Overview

Run the staff panel as a Per-contact Standard application in the Connect agent workspace. It reads the contact ID from the workspace, checks staff access and loads approved contact attributes automatically. Staff choose a task and review the result in the panel.

Both the interface and server remain in your staff host; Voiceflow runs the published read-only assistant. This walkthrough is on demand. It does not listen to calls, ingest a live transcript or send a customer reply automatically.

Prepare the staff deployment

Gather the assistant, staff identity and provider permissions once.

Download and extract the starter bundle. Keep support-platforms, contact-center-assist and amazon-connect as sibling directories. The app reuses your host’s verified staff session; platform registration alone does not create one. Test SSO in the actual embedded browser.

Install the staff entry

Connect the server, generate the platform app and open a sandbox interaction.

1

Connect the provider reader

Install @aws-sdk/client-connect in your existing Node.js 22+ staff server, then use the supplied Connect factory:

Mount the Connect staff routes

requireStaff verifies the host session and tenant. authorizeContact(staff, contactId, { signal }) checks the trusted assignment/case service for this instance and returns the initial contact ID, or null. The loader refuses unauthorized contacts before any AWS read. Do not return the browser ID unchanged without an access check. The server owns the Region, instance and IAM role; the browser receives no AWS credentials.

Mount handleAssistRequest at both same-origin POST paths: /api/agent-assist and /api/agent-assist/context. Apply a 32 KiB body limit, bounded authentication/provider reads and staff rate/concurrency limits. verifyCsrf applies the host’s policy to both routes. Keep Voiceflow and provider credentials on the server.

2

Build and install the platform app

Replace https://staff.example with your staff host’s HTTPS origin, without a path. From the extracted archive root:

Build the Per-contact workspace page

The command creates assist-build/amazon-connect/: panel.html, panel.css, a bundle containing the tested @amazon-connect/app@1.0.15 entry, and SETUP.md. Serve the three panel files together at https://staff.example/support-platforms/amazon-connect/panel.html, behind staff sign-in. No script replacement or separate SDK bundling is needed.

In Connect console → Integrations → Add integration, choose Standard application, enter that HTTPS access URL and select Per contact scope. Associate the pilot instance, grant the application to the pilot security profile and use only approved iframe origins/permissions. Open it from the workspace Apps launcher during an assigned test contact and complete staff sign-in. If the menu is absent, check regional availability and prerequisites.

Use Cache-Control: no-store and a restrictive CSP permitting the actual workspace and SSO ancestors. The setup command refuses to overwrite an existing build directory; append a new output directory after the origin when rebuilding.

3

Check context and choose a task

The saved ticket/contact ID appears automatically. Wait for Case context ready and inspect Approved case context. A denied context check leaves tasks disabled and starts no Voiceflow request. Each later task reloads its snapshot on the server.

The question box is optional for these four actions. Use it to focus a task, or enter a question and choose Ask assistant. No task runs just because a ticket/contact opens. Configure the assistant to honor the selected task’s output format; if a reply contains internal guidance, edit or regenerate it before using it.

Configure case context

Follow the workspace selection while keeping authorization on the server.

The browser entry reads event.context.scope.contactId only when the workspace scope is contact. Idle scope keeps the task controls disabled. Create binds the new contact; destroy clears the ID, approved snapshot and answer, and discards late results. No manual case ID or Finish case step is required.

The backend checks contact entitlement and calls GetContactAttributes with the trusted initial contact ID. Only case_reference, reason_for_contact and product enter the snapshot. These attributes do not include a live transcript or full CRM case. If a CRM record is needed, resolve the reference and recheck that record’s staff/tenant access on the server before adding approved fields. A case reference alone is not authorization.

Connect results remain in the staff panel. Copy suggestion copies completed text for review in the normal workspace; no Connect composer insertion, after-contact save, customer send or call control is supplied. The Summarize task covers only the available attributes unless your authorized loader supplies an interaction excerpt.

Changing the case or task focus clears the previous result. Every task starts a fresh Voiceflow conversation; follow-up facts must come from the refreshed snapshot or current question. The host must unload/reset the panel on logout and account changes. Aborting browser consumption does not prove remote work stopped or transcripts were deleted; set expiry and retention with the owning teams.

Test the staff workflow

Prove automatic context, human review and recovery before expanding the pilot.

1

Verify the automatic context and all four tasks

Open an allowed sandbox interaction without typing an ID. Use an unopened-item case delivered 12 days ago with an approved 30-day policy. Confirm the visible ID and approved snapshot, then run each task with an empty question. Check facts, policy sources and explicit missing information. Summaries must not claim coverage beyond the loaded snapshot.

2

Test the platform lifecycle and draft action

Launch the app while idle: task controls must remain disabled. Open it for a contact, start a task and end that contact while the response is streaming. The answer and context must clear and late text must not reappear. Repeat with a second contact and confirm no first-contact facts carry over.

3

Prove denial and recover from failure

Use the second staff account for a record it cannot access: no context or answer may appear and no runtime request may start. Test an expired session, denied provider read, interrupted stream and unavailable assistant. Partial output must disappear and remain unavailable for copying or insertion. Restore the service and retry manually; normal support must remain usable.

The panel or context does not load

Check installation, embedded sign-in and the approved provider reader.

Confirm the generated URL, app assignment, exact iframe ancestors and staff session. For 401, sign in again. For 403, check CSRF, tenant mapping and record access without weakening them. For 413, narrow the approved snapshot. For 502, inspect a redacted server correlation record for provider/runtime failures. Retry context by reopening/reloading the app after recovery; do not expose the panel publicly to bypass SSO.

The suggestion is incomplete or uses the wrong case

Check the loaded snapshot and current workspace instance before using the result.

Compare the visible ID and approved fields with the actual interaction. Reopen the app to obtain a fresh binding. Check the server’s account/instance configuration and published Voiceflow Environment. Test the switch during a response again; a successful network abort alone does not prove late results are discarded. Review model output against sources before using it.

Release and maintain assistance

Record the accepted build, permissions and rollback before broadening access.

1

Release to the pilot group

Release first to the pilot group. Record the staff-host build, field allowlist, platform registration, tested staff identities and Voiceflow Environment/release. Repeat the allowed/denied interaction check through that released entry. Assign owners for policy freshness, identity/token lifecycle, app/SDK updates, retention and incidents; rerun quality evaluation after changes.

2

Restore the reviewed deployment

Remove the security-profile grant or disassociate the integration before deleting it. Restore the recorded host build and reviewed Voiceflow release, then repeat a fresh-session check.

Local review can use node support-platforms/fixture-server.mjs and http://127.0.0.1:4397/fixture-sign-in?platform=amazon-connect. These clearly labelled fictional SDK/model fixtures exercise binding, task buttons and cleanup; they do not establish live installation, provider permissions, SSO or answer quality. Complete those checks in the actual sandbox before release.

For business-system tools, Import help articles and Human agent hand-off, use the canonical Integration manuals.

Use the up and down arrow keys to select a result, Enter to open it, and Escape to close the search.