The intake call had lasted fifty-one minutes. A new family, a frightened parent, three children, a housing crisis tangled up with a domestic-violence history and a missed medical appointment that a school had flagged. The intake worker had typed as fast as she could, capturing fragments: dates half-heard, a name spelled two ways, the parent's own words about why she left the apartment, a long stretch where the parent cried and the worker just listened and wrote almost nothing. At the end she had four screens of rough, out-of-order notes and a blank intake assessment summary that her supervisor needed before the case could be assigned. She pasted her notes into the agency's AI tool and asked it to produce a clean intake summary. What came back was orderly, complete, and readable in a way her raw notes were not, and that is precisely the danger, because an intake summary is the document every later decision-maker reads first, and the model had quietly turned the family's uncertainty into the worker's confident assertions and filled the silent stretches with assumptions no one had ever stated.

Why the Intake Summary Carries Outsized Weight

Of all the documents a caseworker produces, the intake and assessment summary deserves special care for a structural reason: it is the document that frames the case for everyone who touches it afterward. The supervisor assigning the case reads it. The worker who picks up the case reads it as their orientation to the family. A court, in a child-welfare matter, may read it as the agency's first account of why the family came to attention. In benefits and eligibility work, the assessment summary sets the frame for what programs a person is screened for and what their situation is taken to be. An error in a single late case note harms the record at one point. An error in the intake summary propagates, because it shapes the lens through which every subsequent contact is interpreted.

This is also the document where AI assistance is genuinely valuable, which is why the lesson is not "do not use it." Intake notes are, by their nature, the messiest notes a worker produces. They are taken live, under emotional pressure, often by phone, capturing a person in crisis who is not telling their story in chronological order. Turning four screens of fragmentary live notes into a clean, organized, complete summary is real cognitive labor that can eat thirty to forty-five minutes at the end of an already long intake. An AI tool that drafts that summary from the worker's notes can return a meaningful share of that time, and a well-organized summary genuinely serves the family by making sure the next reader understands the situation. The value is real. The risk is that the very act of cleaning up messy input is where the model introduces its most characteristic intake errors.

The intake summary is the lens every later reader looks through. An assumption the model adds here is not one error; it is the frame for every decision that follows.

The Gap-Filling Failure Mode

Every AI documentation task carries the general hallucination risk: a large language model (LLM, an AI system that generates text by predicting the most likely next word) produces fluent, confident prose whether or not it is grounded in the source. But intake summarization has a specific, intensified version of this risk, and naming it precisely is the heart of this lesson. Call it gap-filling. Intake notes are full of gaps: the parent did not finish a sentence, the worker did not catch a date, a topic came up and was dropped, the parent cried and nothing was recorded. A human reading those notes sees the gaps as gaps. The model does not. The model's entire function is to produce coherent, complete text, and a summary with visible holes is not coherent or complete. So the model fills the holes with the statistically likely content, and it does so seamlessly.

Gap-filling shows up in intake summaries in several recurring shapes, and a worker who knows the shapes can hunt them deliberately.

Uncertainty Becoming Assertion

The most common and most damaging intake gap-fill is the conversion of the family's uncertainty, or the worker's, into a confident assertion. The parent says "I think it was maybe a couple of weeks ago, I'm not really sure." The worker's notes read "left apt ~2 wks ago?? unsure." The model's clean summary reads "The family left the apartment approximately two weeks prior to intake." The hedge is gone. The question mark is gone. The family's own uncertainty has been laundered into the worker's professional statement of fact, and the next reader has no way to know the date was a guess. In a case where timing matters, when a child was last seen, how long a person has been unhoused, when an incident occurred, this kind of false precision can mislead a safety assessment or an eligibility determination built on it.

The verification move is to preserve uncertainty as uncertainty. An intake summary should be allowed to say "the parent reported leaving the apartment approximately two weeks before intake but was uncertain of the exact timing." Hedged language in an intake summary is not sloppy writing; it is accurate documentation of what is actually known. The worker's job in verification is to restore every hedge the model removed.

