination and the specific evidence supporting it." There is no label to accept. The worker has to read the evidence and write a basis, which takes perhaps four to six minutes. The record now shows the worker's reasoning, the evidence it rested on, and a conclusion that is genuinely theirs. The two screens save the same drafting hours. Only one of them produces a decision a court can attribute to a human, and the difference is entirely in how the handoff moment was built.

The fourth part is a record of the decision point itself. The handoff produces an artifact: the human decided X, at this moment, on this verified evidence, and here is the AI's role and where it stopped. This is the audit trail that makes the decision-aid boundary defensible. In the opening unit, once the handoff was redesigned, every substantiation carried a short record showing the worker reached the determination from the evidence, with the AI's drafting role logged and bounded. That record is what would let the agency answer a court that asked whether a human or a model decided.

A clean handoff hands the worker organized, verified evidence and a blank where the determination goes, never a pre-filled conclusion to edit.

Designing Against Automation Bias

The deepest enemy of a good handoff is automation bias: the well-documented human tendency to over-trust the output of an automated system, to accept its suggestion with less scrutiny than one would give a human colleague, and to treat its presence as evidence of correctness. Automation bias is not a failing of careless workers. It is a predictable feature of how people interact with reliable-seeming tools, and it grows stronger the more accurate the tool actually is, because accuracy builds the trust that then gets misapplied to the rare wrong case.

Handoff design fights automation bias structurally, not by exhortation. Telling workers "remember, you are the decision-maker" does almost nothing against a 90-plus percent accurate tool and a 28-family caseload. What works is removing the default the worker would otherwise accept. If the AI never produces the risk label, there is no label to defer to. If the determination field is blank and requires the worker to articulate a basis, the worker has to reason rather than ratify. If the workflow asks the worker to identify the evidence supporting their determination before showing them any AI summary, the worker forms an independent judgment first. The design principle is to engineer the moment so that the path of least resistance is the human actually deciding, rather than the human accepting.

There is a measurable consequence to getting this wrong. In the opening pilot, before the redesign, a quality review sampled substantiation decisions and found that in a meaningful share of cases the worker's narrative simply restated the AI's pre-filled risk label as if it were their own conclusion, with no independent reasoning visible in the record. After the redesign removed the label and required an evidence-based determination, the same review found workers articulating a basis, sometimes agreeing with what the old label would have said and sometimes not. The hours saved did not change. What changed was that the decisions were once again being made by people, and the record could prove it.

Handoffs Across Documentation, Eligibility, and Screening

The handoff looks different depending on what the AI is helping with, and a practitioner designing real workflows needs to see the pattern in each of the field's main use cases.

Documentation

In note and report drafting, the goldmine use case, the AI does the most and the handoff is about verification rather than a single decision. The AI drafts; the human verifies every factual claim against the source and owns the filed document. The stop here is the verification gate: the document cannot be filed until the worker has checked it claim by claim. The handoff artifact is the verification record. The decision the human owns is the accuracy and truthfulness of a legal record bearing their name, and "the AI wrote it" transfers none of that accountability. Even here, where AI contributes the most, there is no consequential case decision embedded in the draft; if the AI's draft of a court report contains a recommendation to the court, that recommendation is a human-only determination that the draft must not pre-decide.

Eligibility and Benefits

In an eligibility determination, the AI may help gather and organize information and even surface which rules appear relevant, but the determination of whether a person qualifies is the human's, verified against the current policy source. The handoff stop is before the determination: the AI assembles and organizes, the human applies the verified policy and decides. The danger mirrors the opening story. A tool that pre-fills "Determination: ineligible" has made the decision, and a worker who accepts it has denied someone food or medical coverage by default. The clean handoff gives the worker the organized facts and the verified applicable rules and requires the worker to make and justify the determination, because a wrong denial that no human truly decided is a due-process failure as well as a harm.

Screening Support

Screening support is the use case where the handoff matters most and is hardest, because the AI's output is itself a risk signal that looks like a decision. The non-negotiable design is that the signal is one audited input under mandatory human review, never a verdict and never a default. The handoff must present the signal as one piece of evidence among many, must require the human to reach the screening decision from the full picture, and must record that the human decided and why, including where they weighed the signal differently than its face value. A screening handoff that lets a risk score flow into a decision unexamined is the exact pattern the field's hardest history, the predictive-tool controversies, warns against, and equity auditing of the signal sits alongside the handoff as a permanent companion control.

