Will live practice
work at the branch?

The product

Employees practice fictional member conversations and use coaching roadmaps. Managers guide practice. FI Strategies uses aggregate patterns to improve training. The immediate technical risk is two-way live voice. The $39,000 Full Program also adds coaching/business-development tracks, a private Field Coach, a six-month Momentum Path and branch-level Results Board. Existing FI refresher video links are included; building a video platform, producing videos and cloning Paul’s voice are excluded.

Every Full Program component

Open the complete product walkthrough →

The demo shows the intended workflows. The network pilot exercises shared transports; it does not claim that each production feature has been built.

Two separate network paths

EMPLOYEE BROWSER
  │
  ├── HTTPS 443 → FI test site / Vercel
  │                ├── Pages, sample roadmap, CSV upload/return, PDF/CSV and media
  │                └── Session setup + text replies → OpenAI API
  │
  └── WebRTC audio + data channel ↔ OpenAI media service

The API key stays on the server. Live audio travels directly between the browser and OpenAI. The custom domain does not make all traffic first-party. No TURN relay is currently configured.

Run these checks in order

1. Use a normal teller/personal banker account on the branch network, with IT permission.
2. Run HTTPS roadmap, upload/return and cookie checks.
3. Run data/report, server AI and sample-video checks. Download and open the CSV and PDF. Then check headset playback and microphone permissions.
4. Run live voice for three minutes. Complete three conversational turns, interrupt once, then stop and reconnect.
5. If helpful, run the five-second HTTPS audio echo. It tests transport only, not an AI fallback.
6. Download the JSON and correlate timestamps with the institution's proxy/firewall logs.

Repeat for VPN/VDI and other branches if used. A CEO laptop or home-network success is not sufficient.

When it fails

Page blocked: check DNS, URL categorization and hosting policy.
API blocked: check POST/body inspection and TLS proxy behavior.
No microphone: check managed browser, OS and VDI permissions.
Session created, no audio: examine ICE/media, UDP restrictions and proxy-only routing.
Drops or poor sound: inspect timeouts, packet loss, jitter, congestion and headset selection.

A 401 is an access-code issue. A 429 or 5xx may be a provider/configuration problem. Do not label every error a firewall block.

Options to evaluate

1. Narrow IT-approved access for the actual site and media service.
2. Authenticated TURN/TLS on TCP 443, proven with a relay candidate and real audio. Not implemented here; TLS TURN can still be blocked by an HTTPS-only proxy.
3. Same-origin WSS audio relay, with duration/reconnect handling. Not implemented here.
4. Push-to-talk over HTTPS with transcription, AI reply and speech playback. The echo test checks upload/download only; those AI stages are not implemented.

Do not disable security controls or use a hotspot as proof the corporate network passes. Obtain API-specific destination requirements; ChatGPT app allowlists are not an API specification.

What to send back

A simple matrix for each institution: device/account/network tested, what passed, what failed, evidence and timestamps, proposed approved change, retest result, and recommended voice transport. Separate technical compatibility from the institution's security/vendor approval.

Not tested: production login/SSO, database persistence, tenant isolation, concurrent load, production scoring or compliance. No real member data. The product walkthrough uses fictional data and scripted text scoring.

Consultant’s production checklist

• Login and roles: choose the real identity provider; test employee, coach, FI and cross-credit-union access. A demo role switch and cookie probe do not prove isolation.
• Storage and knowledge: choose database/document hosting; test saves across reloads, authorized retrieval, retention and deletion. This demo’s sample state is temporary.
• Call capture: prove FI’s Granola → draft → approval → logbook → pre-call-brief pipeline from the FI side. No live Granola connection exists in this demo.
• Videos: start with delivery from the platform’s own domain, as Alex authorized. Obtain FI’s approved files and test playback, sound, seeking and sustained delivery on the CU workstation. If the final setup uses another host/CDN, test that exact destination too. The sample does not prove external hosts work.
• Capacity: agree expected session duration and concurrency, then test provider limits and sustained voice on representative networks. Two successful sessions are not a load test.
• Final retest: any new hostname, direct-storage URL, login redirect, streaming text or voice relay must be added to the destination list and tested.

Keep an endpoint inventory with hostname, purpose, protocol, browser-versus-server origin, owner and evidence. Do not request blanket security exceptions.

Technical references

OpenAI WebRTCBrowser microphone permissionsManaged Edge audio policyCloudflare TURN transportsVercel WebSocket constraints