By Cody Saunders, LMSW · October 11, 2026
Turn a draft into useful work
A quick AI draft is only the start of a task. Someone needs to check the facts, protect the information, and decide whether the result fits its purpose. This lesson helps care teams practice that full workflow. The goal is useful work that leaves more room for people, rather than a larger pile of drafts to fix.
Generative AI creates content such as text, images, or sound. Here, we focus on text drafts. Learners practice with public information and made-up cases. They learn to define a task, check its data limits, compare a draft with the source, and decide whether to use, revise, or hold it. The lesson builds on basic AI literacy; it is about doing the work responsibly, not learning a list of terms.
Who this lesson is for
This free 75-minute module is for clinicians, supervisors, and staff who already know basic AI terms. It can be taught with an approved tool or with printed drafts. It does not approve a product or authorize any new use of client information. It does not promise continuing education credit or certify a learner’s skills.
The facilitator should know the host’s approved tools and reporting process. Before the session, name a contact for clinical questions and a contact for data questions. Prepare the public-text exercise, fictional handout, and trial results below. Keep real records out of the session. If someone starts to describe a real client, ask them to stop and use the supplied case instead.
Learning objectives
By the end, learners should be able to:
- Define a task, source, audience, and permitted information before drafting.
- Review an output for source accuracy, meaning, and practical fit.
- Distinguish permission to draft from approval to use a final document.
- Explain when to pause and report a problem.
- Count review and correction time when testing whether a tool helps.
Timing and facilitator guidance
0–8 minutes: Set the task. Ask learners what a useful final document should accomplish. Show how a vague request such as “write a client guide” leaves too much open. Name the audience and purpose before choosing a tool.
8–20 minutes: Teach the workflow. Explain the steps below. Ask learners to identify the source of truth and the final reviewer for a public staff guide. Then ask how those roles would differ for a care record.
20–35 minutes: Practice source-based drafting. Use Activity 1. Have pairs mark each added or altered claim before rewriting the draft. If using a live tool, give it only the fictional policy below.
35–50 minutes: Review fit for a person. Use Activity 2. Ask learners to discuss why a draft can be accurate in general and still unsuitable for a specific purpose. Keep the discussion tied to the fictional care context.
50–62 minutes: Respond to a workflow change. Use Activity 3. Practice a short report to the named team lead. Explain the host’s actual incident process rather than improvising one for a real event.
62–70 minutes: Count the whole effort. Use Activity 4. Invite learners to compare time, quality, and staff experience. Ask what else must be measured before a decision.
70–75 minutes: Teach back and close. Each pair explains the draft-to-final path and one stop condition. Give the actual help contacts. Assign follow-up practice only within the host’s approved use.
A full draft-to-final workflow
Start with one defined job. State what the final document is for and who will read it. Name the approved source and the person who can resolve unclear points. Keep care decisions with the people qualified to make them.
Next, check the tool, account, feature, and data. An approved public-writing use does not cover visit audio or private notes by default. Do not move information into a personal account because the approved service is slow. Ask for an approved alternative.
Write a prompt, or request, that names the task and limits. For example: “Use only the policy below to make a short new-staff guide. Keep required steps. Do not invent dates or contacts. Mark unclear points for review.” This guides the work but cannot guarantee that the model follows it.
Compare the output with the source. Check numbers, order, names, and the difference between a reported fact and an assessment. Look for missing limits. Open any cited source and check what it actually supports. NIST describes false or invented content as confabulation in its Generative AI Profile. A polished answer may still contain it.
Then check fit. Is the language clear? Does the draft assume resources the audience may lack? Does it label someone unfairly? A document can repeat source facts yet fail to meet the reader’s needs. Revise the draft and check the new version too.
Finally, the assigned owner approves the work through the usual process. Store the final version in the right place. Keep it distinct from drafts. Record problems through the approved channel. A request for better wording is not the same as approval to release a care document.
Activity 1: The draft adds a rule
Use this made-up policy: “Staff request leave through the work portal. A supervisor checks coverage. Approval must be received before travel is booked. Urgent requests go to the supervisor by phone.”
