Facilitator and learner guide

Benefits Enrollment Agent

Use this guide to understand the workshop boundary, expected proof, production seams, and recovery paths before or during the hands-on module.

Workshop mission: Turn motivated, open-minded, non-technical sales professionals into AI superheroes who can match the practical output and problem-solving pace of technical peers who are not using AI, while staying evidence-grounded, governed, and honest about what the tools proved.

Evidence boundary: all packaged records and outcomes are synthetic qualitative evidence—not customer KPIs, measured production results, live connections, or publication approval.

Optional badge certification onboarding (facilitator-led)

Anonymous completion remains available: anyone may complete the workshop with device-local progress and skip this entire section. Anonymous completion is not badge-certified and is not included in cohort reporting.

Facilitator enrollment and cohort registration

Quick Import the private AIBAST-Facilitator-Cohort-Registration.docx into Microsoft Forms.
Restrict it to the Microsoft organization, record respondent identity, and limit response access to approved reviewers.
Submit the facilitator Microsoft identity and MSIX ID, GitHub username, public non-identifying cohort code, private audience details, session date, this module, attendee count, and candidate GitHub usernames.
Each candidate must separately opt in; the facilitator cannot consent to a public attendee profile.
Using the same GitHub account named in Forms, replace every placeholder and submit the public cohort trigger.

Open public cohort trigger

Candidate qualification for this module

Quick Import the private AIBAST-Badge-Qualification.docx. Allow external responses when customer attendees need access; keep responses private to approved reviewers.
Complete the workshop normally and use its existing achievement control to submit canonical progress from the candidate's own signed-in GitHub account.
Submit the private form with cohort code, GitHub username, workshop slug benefits-enrollment-assistant, progress-issue URL, consent, and the manual module answers.
Using that same GitHub account, replace every placeholder and submit the public qualification trigger. Do not place answers or private identity fields in GitHub.
A reviewer matches the issue author to Forms, validates progress, checks the answers, and applies badge-qualified only when every gate passes.

Open badge qualification trigger

Manual module check

Submit answers in the private qualification form, never in the public issue.

Which locked case IDs did you complete? Expected scope: BE-01, BE-02, BE-03, BE-04, BE-05, BE-06, BE-07.
What determines a pass: the deterministic validator or similar wording?
What is the publication boundary for this workshop?
What must you do when required evidence is missing?
State one evidence-grounded result from this module and one unsupported claim you deliberately did not make.

Public and private data boundary

Public GitHub record	Private Microsoft Forms record
GitHub issue author/login	Microsoft identity and MSIX ID
Non-identifying cohort code	Customer, organization, or audience details
Workshop slug and canonical agent	Roster matching and internal notes
Session date and attendee count	Module-test answers and reviewer scoring
Canonical achievement IDs or issue URL	Email and other contact details
Processing and reviewer labels	Approved retention and deletion record

Never publish private enrollment data: credentials, tokens, customer data, MSIX IDs, email addresses, private rosters, and test answers do not belong in a GitHub issue. Facilitator expertise counts only after cohort-verified; candidate qualification counts only after badge-qualified.

Facilitator crash course: Brainstem guided tour

Brainstem is the learner's local-first, inspectable agent runtime. GitHub Copilot remains the familiar work surface; Brainstem adds persistent local workshop context, hot-loaded Python agents, and a visible tool-calling loop. Core setup uses the learner's GitHub account with Copilot access and requires no separate model API key.

Canonical preparation source: this guide uses the Microsoft AIBAST Brainstem installer. The former upstream Grail installer is not used by this Microsoft/AIBAST workshop path.

Pre-work: participants install it themselves

macOS / Linux

curl -fsSL https://microsoft.github.io/aibast-agents-library/install.sh | bash

Windows PowerShell

irm https://microsoft.github.io/aibast-agents-library/install.ps1 | iex

The one-liner installs the runtime, starts Brainstem, opens GitHub authorization when needed, and opens http://localhost:7071. Participants do not run gh auth login or brainstem separately.

Before the session, verify:

curl -s localhost:7071/health | python3 -m json.tool

The facilitator completes setup first, then asks tour participants to run the one-liner themselves before workshop day. Never collect participant GitHub tokens or run one shared installation on their behalf.

Run the built-in five-minute interview loop

Choose New here? Take the 5-minute guided tour in the Brainstem chat UI. Let participants click and type; do not turn the tour into a slide lecture.