Building, Testing, and Sustaining the Handoff

A designed handoff is not finished when it is drawn; it has to be built into the tool and the process, then tested against the way work actually happens. Building it means configuring the AI tool so it does not generate human-only conclusions, structuring the case-management system so the decision field is the worker's to complete, and writing the policy that says which steps are human-only and what each handoff must record. It means training workers not on the slogan but on the moments: here is where you stop, here is what you decide, here is what you log.

Testing the handoff means watching for the drift the opening story describes. A handoff can be well-designed on paper and erode in practice, because workers under pressure find the path of least resistance and tools get updated with new features that quietly reintroduce a default. The practical test is a periodic review of a sample of consequential decisions, asking a single question: can I see, in the record, that a human reasoned to this conclusion from verified evidence, distinct from what the AI produced? If the answer is yes, the handoff is holding. If the records show AI conclusions with human signatures attached, the handoff has failed regardless of what the design document says.

Put numbers on the testing cadence so it is a practice and not a hope. A unit of 15 workers each making perhaps two consequential determinations a week generates around 30 such decisions weekly, or roughly 120 a month. A reviewer does not need to re-examine all of them; a sample of 15 to 20 consequential decisions a month, pulled across workers, is enough to detect drift. For each sampled decision the reviewer asks the single question and records a yes or no: is there visible human reasoning from verified evidence, distinct from the AI's output? If the yes rate is high and steady, the handoff is holding. If it slips, say from 95 percent to 70 percent over two months, that slide is the early signal that a tool update added a default or that caseload pressure is pushing workers back toward acceptance. Catching that slide at 70 percent, before it reaches the floor, is the entire point of testing, because by the time a fabricated or migrated decision surfaces in a courtroom the harm is already done.

Sustaining the handoff is a governance responsibility, not just a worker one. The agency owns the obligation to keep the boundary intact as tools change, caseloads shift, and new staff arrive. That means re-checking the handoff every time the AI tool is updated, because a vendor update can add an auto-populated recommendation that silently turns a clean handoff into the opening story again. It means protecting the time workers need to actually perform the decision rather than rush past it. And it means treating any erosion of the handoff as an incident worth investigating, because a migrated decision in this field is not a process inefficiency. It is a child's safety or a family's benefits decided by a system that was never supposed to decide it.

Key Takeaways

• The cardinal rule that AI informs and humans decide is a principle, and a principle does not enforce itself; the boundary between informing and deciding is protected by workflow structure or it is not protected at all.
• An undesigned handoff produces quiet drift, not dramatic failure: a model's output (such as an auto-generated "Risk level: elevated" label) becomes the human's default starting position, and a consequential decision migrates from a caseworker to a text-prediction system that no one chose to let decide.
• Handoff design begins by naming the human-only decisions (removal, substantiation, safety determination, risk level, benefit denial, recommendation to the court) and protecting them; there is no valid third category in which the AI "recommends" a consequential decision and the human "approves" it.
• A clean handoff has four parts: a deliberate stop where work cannot proceed without a human act, a changed question that asks the worker to decide from the evidence rather than accept a draft, separation of AI-produced material from the human decision, and a record of the decision point that serves as the audit trail.
• Automation bias, the tendency to over-trust an automated system, grows stronger as the tool gets more accurate and is fought structurally, not by exhortation: remove the default, blank the determination field, and have the worker form an independent judgment before seeing the AI's conclusion.
• The handoff differs by use case: in documentation it is a verification gate, in eligibility it is a stop before the determination so the human applies verified policy, and in screening support it requires the risk signal to be one audited input under mandatory human review, never a verdict.
• A handoff must be tested by sampling consequential decisions and asking whether the record shows a human reasoned to the conclusion from verified evidence distinct from the AI's output; AI conclusions with human signatures attached mean the handoff has failed.
• Sustaining the handoff is a governance duty: re-check it whenever the AI tool is updated, because a vendor update can silently reintroduce an auto-populated recommendation, protect the time workers need to decide, and treat any erosion as an incident worth investigating.