# Companion-video brief — Hire and supervise an AI employee

Use the [full course](../appmint-hire-and-supervise-ai-employees.md) for narration. This is a recording brief, not a completed video. The current evidence consists of real interface stills, saved request/readback results and a genuine isolated model task. Capture fresh motion from the matching workflow; do not reuse unrelated assistant, chat or Vibe clips.

## Story

Open on the manager's desired result: a project review that identifies missing information without changing customer records. Introduce the employee's separate identity, then follow that identity through draft, access, runtime, one task, a human decision and pause. Keep “saved”, “connected” and “useful work completed” visually distinct.

Use Appmint navigation for the Appmint version. For the BusinessMade introduction, record **Admin → AI employees** and explain the shared management controls. Do not show BusinessMade users an Appmint-only menu as if it were their own navigation.

## Available shots and narration anchors

All filenames below are in [the actual AI Employee captures](../../application-fixes/assets/ai-employees-local/).

| Scene | Actual still | Explain while recording |
| --- | --- | --- |
| The outcome | draft-created-setup.png | Creating a named record starts the setup; it does not finish a job. |
| BusinessMade entry | businessmade-admin-entry.png | Admin contains the feature in this product. |
| Defaults | settings-current-runtime.png | Read inherited values and the actual provider. Use the corrected Session manager display. |
| Role choices | templates-current-runtime.png | Hire uses a template; our coordinator starts blank. |
| Create | draft-before-create.png | Name, optional handle/supervisor, specific description, hours, budget and four approval categories. Scroll at a readable pace. |
| Saved profile | draft-profile.png | Show persisted values, then reload. Do not pass off an unsaved form as successful setup. |
| Connection milestones | draft-created-setup.png | Highlight Created here, Given access, Provisioned, Switched on, Signed in, First job separately. |
| Access | draft-access-locked.png | No additional groups; mandatory AI identity; login remains locked. |
| Instructions | instructions-current-runtime.png | Platform guidance versus organisation instructions and employee scope. |
| Real model decision | actual-model-approval-waiting.png, actual-model-rejection-completed.png | Human Reject, worker resumes and records no action. Done describes the finished task, not approval of its proposal. |
| Record attachment | attachment-before-send.png, actual-model-attachment-completed.png | Existing Task selected by exact title; marker comes from its note, not the prompt. Read-only operation checked with employee identity. |
| Private document | private-upload-before-send.png, actual-model-private-file-completed.png | Change Public to Private before Upload; show the filename, Send, result marker from the file and the persisted work note. |
| Managed provider | managed-before-provision.png, managed-provider-checked.png | Actual Provision and Check provider; distinguish running process from paused employee login. |
| Real model result | actual-model-completed.png | Show Done, the correct missing contact-form check, recorded steps and rounded cost. This is genuine model output from an inline exercise. |
| Work assignment | give-work-before-send.png, give-work-queued.png, give-work-cancelled.png | First line names the job; Send saves even when paused; cancel the unwanted queued exercise. |
| Persistence | draft-reopened.png | Reopen the original employee rather than creating a duplicate. |

The original draft is `tutorial-coordinator-20260924`; it must remain inactive after filming its setup sequence. A separate manual lifecycle fixture may be used for persisted work-state tests. Label such footage **Local application lifecycle rehearsal — manually driven worker, no model execution**. It is evidence for controls, not footage of autonomous AI solving a business task.

## Genuine worker evidence and motion to capture

A controlled runtime completed the inline fictional launch checklist with a real model. Use [the runtime report](../../application-fixes/ai-employees-actual-model.md) and its exact input/output to reproduce that chapter. A result showing a missing contact-form test is a review of supplied facts, not a live test of the website. Record these actions as fresh motion; do not turn stills into a claim that a full recording already exists:

1. Review access and a permitted budget, then show the actual provider connection without secrets.
2. Switch on; show real sign-in/hello and the worker's connection identity.
3. Give the tested inline checklist task once and open the saved job. Both the record and private-file branches have passed; record their distinct picker/upload sequences. Keep file paths, credentials and temporary download URLs out of the footage.
4. Follow actual steps and result. Compare each reported fact with the original record and verify no unintended change.
5. Show a human approval and rejection with separate harmless tasks. Show the post-decision result; a clicked button is not the end of the story.
6. Pause, show locked access and verify no new work acquisition using the employee's own session. Use managed-final-paused.png and managed-provider-stopped.png for the reviewed ending: empty queue, locked employee, stopped runtime. Treat any already completed external action separately.

Never film the one-time sign-in kit, bearer tokens, passwords, authenticator secret, API keys or private customer data. Use a fictional dataset. Do not manufacture an online badge or completed result for a smoother edit.

## Graphics and pacing

Use the course's simple lifecycle diagram, highlighting the current stage while the real screen stays visible. A second diagram can distinguish **job description → intended responsibility**, **groups → permitted operations**, and **approval → a person's decision**. These graphics explain the interface; they must not impersonate screenshots.

Give each chapter a concrete question: “Who is this employee?”, “What can it access?”, “Is a worker actually connected?”, “What did it do?”, “How do I stop it?” Keep the name, status and section label visible during close-ups. Pause on saved results long enough to read. Include the zero-budget and Only these-with-no-selection pitfalls beside their actual controls.

## Final recording checks

Verify employee names and fixture state before recording; never replay Create against an existing handle. Do not enable global instructions or shared runtime defaults for an illustrative shot. After recording, cancel unwanted training jobs, restore the agreed draft access/instruction state, leave execution fixtures paused and retain their audit history. The reviewed autonomous examples passed; repeat only the explicitly selected fictional exercise when capturing new motion.