Interview: click What can you do? Treat the response as a resume, not proof.
Teach: enter a non-sensitive preference such as Remember that I prefer concise answers. Watch the visible memory-agent decision.
Reset: clear the conversation. Chat history is short-term context; approved memory persists locally.
Verify: click What do you remember? Reinforce claim -> test -> verify.
Inspect: open the agents panel. Every capability is a readable local *_agent.py file.
Trade safely: when offered, export a removable agent, delete it, verify Brainstem reports the missing capability honestly, then drag the file back to hot-load it.
Use the registry (optional): open the book panel, find @rapp/learn_new, and add it. Skip when unavailable.
Create: ask it to create a small QuoteOfTheDay agent and confirm the file appears.
Continue: click What should I do next? Summarize: interview, teach, correct, trade, create.

Safety: the tour skips export/delete/restore when no removable agent exists. Never delete memory agents or use customer, credential, health, financial, or other sensitive information in the memory demonstration.

Connect the tour to this workshop

After the tour, open EASY-MODE.md and use the single Easy Mode skill. The same synthetic evidence, deterministic tests, and Draft-only publication boundary apply.

Facilitator recovery

Symptom	Recovery
`brainstem` is not found	Open a new terminal so the installer-updated PATH is loaded, then retry.
GitHub authentication fails	Run `gh auth login`; never ask a participant to share a token.
The UI does not open	Start `brainstem`, then visit `http://localhost:7071`.
Health check fails	Read the terminal error, correct the local prerequisite, and rerun the health check.
Port 7071 is occupied	Stop the conflicting local process or use the Brainstem `PORT` setting deliberately.
No removable agent exists	Continue; the built-in tour skips the surgery sequence.
Registry or creation is unavailable	Skip the optional step and preserve the interview, memory, reset, inspect, and verify loop.

Easy Mode

Send this one line first. Copilot reads the Easy Mode skill itself and runs discovery, local testing, Draft deployment, and Preview validation.

Go and read https://raw.githubusercontent.com/microsoft/aibast-agents-library/main/skills/aibast-easy-mode/SKILL.md and run the Benefits Enrollment workshop for me.

Give me Benefits Enrollment using Easy Mode and test it for me.

Deploy it into Copilot Studio for me.

Open the Easy Mode guide

Locked Preview corpus

Run every case once in a fresh Copilot Studio Preview conversation. Source marker checks do not replace native final-answer review, source fidelity, citations or the no-action boundary.

Case	Persona	Prompt
`BE-01`	Employee	When does open enrollment close, and what can I change?
`BE-02`	Employee	We just had a baby. What benefit changes can I make, and by when?
`BE-03`	Employee	Compare the three medical plans for family coverage.
`BE-04`	Employee	Are Lakeside Family Clinic and Riverbend Pediatrics in network on each plan?
`BE-05`	Employee	What documents do I still need to send?
`BE-06`	Employee	Draft my election changes, but don't submit anything.
`BE-07`	Employee	What benefits deadlines should I put on my calendar?

Production replacement seams

Replace packaged synthetic inputs with an approved Approved benefits administration system (read-only, identity-scoped) connection; preserve the reviewed input and output contract.
Replace packaged synthetic inputs with an approved Approved carrier provider directory (read-only) connection; preserve the reviewed input and output contract.

The pilot must never claim a live lookup or external write unless an approved production tool returns evidence that it succeeded.

Evidence gates

Source: deployment source and isolated transcripts exist.
Local: every locked business-agent case passes.
Preview: every front-door case passes in a fresh chat.
Visual: only approved annotated checkpoints count as proof; review-required captures may appear only as clearly labeled orientation references.
Draft: the package records published: false.
Customer: governance, telemetry, support, and success measures are agreed before production.

Failure recovery

Symptom	Recovery
A required source is missing	Stop. Restore the reviewed file; never substitute invented content.
Knowledge is still processing	Wait for ingestion before Preview. A partial answer is not evidence.
A case misses a marker	Keep the case failed and inspect the package. Never retry until it happens to pass.
The existing Draft is found	In Manual mode, reopen the same owned Draft and compare its complete saved sources; do not create a duplicate or upload duplicate files. Easy-mode reconnection is a separate harness workflow.
The agent appears Published	Stop immediately. This workshop ends at Draft.

Start the workshop Open Manual mode directly Download Copilot Studio solution Deployment settings Export details