The fictional AI draft says: “Book travel first. Request leave at least two weeks ahead. Urgent requests can be sent by text.”
Ask pairs to mark each change. Then write a short guide that follows the policy. Finish with one question for the policy owner.
Answer and debrief: The draft reverses the travel rule, invents a two-week deadline, and changes phone to text. A correction should keep the four source steps. A useful question is whether the portal has an approved link or whether any request deadline exists. Do not fill that gap from memory. This office task contains no client data, but an incorrect guide can still disrupt staffing.
Activity 2: The draft does not fit
A fictional therapist wants to simplify a handout about a bedtime routine already discussed with a fictional client. The source includes choosing a regular time and reducing distractions where possible. The client works changing shifts and shares a busy home. The AI draft says: “Go to bed at 9 p.m. every night. Use your private bedroom. If you cannot do this, you are not trying hard enough.”
Ask learners to identify added instructions, practical barriers, and judging language. What needs the therapist’s judgment rather than another automatic rewrite?
Answer and debrief: The fixed time and private room are unsupported assumptions. The blame statement is inappropriate. The therapist needs to adapt the handout to the person’s actual circumstances and agreed care plan. A request to use respectful language may help with wording, but it cannot decide the right clinical advice. The exercise is about reviewing fit, not recommending sleep treatment. Social workers’ duties around competence and client participation are addressed in the NASW Code of Ethics. These are professional standards, not a substitute for applicable law or team policy.
Activity 3: A new feature asks for more data
A fictional team approves a tool for public staff guides. During the exercise, the tool offers a feature that records meetings. A learner wants to use it at the next case meeting. The team has not reviewed the recording feature or its data recipients.
Ask learners to write a short message to the team lead using only the facts above. Name what should pause and what can continue.
Answer and debrief: Hold the new case-meeting recording use. Keep the public-writing use within its existing approval. A useful report names the feature, proposed use, expected information, and missing review. Do not include client details in the report. If a real data event occurs, follow the host’s established incident process promptly; this fictional exercise does not determine whether a breach occurred or what legal notice is required.
The HHS cloud guidance explains HIPAA duties for covered organizations and business associates. Appropriate agreements and safeguards may be needed for handling protected health information. Team approval must account for applicable requirements; a patient’s permission or a product’s familiar name alone does not settle them.
Activity 4: Did the tool save effort?
Use these made-up results for one public-guide task. The usual workflow took 20 minutes to draft and 5 minutes to review. The AI workflow took 4 minutes to draft and 18 minutes to review and fix. Both final versions met the same quality check.
Ask learners to calculate total time. What can this small example tell the team? What can it not tell them?
Answer and debrief: The usual task took 25 minutes. The AI task took 22 minutes. The difference is 3 minutes for this example, not 16. It does not establish savings across the team or benefits to clients. Repeat a fair trial with comparable tasks. Track errors, missed details, staff effort, and whether review fits the workload. A workflow that frees time for care is worth exploring, but its value must be shown in use.
Check learning and carry it forward
Ask each learner to explain the workflow using a new public-text task. Listen for the approved account and data, source comparison, practical fit, final owner, and a reporting path. If a learner says the tool can approve its own answer, return to Activity 1 and ask who checked the invented rules.
For follow-up, choose one approved task and a small set of public examples. Agree on the quality check before drafting. Keep total effort and corrections visible. Name a date and owner for the review. Stop if the task or data changes beyond its approved limits. These are suggested teaching steps, not a validated skills test.
References and related resources
Sources checked October 11, 2026. All cases, policies, drafts, and time figures are fictional. The lesson does not establish clinical benefit, billing improvement, cost savings, legal compliance, or certification.
NIST Generative AI Profile, July 2024: voluntary guidance on risks and review actions. It is not product approval.
NASW Code of Ethics: Responsibilities to Clients: professional standards on client participation and competence; practitioners must also consider their own profession and setting.
HHS guidance on HIPAA and cloud computing: guidance on duties for covered organizations and business associates. It does not approve a specific AI service.
Use the data-handling worksheet and purchasing checklist before considering sensitive uses. See Clinician AI Literacy for introductory teaching and Work With Me to discuss free training.