Field Report: 90 Days of AI-Assisted Intake at a Small Workforce Program

This is the kind of post we promised when we launched: an honest field report, including the parts that didn’t work. The pilot described here is composited from two grant-funded workforce program engagements that ran in early 2026. Details are changed to protect participant privacy and program identity, but the workflow, metrics, and lessons are real.

The Setup

Organization. Workforce program serving roughly 180 active participants annually, six full-time staff including two case managers, one job developer, an intake coordinator, a program director, and a fiscal officer.
Funding. Mixed: WIOA Title I formula funds, a state workforce-development grant, one local foundation award.
Workflow piloted. The first 45 minutes of new participant intake — specifically, turning the intake interview into the structured record their case management system requires.
Tool. A single enterprise AI subscription with a no-training data clause. No PII pasted into consumer tools.
Duration. 90 days, with 41 new intakes processed during the pilot window.

The Workflow

The intake coordinator continued doing the intake interview as before — in person, in a private office, with the participant. After the interview, instead of typing the intake record directly into the case management system, she dictated a 2–3 minute summary into a voice recorder. The recording was transcribed locally, the transcript was de-identified (the participant’s name replaced with initials), and the de-identified transcript was passed through a fixed prompt that produced a structured first draft of the intake record. The coordinator reviewed the draft, corrected it, and then typed the final version into the case management system as always.

What Worked

Time saved per intake. Average 18 minutes saved per intake by week 6, settling at 22 minutes by week 12. That’s roughly 30% off the post-interview documentation step. Over 41 intakes, the program recovered about 15 hours of coordinator time.

Consistency. Random review by the program director found intake records were noticeably more uniform in structure and tone. New participants whose intakes happened on busy days no longer received shorter, more abbreviated records than those who arrived on quiet days.

Staff comfort. The coordinator went from a 2/5 on AI comfort at week 1 to a 4/5 at week 12. Both case managers asked, unprompted, whether the same workflow could be used for follow-up case notes.

What Didn’t Work

Barrier-coding accuracy. The structured intake record requires the coordinator to code each participant’s identified barriers using a specific funder-defined list. The AI’s first drafts coded barriers correctly about 75% of the time, but when it was wrong, it was wrong in a confident way — picking a plausible-sounding but ineligible barrier label. The team responded by adding a hard rule to the prompt: “Do not assign barrier codes. List the participant’s stated and observed challenges in plain language, and I will code them manually.”

Inferred employment history. In about a quarter of cases, the AI “helpfully” filled in employment history details that the participant had not actually stated — plausible inferences from context (e.g., assuming a stated job had ended when it hadn’t). This was caught in review every time but added cognitive load to the review step. The fix was an explicit instruction: “Use only what the participant stated. If something is unclear, list it under ‘Items to clarify’ instead of inferring.”

First-week awkwardness. The coordinator reported feeling “watched” during the dictation step in the first two weeks — like she was performing rather than working. This faded by week 3, but if we ran this again, we’d warn staff about the adjustment period in advance.

The Three Warnings

Don’t let AI code anything that has a funder-defined data dictionary. Plain-language summaries that staff then code are safer.
Plan for at least three weeks of friction. The numbers don’t move until staff stop second-guessing the tool.
Make your AI use policy visible before you start. A staff member asked, mid-pilot, what would happen if a participant asked whether AI was used. The team hadn’t agreed yet. They do now.

Documentation and Compliance

The AI subscription was a direct cost to the workforce grant, documented per the template in Documenting AI Costs in Federal Grant Reports. The dictation recordings were retained for the same period as the intake record itself, encrypted at rest, and access was limited to the intake coordinator and program director. No participant identifiers were ever pasted into the AI tool.

What’s Next

The program has extended the workflow to follow-up case notes for the next 90 days, using the structure described in AI for Case Notes. We’ll publish the follow-up field report when that round is complete, including the moments where it didn’t work.

StrategicalAI.org