Reported Becoming Observed, and Said Becoming Confirmed

A second shape is the collapse of the crucial distinction between what was reported to the worker and what the worker established or observed. In an intake, almost everything is reported: the parent says the other parent was violent, the parent says the children have been missing school, a school says a medical appointment was missed. These are reports, and an accurate intake summary attributes them clearly: "the parent reported," "according to the school," "the parent stated." The model, smoothing toward clean prose, frequently drops the attribution and converts a report into a flat statement of fact. "The parent reported that the father was physically abusive" becomes "The father was physically abusive." A serious, contested, unproven allegation has just been written into the agency's first record as an established fact, which is both inaccurate and a due-process problem, because the family's right to contest an allegation depends on it being recorded as an allegation.

The verification move is to restore attribution to every reported claim and to make sure the summary distinguishes the three different epistemic levels in an intake: what the worker directly observed (rare in a phone intake), what was reported by the family, and what was reported by a third party such as a school or a hospital. Each carries different weight, and collapsing them into undifferentiated "facts" misrepresents the case from its first document.

Inventing the Unspoken

The third shape is pure invention in the service of completeness. Intake assessment formats often have standard sections: presenting concerns, household composition, history, protective factors, immediate needs. If the worker's notes are silent on one of those sections, because the topic did not come up or the parent did not get to it, the model knows the section should exist and may populate it with plausible content. A "protective factors" section that the parent never discussed gets filled with generic protective factors. A "substance use" line that was never raised gets a "no concerns reported" that was never actually asked or answered, which is itself a false statement: "no concerns reported" implies the question was asked. A household-composition section gets a relationship inferred from context that the parent never stated.

The verification move here is to treat every section against the notes and to let the summary say "not assessed" or "not discussed at intake" where the topic genuinely was not covered. A truthful gap is far more useful to the next worker than a fabricated completeness, because the next worker needs to know what still has to be asked. An intake summary that looks complete but quietly invented its completeness sends the next worker into the home believing questions were answered that were never even raised.

Grounding the Draft Before You Verify It

Verification is the safeguard, but the work starts earlier, at how the draft is generated. The same prompting discipline that produces accurate case notes applies with extra force to intake summaries, because the input is messier and the gap-filling temptation is stronger. Two practices reduce the number of fabrications you will have to catch later.

First, instruct the model to mark gaps rather than fill them. A prompt that says, in effect, "summarize only what is in these notes; where a standard intake section is not addressed in the notes, write 'not discussed at intake' rather than supplying typical content; preserve all expressions of uncertainty; attribute every claim to its source as reported, observed, or third-party" changes the model's behavior meaningfully. It does not eliminate gap-filling, because the model's underlying drive toward completeness remains, but it tilts the draft toward honesty about what is missing and gives you a cleaner starting point. This is grounding through instruction: constraining the model to the record and to the epistemics of an intake.

Second, give the model structure that mirrors how the family actually presented, not a generic template it will feel obligated to complete. If the agency's intake format has ten sections and the call covered four, a model handed the ten-section template will work to fill ten. A worker who organizes the prompt around what was actually discussed, and explicitly flags the rest as not covered, removes the structural pressure that produces invented sections. The model fabricates most where the template demands content the conversation did not provide.

Even with both practices, the draft is a draft, and the verification standard is unchanged: every claim traced to the notes before the summary is filed. Grounding reduces the volume of fabrications; it never removes the obligation to check.

The Intake Verification Pass

Verifying an intake summary uses the same adversarial, claim-by-claim discipline as any AI-assisted document, with three intake-specific checks layered on. The worker reads the draft beside the raw intake notes, assuming the model has smoothed something it should not have, and works through each claim.

• The hedge check. For every statement of fact in the summary, ask: was this certain in my notes, or has a hedge been removed? Restore every "approximately," "reportedly," "unclear," and "the parent was uncertain" that the model stripped out. A date, a duration, or a count that the family was unsure about must read as uncertain in the summary.
• The attribution check. For every claim, ask: who is the source, and does the summary say so? Restore "the parent reported," "the school stated," "according to the caller" to every claim that is a report rather than a worker observation. Make sure no allegation is written as established fact.
• The section check. For every section of the assessment format, ask: did we actually discuss this, or did the model populate it? Replace any invented content with "not discussed at intake" or "not assessed" so the next worker knows what remains to be asked.
• The standard claim check. As with any case document, trace every name, date, age, address, and figure to the notes, and verify any policy or program-eligibility statement against the current source rather than the model's assertion.

