Good AI decisions start with the person receiving care. A tool may save time. Yet that gain matters less if a client feels watched, unheard, or pushed aside. This lesson helps teams work through those tradeoffs together.
These cases are made up. They are for learning, not advice about a real client. This is a free training resource. It does not grant continuing education credit or prove that a team meets legal rules.
Who this lesson is for
Clinicians, supervisors, intake staff, and leaders can use this 90-minute lesson. Mixed groups help. A front desk worker may spot a privacy problem that a clinician misses. A clinician may see a care risk that a tool buyer misses.
The aim is thoughtful action. People should leave knowing what to do, who to ask, and how to explain a choice to a client.
Learning goals
By the end, participants can:
- Separate known facts from an AI tool's guesses.
- Explain a proposed use in words a client can understand.
- Name a concern about privacy, choice, fairness, or care.
- Choose a next step and name the person responsible for it.
- Write a short reason for the choice and a plan to check the result.
Facilitator preparation
Set aside 90 minutes: 10 for the lesson, 10 for the decision method, 40 for four cases, 20 for group review, and 10 for a closing exercise. For each case, use six minutes in pairs and four minutes for discussion.
Bring the team's current AI and recording rules. If there are no clear rules, say so. Do not ask participants to fill that gap with guesses about the law. Use only these made-up cases. Keep client names, records, and staff performance details out of shared notes and AI tools.
Make it safe to disagree. Ask people to explain their reasons. A fast answer is not always a sound answer. No participant needs to share a personal health story.
Use the AI Data-Handling Worksheet to trace where information goes. Use the purchasing checklist when a case raises questions about a vendor. Those tools support review; they do not approve a product.
Short lesson: useful work still needs care
Generative AI is software that makes new text or other content from patterns it learned. It can produce a clear draft that contains a false claim. It may also repeat unfair patterns. A confident tone does not prove that a statement is true. NIST's voluntary guide describes risks such as false content, harmful bias, and privacy loss. It is guidance, not a law or a clinical practice standard. NIST Generative AI Profile, July 2024
Administrative work includes tasks such as arranging schedules. Care work includes interpreting symptoms and choosing treatment. A task can cross that line. For example, a note tool may organize words, then add a diagnosis that the clinician never made. Review must catch that change.
Professional ethics also matter. The NASW Code addresses informed consent, client choice, competence, and privacy. These are professional standards for social work. Other professions have their own standards. They are not one set of legal rules for every worker. NASW ethical responsibilities to clients
HIPAA is a federal health privacy law that applies to covered health organizations and their business associates. A cloud vendor that handles protected health information for a covered organization can need a business associate agreement and other safeguards. A client's permission alone does not replace those duties. Confirm what applies to the specific use. HHS cloud computing guidance
A shorter workday or faster billing process can be worth pursuing. Measure those gains beside review time, client trust, errors, and staff workload. Never add a service or fact that did not occur to make a record look complete.
A method for working through each case
First, state the facts. Then name what is missing. Ask who could be helped and who could be harmed. Give the client a place in that discussion.
Use the five domains in the Behavioral Health Responsible AI Framework:
- Clinical Appropriateness: Does this use fit the care need and the setting?
- Privacy & Security: What information goes where, and who can access it?
- Reliability & Safety: What could be wrong, unfair, or left out?
- Human Oversight: Who checks the work and can stop it?
- Governance & Accountability: Who owns the choice, the rules, and follow-up?
This is a working framework, not a validated test or certification. A team may record its decision as Hold, Pilot with limits, Approved with conditions, or Paused or ended. The case discussions below help explain the reasons behind a choice. They do not approve tools for real use.
Case 1: A client declines recording
A clinic has approved a recording tool that drafts visit notes. Morgan, a made-up client, says, “I do not want this visit recorded.” The clinician is behind on notes. A supervisor suggests asking Morgan again because the tool could save time. A manual note option exists.
Activity: Practice a two-minute response to Morgan. Then decide what the clinician should do today. Name one question to send to the supervisor after the visit.
Answer and debrief: A useful response is, “Thank you for telling me. We can use our manual note process today. I can answer questions about the tool if you want.” Avoid making Morgan defend the choice. Repeated requests can feel like pressure, even when the team means well.
Follow the applicable recording, consent, and care rules. In this case, use the available manual process. Ask the supervisor how the team will support staff when clients decline. Do not make one client's choice carry the burden of fixing a workload problem. NASW's code specifically addresses consent before audio or video recording; that is a professional standard, while state recording laws and team rules need separate review. NASW Code, section 1.03
Ask the group how it would check whether clients feel free to say no. An anonymous question about comfort may be more useful than counting how many clients accept recording.
