The risk-screening tool flagged the case orange. On the supervisor's screen it was a small colored band next to the family's name, the output of a predictive model that had read the prior referral history and returned a number the agency translated into a tier. The investigator who had actually been to the home that morning sat across the desk and said the family was struggling but safe: the apartment had food, the toddler was bonded to the mother, the older child was in school and doing well, and the one prior referral two years ago had been screened out as unfounded. The supervisor agreed. The case stayed open for services, not removal. It was, by every measure that mattered, the right call, made by two people who had weighed a machine's signal and decided it did not change what they saw. And then they wrote it down. Or rather, they did not, not really. The case note said the family was offered voluntary services and the case would remain open. It said nothing about the orange flag, nothing about how the screening score had been considered, nothing about why two professionals had looked at an elevated risk tier and concluded it did not warrant the action the tier implied. Eight months later, when a different incident reopened the case and an attorney requested the full record, the orange flag was discoverable in the system logs but the human reasoning that had answered it was nowhere. The decision had been right. The documentation could not prove it.

Why the Signal Itself Must Be Documented

This lesson is about a quiet failure that does not look like a failure at the time it happens. When a caseworker or supervisor uses an AI risk signal correctly, treating it as one input under mandatory human review and not as a verdict, the natural instinct is to document the decision and move on. The decision is what matters, after all. The family is getting services or not, the child is staying home or not, the case is opening or closing. The signal feels like scaffolding: useful for getting to the decision, but not part of the building once the decision is made. So it goes undocumented. And that is exactly the mistake.

In a field governed by due process, the right to challenge a determination is meaningless if the basis of the determination cannot be reconstructed. Due process here means notice, a fair hearing, and the right to challenge a decision the government makes about a family. When an AI signal touched a case, that signal is part of the basis, whether or not it changed the outcome. An advocate has the right to know it was there. A court has the right to understand how it was weighed. An equity auditor reviewing whether a screening tool produces disparate outcomes across racial or economic lines needs to see not just the scores the tool generated but how workers responded to them. A signal that was used but not documented is a signal that operated invisibly, and invisible influence on a consequential government decision is precisely what due process exists to prevent.

There is a second reason, more protective of the worker. The decision-aid rule, the program's cardinal rule that AI informs and humans decide, is only defensible if you can show that a human actually decided. An undocumented signal leaves an ambiguous record: a model produced an elevated risk tier, a worker took the action the tier pointed toward, and there is nothing in the file showing the worker exercised independent judgment rather than deferring to the machine. Even when the worker did exercise judgment, the silent record cannot prove it. The documentation of how the signal was used is what converts the abstract principle "humans decide" into a concrete, auditable fact about this specific case. Without it, the worker who did everything right is indistinguishable, on paper, from the worker who rubber-stamped the algorithm.

A right decision that cannot be reconstructed is not a defensible decision. In a due-process field, the reasoning is part of the record, not a private step on the way to it.

The Four Things a Signal Note Must Capture

Documenting how a signal was used is not a vague aspiration. It is a concrete set of facts that belong in the record every time an AI risk signal touches a case decision. There are four of them, and a note that captures all four turns an invisible influence into a transparent, challengeable, auditable part of the case. Think of it the way you would think of citing a source in any professional document: you name what informed you, what it said, what you did with it, and why.

One: That a Signal Existed at All

The first and most basic fact is that an AI risk-screening tool produced an output for this case. This sounds obvious, but it is the fact most often omitted, because to the worker the signal is just part of the interface they work in every day. The note must state plainly that a predictive risk-screening tool was run, what tool it was, and when. A note that says "a structured decision-support screening was completed on intake at 9:14 AM using the county's risk-screening tool" does in one sentence what the orange-flag case in the opening failed to do across an entire case file. It establishes, for anyone who reads the record later, that AI was in the loop. An advocate cannot challenge a signal they do not know existed; an equity audit cannot account for a score that never appears in the narrative record.

Two: What the Signal Actually Said

The second fact is the content of the signal: the score, tier, flag, or category the tool produced, recorded exactly as it was, without the worker's interpretation folded into it yet. If the tool returned a high-risk tier, the note records a high-risk tier. If it returned a numeric score within a band, the note records the score and the band. The discipline here is to capture the raw output faithfully, because the raw output is what an auditor will compare across cases when checking for disparate impact, and it is what an advocate will examine when asking whether the tool treated this family the way it treated others. Softening or paraphrasing the signal ("the screening suggested some concern") corrupts the record. The signal said what it said. Write it down.