A Worked Intake Verification

Return to the fifty-one-minute call. The worker has her clean AI draft and her four screens of raw notes, and she runs the intake pass, which takes her about twelve minutes against the thirty-five-plus the summary would have taken to organize from scratch. The draft opens: "The family experienced domestic violence and fled the apartment two weeks ago." She stops on the first sentence. Her notes say the parent reported a history of violence by the other parent, and that the parent was unsure of exact timing. Two corrections: this is reported, not established, and the timing is uncertain. She rewrites it as "The parent reported a history of domestic violence by the children's father and stated she left the apartment approximately two weeks before intake, though she was uncertain of the exact date." Further down: a "Substance Use" section reads "No substance use concerns reported." Her notes have nothing on substance use; it never came up. She replaces it with "Substance use: not discussed at intake." A "Protective Factors" section lists a supportive extended family. The parent never mentioned extended family. She deletes the invented content and writes "Protective factors: not yet assessed; to be explored at first home visit." Each correction takes seconds. Together they are the difference between a summary that honestly frames a family in crisis and one that recasts an uncertain, half-told story as a confident set of findings the next worker will act on.

The Decision-Aid Boundary at Intake

Intake is also where the cardinal rule of this whole program, AI informs and humans decide, has to be stated with particular clarity, because intake summaries increasingly sit next to AI screening tools, and the two must not be confused. A clean AI-drafted intake summary is a documentation aid: it organizes what the worker gathered. It is not, and must never be allowed to become, a screening decision about whether the case warrants investigation, what its risk level is, or whether a family should be referred for a particular response. Those are human decisions, made by the worker and the supervisor under the agency's structured decision-making process, and they are bound by due process and equity.

The reason to be explicit is that a well-organized AI summary can create a subtle pressure toward agreeing with its framing. When the summary reads cleanly and confidently, a busy supervisor may anchor on its account and its implicit emphasis. This is exactly why the verification work matters beyond catching individual fabrications: the summary the supervisor reads must be the worker's accurate account, with its uncertainties and attributions intact, not the model's smoothed and subtly editorialized version. A summary that has quietly converted allegations into facts and uncertainty into precision will push the human decision in a direction the actual intake did not support. Keeping the summary honest is part of keeping the decision human.

And the accountability is unchanged. The intake summary goes into the record under the worker's name as the agency's first account of the family. If it contains a fabricated finding, an unattributed allegation written as fact, or an invented section, the worker owns it, the same as any document. "The AI organized my notes" is no more a defense than "the AI wrote it." The tool drafts; the worker verifies; the human decides; and the record reflects what actually happened at intake, uncertainties and all.

A clean intake summary organizes what was gathered. It never decides what the case is. The screening call belongs to the worker and the supervisor, never to the document.

Key Takeaways

• The intake and assessment summary carries outsized weight because it frames the case for every later reader: the assigning supervisor, the next worker, a court, an eligibility screen. An error here propagates into every subsequent decision, unlike an error in a single later note.
• AI assistance is genuinely valuable at intake because intake notes are the messiest a worker produces (live, emotional, out of order), and a clean summary can return thirty to forty-five minutes. The value is real; the risk is that cleaning up messy input is exactly where the model fabricates.
• The intake-specific failure mode is gap-filling: the model fills the silences, hedges, and missing sections in intake notes with statistically plausible content, because its function is to produce complete, coherent text and it does not see gaps as gaps.
• Gap-filling takes three recurring shapes: uncertainty becoming assertion (a hedged guess written as fact), reported becoming observed (an allegation written as established fact, dropping attribution), and inventing the unspoken (populating a standard section the conversation never covered, including a false "no concerns reported").
• Ground the draft before verifying it: instruct the model to mark gaps as "not discussed at intake" rather than fill them, to preserve all uncertainty, and to attribute every claim as reported, observed, or third-party. Organize the prompt around what was actually discussed, not a generic template the model will feel obligated to complete.
• The intake verification pass adds three checks to the standard claim-by-claim review: the hedge check (restore every removed uncertainty), the attribution check (restore every "the parent reported" and write no allegation as fact), and the section check (replace invented sections with "not discussed at intake" so the next worker knows what remains to ask).
• A truthful gap is more useful than a fabricated completeness. A summary that looks complete but invented its completeness sends the next worker into the home believing questions were answered that were never raised.
• The decision-aid boundary applies at intake: a clean AI summary organizes what was gathered; it never decides whether the case warrants investigation or what its risk level is. Those are human decisions bound by due process and equity, and a smoothed, editorialized summary can improperly push them. The worker owns the summary under their name; "the AI organized my notes" is not a defense.