The customer portal is the most-built business app there is, and the new tools change what 'finished' means. A portal used to be logins, scoped pages, and a few tables. Now your customers expect to ask their data questions and have the app do work for them. Here is how to build that version — secure, scoped, and AI-powered — in an afternoon.
What we are building
A portal where each customer logs in and sees only their own data: projects, invoices, and files. On top of that, two AI capabilities your customers actually feel — an Ask box so they can query their account in plain English, and an agent that quietly handles the repetitive follow-ups so your team does not have to.
Step 1 — The data and the rules
Start with the backend, because the rules are the product. Four tables: Customers (your login table), Projects, Invoices, and Files, connected so every record traces back to a customer. Then the one rule that makes a portal a portal: row-level security scoping every record to the logged-in customer.
Step 2 — Describe the portal
Now let the AI build the frontend. Describe what you want in plain English and watch it assemble against your tables:
Build a customer portal home page. Greet the logged-in customer by name, show three cards (open invoices, active projects, files this month), and a table of their five most recent invoices with status pills and a View PDF button.Because the backend already enforces row-level security, every screen the agent generates is private by construction. You are describing layout, not re-implementing access.
Step 3 — Turn on Ask for your customers
Enable Ask on the portal interface and your customers can now query their own account conversationally — 'what did I pay last quarter?', 'which projects are still open?' — and get a table, chart, or number back. Each query runs under that customer's permissions, so one customer can never surface another's data.
Step 4 — An agent that does the busywork
Finally, add an agent for the repetitive work. Point it at a job — 'when an invoice goes 7 days overdue, draft a polite reminder to the customer and flag it for a team member to send.' The agent reads, drafts, and pauses at the approval gate so a human okays the send. Every action is logged.
That is the difference between a portal that displays data and one that works on your behalf — and it is a few sentences, not a sprint.