Three: How the Human Weighed It Against the Evidence

The third fact is the heart of the note and the part that proves a human decided. It is the worker's reasoning: what the signal indicated, what the worker actually observed and gathered, and how those two things were weighed against each other. This is where the orange-flag supervisor's missing sentence belonged. It would have read something like: "The screening tool returned an elevated risk tier driven primarily by a two-year-old prior referral. On investigation, that referral was an unfounded report that had been screened out. The current home assessment found adequate food, a securely bonded toddler, and an older child attending school and performing well. The worker and supervisor determined that the elevated tier reflected historical data that the present assessment did not corroborate, and that the family's current circumstances did not meet the threshold for removal."

Notice what that paragraph does. It does not hide the signal; it engages it. It names what drove the score, tests that driver against present-day observation, and shows the judgment as a weighing rather than a deference or a dismissal. A worker spends maybe four extra minutes writing this. Those four minutes are the difference between a record that proves human review happened and a record that merely asserts a decision was made. Across a caseload of 24 families with several AI-touched decisions a week, that discipline costs perhaps an hour a week. Set against a fair-hearing reversal, an equity-audit finding, or a deposition where a worker cannot explain why they acted on a flag, that hour is the cheapest insurance in the workflow.

Four: The Decision and That It Was the Human's

The fourth fact is the decision itself, stated as the worker's and the supervisor's, not the tool's. The language matters. "The case was screened in for investigation" is weaker than "Based on the assessment above, the worker and supervising investigator decided to screen the case in for investigation." The active human subject in the sentence is not a stylistic preference; it is the documented assertion that a person, not a model, made the consequential call. The decision-aid rule lives or dies in this sentence. When the record names a human as the decider and shows the reasoning that led there, the rule is proven on the face of the document. When the record is passive or attributes the action to the screening outcome, the rule is undermined no matter how careful the worker actually was.

Documenting When You Disagree With the Signal

The hardest and most important documentation happens when the human decision goes against what the signal pointed toward, in either direction. These are the cases where the record is most likely to be scrutinized and where careful documentation does the most work, both for the family and for the worker.

Consider the override downward: the signal indicates elevated risk and the worker, on the evidence, declines to take the action the elevation implies. This is the orange-flag case. Here the documentation must be especially thorough, because the worker is choosing the less cautious path relative to the algorithm, and if something goes wrong later, the question will be why the elevated signal was not acted on. The note that captures what drove the score, why the present evidence did not support it, and who made the decision is the worker's complete answer to that question. A worker who can point to a contemporaneous note saying "the elevated tier was driven by a prior unfounded referral that the current assessment did not corroborate" is in an entirely different position than a worker who acted correctly but left the file silent. The first worker documented a defensible judgment. The second worker, identical decision, looks like someone who ignored a warning.

Now consider the override upward: the signal is low or moderate, but the worker sees something the model did not and decides on a more protective action. This matters just as much, because it is the clearest possible demonstration that the human is not deferring to the tool. The note should record that the screening returned a lower tier, what the worker observed that the tool's inputs did not capture (a fresh disclosure, a present-day safety concern, a parent's account that contradicted the historical data), and that the worker's judgment, not the score, governed the action. Documenting the upward override is how an agency proves to an equity auditor and a court that its workers treat the tool as a floor for attention, never a ceiling, and that low scores never suppress human concern.

Document the agreement, document the override, and document the override most of all. The case where the human and the machine diverge is the case where the record has to carry the most weight.

There is a subtle trap to name here. When the human decision happens to match the signal, there is a temptation to let the signal stand in for the reasoning, as if agreement means the work is done. It does not. A note that says "high-risk screening confirmed; case screened in" has documented agreement but not judgment. It reads as the worker ratifying the model, which is the opposite of what the decision-aid rule requires. Even when you agree with the signal, the note must show that you reached the decision by weighing the evidence, and that the agreement is a result of your judgment rather than a substitute for it. The reasoning is what you document; the signal is only one of the things the reasoning weighed.

The Language Traps That Undermine the Record

How a signal note is worded can quietly defeat its purpose. Certain phrasings, which feel natural and efficient to a time-pressured worker, transfer agency from the human to the machine on the face of the record. Learning to spot and avoid them is a core skill.

The first trap is the passive verdict. "The case was flagged as high risk and screened in." There is no human in that sentence. It reads as though the flag did the screening. Rewrite it so a person is the subject of the consequential verb: "After reviewing the high-risk flag against the home assessment, the worker screened the case in." The second trap is treating the score as a finding rather than an input. "The risk assessment determined the family to be high risk" presents the model's output as a determination, which it is not. The model produced a signal; the determination is the human's. Write "the screening tool returned a high-risk score" to keep the score as an input and reserve "determined" for what people decide. The third trap is the unexamined confirmation, already discussed: "screening confirmed concern" hides the reasoning behind a word that implies the model validated the worker rather than the worker validating, or not, the model.

