A practical talk about fitting AI into care workflows while keeping people in charge. The focus is how a team sets limits, protects information, checks results, and tests whether a tool helps. This is a proposed session for learning, not a claim that it has already been delivered.
Program abstract
A fast AI draft can look helpful. The real test is what happens next. Does it keep the client's meaning? Can a trained person check it? Does the whole process reduce work, or add more fixes?
In this talk, Cody Saunders, LMSW, helps care teams plan responsible AI use around one clear task. Learners follow a made-up note workflow from its purpose to its final review. They consider what data the tool receives, who can see it, and who approves the output. They also see why an office task, a note draft, and a treatment suggestion need different limits.
The session uses the five areas of the Behavioral Health Responsible AI Framework: Clinical Appropriateness, Privacy & Security, Reliability & Safety, Human Oversight, and Governance & Accountability. Learners apply those areas to a small pilot, or trial. They practice naming a missing control and choosing a next step.
The goal is useful support for care: less busywork, clear records, and more room to connect with clients. Benefits must be tested, including review time, fixes, and costs. Learners leave with a short plan for one task and a better sense of when to proceed, narrow the use, or wait. No tool is treated as a substitute for clinical judgment.
Audience and format
For clinicians, supervisors, care operations staff, and leaders. Suggested length: 40 minutes, including a workflow exercise and questions. Suitable for an in-person or online session. A basic understanding of AI drafts helps, but terms are explained as they arise. No live AI account or client data is needed.
This talk focuses on the full care workflow. The AI Basics talk introduces key terms. The Human Review talk focuses more deeply on checking output. The Clear AI Rules talk focuses on team policy and ownership. This session connects those needs through a practical care-use plan.
Learning objectives
After the session, learners should be able to:
- Define one AI task with a clear allowed role and limit.
- Identify a data or review gap that must be resolved before live use.
- Apply the master framework's five areas to a made-up workflow.
- Choose one benefit measure and one reason to pause a pilot.
Proposed session flow
- First 5 minutes: define the work need. Compare a public office draft, a care note, and a treatment suggestion.
- Next 10 minutes: walk through the five review areas. Connect each to people, data, and the final result.
- Next 10 minutes: complete the made-up case below. Name the gap and assign a next action.
- Next 10 minutes: build a pilot plan. Include total work time, errors, client and staff feedback, costs, and a backup process.
- Final 5 minutes: questions and a task-specific next step.
Sample case and debrief
Made-up case: A team wants to reduce after-hours note work. Its proposed AI tool will turn visit information into a note draft. Clinicians will review before signing. The team has tested made-up cases. The tool sometimes adds findings that were not supplied. The team has not shown that reviewers can reliably catch these errors. It has also not finished reviewing the tool's data storage and access.
Ask learners to name the gaps in the five areas. Then choose a decision state for live use: Hold; Pilot with limits; Approved with conditions; or Paused or ended. These are the same proposed states used in the master framework.
Answer notes: Live use remains on Hold. The data review is unfinished. Important output errors and the review process also need work. A planned launch date does not resolve those gaps.
The use owner should bring privacy and security staff into the data review. Clinical reviewers should test whether they can find and correct unsupported findings. The team can keep using made-up cases under its rules. It should narrow the tool's role or change the process if the controls cannot support the task.
For a later approved pilot, name users, data types, reviewers, a review date, and stop rules. Keep the usual note process available. Compare total note time, including review and fixes. A faster first draft alone does not establish a benefit.
Patient relationships and daily work
The session asks teams to consider how the tool affects the visit. Staff should be able to explain the approved use in plain words and handle client questions under the rules that apply. A device should support attention rather than pull it away.
Billing examples stay tied to actual care. A draft may help organize information or flag missing fields. It must not add services or clinical facts to fit a claim. Clinical and billing review have distinct roles. Lower costs, fewer claim errors, and lighter workloads are goals to measure, not promises made by this talk.
Evidence and scope for the presenter
HHS explains that cloud services handling electronic protected health information on behalf of a HIPAA-covered entity have business associate duties. A suitable agreement and other applicable HIPAA safeguards are needed. A seller's “secure” label alone does not approve the task. Have the appropriate staff review the service, data path, rules, and contracts. HHS HIPAA cloud guidance.
NIST's voluntary generative AI guidance discusses made-up output and actions for managing risk. It does not establish a product's error rate or prove that a short local pilot is safe for every client. NIST Generative AI Profile, AI 600-1, July 2024.
The five areas and decision states come from Cody's Behavioral Health Responsible AI Framework, version 1.0. It is an evolving planning guide, not a validated test, certification, or legal opinion. The case and pilot steps are proposed teaching examples. Laws, professional standards, and local policies still need setting-specific review.
Resources to share
- AI and Clinical Documentation: examine the full note workflow.
- Can I Put Client Information Into an AI Tool?: understand data review needs.
- AI data-handling worksheet: map data types without adding real client details.
- 25-question purchasing checklist: request proof before a purchase.
Sources checked October 6, 2026. No continuing education credit, certification, legal compliance, care outcome, or savings is promised. Use made-up cases. Recheck sources and host policies before presenting.