The letter was supposed to be kind. A case manager at a county aging-services unit needed to tell Mrs. Alvarez, eighty-one, that the home-delivered meals program she relied on was changing its delivery schedule, and that she would need to be home on a different day or arrange for a neighbor to receive the meals. It was a small message with a large weight: Mrs. Alvarez lived alone, did not drive, and had told the case manager more than once that the meal delivery was sometimes the only human contact in her day. The case manager had eleven other letters to send before five o'clock, so she opened the agency's AI writing assistant and asked it to draft a warm, clear notice explaining the schedule change. The draft came back polished and gentle. It opened with a line about how much the program valued Mrs. Alvarez. It explained the change in plain language. And in the second paragraph it included a sentence the case manager had not asked for and that was not true: "If you are unable to be home, you may visit our downtown office between 9 and 5 to pick up your meals in person." There was no pickup option. The downtown office was four miles away with no bus route Mrs. Alvarez could use. The model had invented a helpful-sounding alternative because letters like this one usually offer an alternative, and an eighty-one-year-old woman who could not drive was one paragraph away from being told to do something impossible.

Why Client Communication Is Its Own Kind of Risk

Most lessons about AI in human services focus on the case record: the note, the court report, the eligibility determination. Client communication is different, and the difference is what makes it dangerous in a way that is easy to underestimate. A case note is read by colleagues, supervisors, and courts, people trained to be skeptical and equipped to verify. A letter, a text message, a portal message, or an email to a client is read by the person whose life the message affects, often a person in crisis, often a person with limited time, limited options, and no way to tell that a sentence was generated by a text-prediction system rather than written by the worker who knows their case.

That asymmetry is the whole problem. When AI drafts a case note and invents a detail, the error has several layers of professional review between the keyboard and any consequence. When AI drafts a client message and invents a detail, the error can go straight from the model to a vulnerable person who will act on it. Mrs. Alvarez might have spent a morning trying to arrange transport to an office that would turn her away. A parent told in an AI-drafted message that a court hearing is "rescheduled to next month" when it is not may miss the actual hearing and lose ground in a custody case. A benefits applicant told their interview is "optional" when it is mandatory may forfeit the benefit. The client cannot verify the message. They trust it because it came from the agency, on the agency's letterhead, in the agency's voice. The trust is exactly what makes the error land.

A case note's reader is trained to doubt. A client is trained to trust. That is why an invented sentence in a client message is more dangerous than the same sentence in a case note.

There is a second feature that makes client communication distinct: tone carries meaning. A case note is judged on accuracy. A client message is judged on accuracy and on whether it lands as intended with a human being who may be frightened, grieving, defensive, or exhausted. AI is genuinely good at producing warm, clear, well-structured prose, which is precisely why it is appealing for this task and precisely why it can lull a worker into trusting the content because the tone is right. A message can be perfectly warm and perfectly wrong. The warmth is not evidence of accuracy. It can be camouflage for an invented fact.

Where AI Genuinely Helps With Client Communication

None of this means AI has no place in client communication. It has a real and humane place, and naming it precisely is what keeps the use safe. The pain is concrete: a case manager carrying thirty to forty clients, or a benefits worker processing dozens of cases, spends a substantial part of every week writing the same kinds of messages, appointment reminders, status updates, requests for documents, explanations of a decision, referrals to a service. Each one needs to be clear, accurate, and humane, and each one takes time the worker would rather spend with the people in front of them. AI can return some of that time, if it is pointed at the parts of the task where the worker, not the model, holds the facts.

AI helps most with the form and the tone, not the facts. Consider a worker who knows exactly what she needs to say to a client: the appointment moved to Thursday at 2 p.m., the client needs to bring two recent pay stubs and a utility bill, and the office is on the third floor. Those are her facts. She can give the model those facts and ask it to draft a warm, plain-language message at a sixth-grade reading level that conveys them clearly. The model's job is to phrase, structure, and soften, not to supply any of the substance. Used this way, the worker remains the source of every fact and the model is a writing aid, like a very capable assistant who drafts from her dictation and whose draft she reads before it goes out.