Case 2: A draft labels a client unfairly
A draft note says, “The client was resistant and unmotivated.” The visit facts say that Avery missed two visits because the bus route changed. Avery asked for an evening time. The AI draft leaves out both facts.
Activity: Rewrite the passage using only the stated facts. List two checks the reviewer should make elsewhere in the draft.
Answer and debrief: A grounded version is, “Avery reported missing two visits after a bus route change and asked about evening visits.” Do not add a promise, diagnosis, or reason beyond the facts.
Check for other unsupported judgments and missing context. Review whether the tool changes meaning for people with different speech patterns or access needs. One error does not prove a measured rate of bias. It does justify examining the draft and looking for a pattern in a proper review.
The clinician corrects the record before approval. If this error repeats, the team should consider pausing that use while it investigates. Describe the issue through the approved reporting process. Share the least information needed. Discuss how a label could affect future care when another worker reads it.
Case 3: A client brings chatbot advice
Jordan brings a chatbot conversation to a visit. It says that Jordan's symptoms “definitely mean” a certain disorder. Jordan feels relieved to have an answer and worried that the clinician will dismiss it. The clinician has not yet assessed the concern.
Activity: Write an opening response. Name what to explore next and what the chatbot cannot establish.
Answer and debrief: Try, “It sounds like that answer gave you some relief. Let's look at what felt right to you and what has been happening.” Explore Jordan's goals and experiences. Explain that fluent text is not an assessment. The clinician uses their usual care process rather than treating the chatbot's label as a finding.
Do not shame Jordan or turn the visit into a debate about technology. Ask how the advice affected choices or feelings. If reviewing the chat would help, discuss what Jordan wants to share. Do not require an entire chat history that may include unrelated private details. Avoid uploading it into another tool without an approved basis and process.
Ask the group how this approach can protect the working relationship while still checking an unsupported claim. See the guide to clients using ChatGPT for more context.
Case 4: Pressure to finish a billing note
A note draft includes “practiced a coping skill for 15 minutes.” That activity did not happen. A staff member says the phrase helps show why the visit should be paid. The clinician is asked to sign before leaving.
Activity: Draft a response to the staff member. Decide what must happen before the note is approved. Name a longer-term team action.
Answer and debrief: Say, “That activity did not happen. I need to correct the note before signing.” Remove the false statement. Check the remaining draft against the visit facts. Follow the normal review process for billing questions and record corrections.
A payer's documentation needs do not justify invented care. If the team needs clearer notes, improve the process for recording what actually happened. Ask a qualified billing lead to review the real requirements. This lesson does not decide payment eligibility or interpret a particular payer contract.
Discuss the power difference in the case. Staff need a way to raise concerns without being pushed to sign. A faster process is useful only if the final record is sound.
Group review: turn a choice into action
Each pair shares one case using four short sentences: the concern, the immediate action, the person responsible, and the follow-up check. Other pairs ask what fact could change that choice.
For example: “The draft adds an activity that did not occur. The clinician will correct it before signing. The supervisor will review why it was added. The team will check a small set of approved practice drafts for the same issue.”
Do not use a live client record for this exercise. A small review can reveal a problem; it cannot prove that every future draft will be accurate.
Closing exercise and answer notes
Ask each person to complete this sentence: “If a tool's output conflicts with the visit facts, I will…” A sound answer includes stopping approval, checking the facts, correcting or replacing the draft, and reporting a repeated problem through the team's process.
Then ask, “What if a client does not want a tool used?” A sound answer includes listening, explaining real options, following applicable rules, and using an approved alternative when available. It avoids promising an option the team cannot provide.
If an answer focuses only on saving time, return to client trust and record accuracy. If it says AI must never help with any task, ask the group to name a bounded use that could be reviewed well. The goal is a careful choice based on the task and setting.
After the lesson
Within two weeks, the facilitator and a named supervisor should review one unresolved question. Choose a small next step, such as making the recording choice clearer or fixing the error-reporting route. Tell staff what changed.
Ask whether the change makes care and daily work easier. Check for added review time as well as time saved. Keep listening to clients. A good process makes room for both useful tools and human judgment.
References and scope
Sources checked October 11, 2026. The NIST guide is voluntary guidance. HHS explains HIPAA duties for organizations within its scope. NASW provides social work professional standards. Local law, other professions' standards, payer rules, and team policies may add requirements. The cases and suggested training steps are original teaching examples.