How I Modernize Outdated Enterprise Software With AI

What the video demonstrates (enterprise modernization with AI)

The creator shows how to replace one legacy workflow with an AI-driven agent system that can:

1. Accept a phone number + task/appointment description via a web portal
2. Have an AI voice agent place an outbound call, carry out a scripted confirmation conversation, and end the call
3. Run post-call evaluation and update the system of record so the portal reflects results

---

Tools/products used (3 no-code/low-code components)

• Claude Code (via Cursor): used to generate/modify the portal UI/code and backend integration logic.
• Retail AI: used to build/configure the voice phone-calling agent, including prompts, variables, phone numbers, and webhook events.
• n8n (mentioned as “N”): used for automation after the call (webhook handling, transcript processing, database updates).

---

Live example workflow (end-to-end)

1) User input via the portal

In the portal, the user submits:

• Phone number
• Short description of what the agent should confirm (e.g., appointment details)

2) Automated outbound call

After submission:

• The AI system places the outbound call
• The customer conversation follows the prompt (greet by name, confirm correct person, state appointment time, ask confirmation, then end the call)
• The call log is updated and later marked complete with a confirmed/not confirmed outcome

---

Portal + backend design (Supabase as shared memory/system)

Supabase is used for:

• Authentication for portal access (login/dashboard template)
• A shared calls table to store call metadata and outcomes
• A Storage bucket to hold audio/recording assets (kept private)

Portal implementation notes

• The creator starts from an app template (login/dashboard/logout) and wires it to Supabase auth + storage.
• The portal page includes a form (phone number + appointment description) that:
  • Inserts a row into a Supabase table (status begins as pending/calling)
  • Triggers the outbound call (later, via Retail AI API from the portal logic)

---

Data model / storage implemented

• A Supabase SQL migration/table is generated (by Claude Code) with fields such as:
  • timestamps (created/updated_at)
  • user_id
  • phone_number
  • appointment details / confirmation status
  • call status (e.g., calling → complete)
  • transcript
  • recording URL (audio link)
  • additional metadata used for filtering/UI display
• A Supabase storage bucket is created (named for call recordings) with policies applied (private by default).

---

Retail AI voice agent setup (prompt + variables + phone number)

• The Retail AI agent is created as a single prompt / simple prompt agent (recommended for easier debugging early on).
• The prompt instructs the agent to:
  • Wait for greeting
  • Greet by name
  • Confirm correct person and appointment details
  • Ask for confirmation
  • Use an end-call function to close the call
• The prompt uses custom variables (e.g., appointment details) passed from the portal.
• The creator configures:
  • An outbound phone number (option to buy one for demos; recommended to match area code)
  • An outbound call agent deployment that is published/activated

---

API-triggered call from the portal

• After the portal inserts the task into Supabase, Claude Code adds logic to call the Retail AI “create phone call” API.
• Required parameters include:
  • from number
  • to number (customer phone)
  • agent ID
  • variables such as appointment details (must match the variable name expected by the Retail AI prompt)
• Environment secrets are configured in env.local:
  • Supabase URL + publishable key + service role key
  • Retail AI API key

---

Post-call automation with n8n (webhook → transcript analysis → DB update)

• Retail AI is configured with a post-call webhook pointing to an n8n workflow.
• In n8n:
  • The workflow captures:
    • called number / customer number (to locate the correct Supabase row)
    • transcript
    • (optionally recording URL, though the demo focuses on transcript to move faster)
  • An LLM step analyzes the transcript to extract the outcome:
    • outputs “confirmed” if the user will attend
    • outputs “not confirmed” if they need to reschedule
  • The workflow then updates the relevant Supabase record:
    • changes call status from calling → complete
    • sets confirmation status (e.g., confirmed)

The portal UI refreshes to show updated complete/confirmed tags.

---

Key implementation notes / guidance

• Keep the voice prompt simple at first; test with ~10–30 calls before optimizing.
• Webhook/event stages may be emitted multiple times during a call; production workflows should filter events so only the final post-call event updates the database.
• Suggested iterative rollout:
  1. Run 10–20 calls
  2. Iterate on portal + Supabase workflow
  3. Refine the Retail AI prompt
  4. Improve n8n post-call analysis and metadata filling
• Mentions offering resources via a community/school if there’s interest.

---

Main speakers/sources

• Primary speaker: The video creator (“Hello legends…” narrator)
• Systems / sources referenced: Claude Code (Cursor), Retail AI, Supabase, n8n (N)