AI also helps with reading-level and plain-language conversion. A worker who has written an accurate but bureaucratic paragraph can ask the model to rewrite it so a person in crisis, or a person reading in their second language, can understand it. This is a real equity contribution: dense, jargon-laden client communication excludes the people least able to decode it, and plain-language rewriting widens access. The worker still verifies that the rewrite did not change any fact, because a plain-language rewrite can quietly alter a deadline or a requirement while it simplifies the sentence. But pointed at clarity rather than substance, AI can make agency communication more humane and more accessible at once.

Translation: A Special and Higher-Risk Case

Language access deserves its own warning because the temptation is enormous and the risk is acute. AI translation is fast, cheap, and available in dozens of languages, and the population it would serve, clients with limited English proficiency, is exactly the population least able to detect a translation error and most harmed by one. A mistranslated deadline, a mistranslated legal right, or a mistranslated instruction can cause real harm to a person who has no way to check it against the original. AI translation can be a genuine aid, but only when the output is reviewed by a qualified human translator or bilingual staff member before it reaches the client, and only when the agency understands that legal-rights notices and consequential determinations often carry formal language-access obligations that an unreviewed machine translation does not satisfy. The general rule holds with extra force here: the model can draft; a qualified human must verify before it reaches a person who cannot verify it themselves.

The Failure Modes in a Client Message

The failures in client communication rhyme with the failures elsewhere in AI-assisted work, but they land differently because the reader is the affected person. Naming them makes them catchable.

The invented option or instruction. This is the Mrs. Alvarez failure. The model adds a helpful-sounding alternative, resource, or instruction that does not exist, because messages of this type usually contain one. The harm is a client acting on a false instruction: traveling to an office that will turn them away, attempting a process that is not available, or relying on a deadline extension that was never granted.

The wrong fact, fluently stated. The model states a date, time, address, amount, or requirement incorrectly. If the worker gave the model the right facts and the model transcribed them wrong, or if the worker relied on the model to supply a fact it did not actually have, the message carries a confident error. A wrong hearing date or a wrong document requirement can cost a client a benefit or a court position.

The overpromise. The model, drafting warmly, produces language that commits the agency to something it cannot guarantee: "Your benefits will be restored within five business days," "We will approve your application," "You will be contacted by Friday." Warm drafting tends toward reassurance, and reassurance can become a promise the agency cannot keep and the client will rely on. An overpromise also creates a due-process and accountability problem when the agency cannot deliver what its own letter committed to.

The wrong tone for the situation. A model does not know that this particular client just lost a child to foster placement, or that this family is in acute crisis. It may produce a chipper, upbeat message where the situation demands gravity and care, or boilerplate empathy that reads as hollow to someone in genuine distress. Tone failures do not cause the same concrete harm as a wrong date, but they damage the relationship and the trust the work depends on, and they can make a hard moment harder.

The privacy leak. A worker rushing to draft messages may paste a client's sensitive information, a child-abuse allegation, a medical condition, an immigration detail, into an AI tool that is not approved for personally identifiable information (PII, the data that identifies a specific person). Or a model drafting one client's message may carry over a detail from another. Client communication touches the most sensitive data in the field, and the convenience of a quick AI draft must never override the obligation to protect it.

Verify Before It Reaches a Person in Crisis

The control is the same control that governs all AI-assisted work in this field, adapted to the stakes of a message that goes straight to a vulnerable reader: nothing reaches the client until a person has verified it. But verification of a client message has its own specific checklist, because the things that can go wrong are specific.

Verify Every Fact Against the Source, Not the Draft

Every concrete claim in the message, every date, time, address, amount, deadline, requirement, and instruction, must be checked against the worker's own knowledge of the case and the authoritative source, not against the fluent draft. The appointment date should match the calendar. The document requirements should match the actual policy. The deadline should match the real deadline. If the message says the client can do something, the worker confirms that option actually exists. The Mrs. Alvarez letter would have been caught in five seconds by a worker asking: is there a pickup option? There was not. The sentence should not have been there.

Hunt for the Helpful Invention

The most dangerous addition is the one that sounds helpful, because it does not trigger suspicion. Workers should read every offered option, alternative, and resource in an AI draft with the specific question: did I provide this, or did the model add it? An invented phone number, an invented office hour, an invented alternative process, an invented extension, all of these read as kindnesses and all of them can send a client toward a wall. The fluent helpfulness of the addition is exactly what makes it slip past review.