A fourth trap is silence about the score's drivers. When the tool exposes what factors drove a score, the note should engage them, especially when those factors are demographic, geographic, or historical proxies that equity audits exist to catch. A score driven mainly by a family's zip code or by a years-old screened-out referral deserves explicit notice in the record, because that is precisely the kind of driver that encodes inequity, and a worker who flags it in the note is doing the equity work in real time and leaving evidence that they did. Saying nothing about the drivers leaves the auditor to wonder whether the worker even saw them.

The fifth trap is documenting the signal somewhere the human narrative does not reach. Many systems log the score automatically in a structured field or an audit log. That automatic log is necessary but not sufficient. A score sitting in a system field, disconnected from the worker's narrative reasoning, tells a reader what the tool said but not what the human did with it. The orange-flag case had exactly this: the flag was discoverable in the system logs, but the reasoning was nowhere in the narrative record. The signal note must live in the case narrative, where the reasoning lives, so that the score and the human response to it can be read together as one connected account.

Building the Signal Note Into the Workflow

Individual discipline is necessary and not sufficient, because the pressure that produces silent records is structural. A worker carrying 24 to 30 families, processing several AI-touched decisions a week, under a documentation burden that already consumes half the day, will not reliably remember to write a four-fact signal note from scratch every time unless the workflow makes it the path of least resistance. The fix is to standardize the note so it becomes habit rather than improvisation.

The simplest mechanism is a short structured template that travels with every AI-screened case, prompting the four facts in order: signal existence and source, the raw signal, the human weighing, and the human decision. A worker who fills four prompted fields produces a complete, defensible signal note in the same four minutes that improvisation would have taken, but without the risk of omitting the fact they happen to forget under pressure. The template is not bureaucracy for its own sake; it is the operational form of the decision-aid rule, the thing that makes "humans decide" provable case by case without depending on memory.

Supervisory review closes the loop. A supervisor reviewing an AI-touched decision should be reading the signal note specifically, checking that all four facts are present, that the reasoning genuinely weighs the evidence rather than ratifying the score, and that the language names a human as the decider. A unit that audits its own signal notes on a sample basis, the way it audits any other documentation, catches the drift toward passive, deferential records before an outside auditor or an attorney does. And the agency that can show its own internal review of how signals are documented is in a far stronger position with an equity audit or an oversight body than one that discovers its documentation gaps only when challenged.

Finally, the time has to be there. The same caseload pressure that makes AI screening attractive as a triage aid is the pressure that produces silent records. If the only way to keep up is to act on the flag and move to the next case, the signal note will be the step that gets skipped, and the agency will have automated its decisions without documenting its judgment. Building the signal note into the workflow means budgeting the few minutes it costs, treating those minutes as non-negotiable in the same way the human review itself is non-negotiable, and recognizing that a screening program without signal documentation is not a governed program. It is an undocumented one wearing the appearance of governance.

Key Takeaways

• When an AI risk signal touches a case decision, the signal and the human response to it are part of the basis of that decision and must be documented, even when the signal did not change the outcome. A right decision that cannot be reconstructed is not a defensible one in a field governed by due process: notice, a fair hearing, and the right to challenge.
• A complete signal note captures four facts: that a screening tool ran and which one, what the signal actually said (the raw score or tier, not a paraphrase), how the human weighed it against the gathered evidence, and the decision stated as the human's and the supervisor's.
• The reasoning is the heart of the note and the proof that the decision-aid rule held. It costs roughly four minutes per AI-touched case, about an hour a week across a typical caseload, and it is the cheapest insurance against a fair-hearing reversal, an equity-audit finding, or an unanswerable deposition.
• Documenting an override matters most. A downward override (declining to act on an elevated signal) needs thorough reasoning because it is the path most likely to be questioned later. An upward override (acting more protectively than a low signal suggested) is the clearest proof that the worker does not defer to the tool.
• Agreement with the signal still requires documented reasoning. "Screening confirmed; case screened in" ratifies the model instead of showing human judgment, which is the opposite of what the cardinal rule (AI informs, humans decide) requires.
• Avoid the language traps: the passive verdict ("the case was flagged and screened in"), the score-as-finding ("the assessment determined the family high risk"), the unexamined confirmation, silence about a score's drivers (especially demographic or historical proxies), and burying the score in a system log disconnected from the narrative reasoning.
• When a score's drivers are demographic, geographic, or stale historical proxies, name them in the note. That is equity work done in real time, and it leaves evidence for the equity audit that these tools require because history shows they can encode inequity.
• Make the signal note structural, not heroic: a short four-field template that travels with every AI-screened case, supervisory review that reads the signal note specifically, internal sampling audits, and the protected minutes to write it. A screening program without signal documentation is an undocumented program wearing the appearance of governance.