AI tools can save case managers real time, but case files hold some of the most sensitive data a workforce program keeps. Before any tool touches participant records, answer five questions: what data it sees, who checks its work, where a person must decide, what the vendor keeps, and how participants are told. The NIST AI Risk Management Framework gives you a free, credible structure for that conversation.
"Can we use AI for case notes?" It is a fair question. The honest answer is that the tool is rarely the hard part. The hard part is deciding, in writing, what the tool is allowed to touch and who stays accountable for the result.
Case files in workforce and human services programs carry Social Security numbers, income and benefit records, disability and health information, criminal history, and the story of a person's hardest season. That information was shared to get help. It deserves a plan before it meets any AI tool.
Here are the five questions we would put in front of any board, nonprofit, or agency team before a pilot starts.
1. What data will the tool actually touch?
Start with an inventory, not a demo. List every field the tool would see, read, or store: names, Social Security numbers, dates of birth, eligibility documents, case notes, assessment results, wage information.
Then sort each proposed use into one of three groups:
- No participant data. Drafting a job fair flyer, summarizing a public policy document, outlining a workshop. Lowest risk and a sensible place to start.
- De-identified data. Looking for patterns across a cohort with names and identifiers removed. Moderate risk, because poorly de-identified data can still point back to a person.
- Identifiable participant data. Summarizing a case file, drafting an employment plan, flagging missing eligibility documents. Highest risk. These uses need a full answer to all five questions.
If you receive U.S. Department of Labor Employment and Training Administration funds, read TEGL 39-11, DOL's guidance on handling and protecting personally identifiable information (PII), alongside your state and local PII policies. Then ask a plain question: does pasting a case note into this tool fit those rules? If no one on your team can answer with confidence, the answer is not yet.
2. Who reviews what the tool produces?
Generative AI can produce text that sounds right and is wrong. NIST's Generative AI Profile calls this risk "confabulation": false content stated with confidence. In a case file, that could be a wrong date, an invented barrier, or a credential the participant never earned.
Name a reviewer for every approved use. Decide what they check (facts, tone, completeness), how they record that the review happened, and what happens when they find an error.
A summary nobody reads closely is not a time saver. It is a liability with a timestamp.
3. Where must a person make the decision?
Some decisions belong to people, no matter how good the software gets. In workforce and human services programs, we draw the line here:
- Eligibility determinations and enrollment
- Sanctions, terminations, and reductions in services or support
- Approvals of training funds, supportive services, or participant payments
- Anything a participant can grieve or appeal
AI can help assemble information for those decisions. A trained staff member makes them, documents the reasons, and owns them. Write that line into your policy so it does not erode quietly when caseloads spike.
The NIST framework treats this as a design choice, not an afterthought: it expects organizations to define, assess, and document their processes for human oversight.
4. What do the vendor's terms say about your data?
Free and consumer versions of AI tools can carry different data terms than business or government agreements. Read the actual terms of service and data agreement, not the marketing page, and get answers in writing.
| Ask the vendor | Why it matters |
|---|---|
| Is our data used to train or improve your models? | Participant information should not become part of someone else's product. |
| How long do you keep what we enter and what the tool produces? | Retention should match your records schedule, not outlast it. |
| Where is data stored, and who can access it, including subcontractors? | You stay responsible for data you share. |
| How and when will you notify us of a breach? | Your funder and state may set their own notice requirements. |
| Can we export and delete our data when the contract ends? | Exit terms are easier to negotiate before you sign. |
| Can you meet the terms of our grant and data-sharing agreements? | Your obligations do not disappear because a vendor is involved. |
5. How will participants know?
People who come to a workforce program are often asked to share a great deal to get help. They deserve to know how their information is used. Decide up front:
- Will your intake and consent forms mention AI tools, in plain language?
- Can a participant ask that their file not be processed by an AI tool, and what happens if they do?
- Who answers when a participant has questions?
NIST lists accountability and transparency among the characteristics of trustworthy AI, alongside privacy and fairness. It is also simple respect. Trust is the real infrastructure of a workforce program.
A framework to organize the work
The NIST AI Risk Management Framework (AI RMF 1.0), published in January 2023, is voluntary, free, and written for organizations of any size. It organizes AI risk work into four functions:
- Govern: the policies, roles, and accountability that apply across every AI use.
- Map: the context, the people affected, and what could go wrong.
- Measure: testing and tracking how the tool performs.
- Manage: acting on what you learn, including stopping a use that is not working.
The five questions above are a practical first pass at Govern and Map. Measure and Manage follow once a pilot is running. NIST's companion Generative AI Profile (NIST AI 600-1), published in July 2024, adds risks specific to chatbots and text generators, including data privacy and confabulation.
AI can give case managers back time for the work only people can do: listening, problem solving, and showing up for participants. Getting the guardrails right first is how you keep that promise. If your team is weighing an AI pilot and wants a second set of eyes on these questions, start a conversation with us.
Sources
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023
- NIST, AI Risk Management Framework overview and resources
- NIST, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, July 2024
- U.S. Department of Labor, Employment and Training Administration, TEGL 39-11, Guidance on the Handling and Protection of Personally Identifiable Information, June 28, 2012
Maria J. Posey is CEO and Founder of Viking Impact Advisors. Contact Us.