Check the Tone Against the Human Behind the Case

The worker knows the client; the model does not. Before a message goes out, the worker should read it as the client will read it, with the client's actual situation in mind. Is the warmth appropriate or does it ring false given what this family is going through? Does the message acknowledge the difficulty of the situation, or does it breeze past it? Does it commit the agency to anything it cannot deliver? This is the judgment the model cannot supply because it does not carry the relationship. The worker does, and verifying tone is part of verifying the message.

Protect the Data Before You Draft, Not After

The privacy check happens before the worker types anything into the tool, not after the draft comes back. The question is whether the tool is approved for this kind of client information, and whether the worker needs to put identifying or sensitive details into the prompt at all. Often a worker can draft a message using only the structural facts (the appointment moved, these documents are needed) without pasting the client's name, condition, or allegation into the tool. Minimizing what sensitive data touches an AI tool is a discipline that protects the client whether or not the tool is approved.

The Worker Owns Every Word That Goes Out

The message goes out under the agency's name and the worker's responsibility. As with a case note or a determination, "the AI drafted it" is not a defense if the message misinformed a client, overpromised, or breached privacy. The worker who sends the message owns every word in it. That ownership is not a burden added to the AI workflow; it is the thing that makes the AI workflow safe. The model produces a draft fast; the worker makes it true and makes it land, and then it is the worker's message, not the model's.

The Time It Returns, and Where It Should Go

It is worth being honest about the math, because the case for AI client communication rests on it. A worker who writes twelve client messages in an afternoon, each taking ten or fifteen minutes to compose carefully, can use AI to produce the drafts in a fraction of that time and then spend the bulk of the saved time on verification and on the messages that need a human's full attention. The drafting of routine, structural messages is genuine toil, and removing it is a genuine gift.

But the saved time has to go somewhere honest. If an agency sees that AI cuts drafting time and responds by raising the number of clients each worker carries, the verification step will be the first thing squeezed, and the agency will have built a faster pipeline for sending clients wrong information. The Mrs. Alvarez letter is not caught by a better model; it is caught by a worker who has the time and the discipline to ask whether the pickup option exists. Protecting that time is the difference between AI that makes client communication more humane and AI that industrializes a new kind of error.

The deepest version of the case is this: the point of returning the time is to return the worker to the client. The hours saved on drafting routine notices should buy more of the thing the routine notices were always a poor substitute for, the actual human contact, the call to check in, the visit, the moment of being present with a person in a hard season of their life. Mrs. Alvarez did not most need a better-worded letter. She needed someone to notice that the meal delivery was sometimes her only human contact, and to act on that. AI that drafts the letter so the worker can make the call is AI used well. AI that drafts the letter so the worker can draft eleven more is AI used to deepen the very problem it was supposed to solve.

Key Takeaways

• Client communication is a distinct, higher-stakes AI risk because the reader is the affected person, often in crisis, who trusts the message and cannot verify it. An invented sentence in a client message can go straight from the model to a vulnerable person who acts on it, with no layer of professional review in between.
• AI helps most with form and tone, not facts. The worker supplies every concrete fact (date, time, address, requirement, instruction) and the model phrases, structures, and softens. A warm, well-written message can be perfectly worded and perfectly wrong; warmth is not evidence of accuracy.
• Plain-language and reading-level conversion is a genuine equity contribution, widening access for people in crisis and people reading in a second language, as long as the worker verifies the rewrite did not silently change a fact.
• AI translation is the highest-risk subcase: fast and tempting, serving the people least able to detect an error. It requires review by a qualified human translator or bilingual staff, and unreviewed machine translation does not satisfy formal language-access obligations.
• The client-message failure modes are the invented option or instruction (the most dangerous because it sounds helpful), the wrong fact fluently stated, the overpromise, the wrong tone for the situation, and the privacy leak of sensitive PII into an unapproved tool.
• Verification has a specific checklist: check every fact against the source not the draft, hunt for the helpful invention, check the tone against the real human behind the case, protect the data before drafting by minimizing sensitive details in the prompt, and remember the worker owns every word.
• "The AI drafted it" is not a defense if a message misinforms a client, overpromises on the agency's behalf, or breaches privacy. The worker who sends the message is accountable for it.
• The time AI returns must be protected for verification and, above all, for returning the worker to direct human contact with clients, not converted into a higher caseload that industrializes a new kind of error.