How to build a client portal with an AI practice coach inside it
A portal each client logs into to see where they are in your process, pull their resources, and roleplay the hard conversations out loud with an AI that grades them.
Giving every client their own software means a portal with their place in your process, their resources, and an AI voice coach that roleplays hard conversations and grades them. Sageridge built a working demo with Codex and deployed it to Vercel in about forty minutes of agent time.
What giving every client their own software means
The idea is simple. Instead of a folder of links and a calendar invite, each client gets a place to log in: their own portal, with their own access code. Inside it is where they sit in your process, what comes next, the resources that go with this stage, and somewhere to upload documents.
Sageridge runs this internally. A new client gets a portal when they sign. They can see exactly where they are, upload their documents, sign the agreement, pay the invoice, and read their proposal, all in one place. Once the build starts, the same portal shows progress.
None of that is new as a concept. Client portals have existed for twenty years. What changed is the cost of building one that is specific to your process rather than generic to your industry.
The part clients actually use is the practice room
The demo portal has a roadmap, a progress view, and a resources shelf. Useful, but not the reason someone logs in twice.
The reason is a live voice roleplay. The client enters an access code, starts a practice session, and talks to a conversational AI playing the other side of a hard conversation. In the demo the scenario was a manager and an employee who does not want to sell: the employee thinks the pitch is inauthentic and will cost them the client's trust, and the manager has to get somewhere without falling back on "that's just the way it is around here."
That is a conversation real managers handle badly, and reading about it does not help. Saying it out loud, badly, to software that will not hold it against you socially, does.
Everyone wants things personalized to them and their specific problems.
It grades the conversation
When the session ends, it scores the attempt. In the demo run, curiosity came back low, empathy 66, commitment 48. The written feedback named one thing that worked, which was that the manager created space, and one thing to do differently: finish by naming the observable behavior you want next time rather than the attitude.
A full transcript sits alongside the score, so the client can see exactly where the conversation went sideways.
The numbers matter less than the fact that there are numbers. A client who practices once and gets a 48 has a reason to practice again. A client who is handed your training PDF has no feedback loop at all, which is why the PDF does not get opened twice.
How it got built
This was built with Codex, because the live voice piece uses OpenAI's conversational voice model. The portal itself could be built with any capable coding agent; the voice roleplay is what tied this particular build to OpenAI.
The prompt was not elaborate. Roughly: build a demo portal I can send to a client so they can reach all our resources, put a live roleplay voice coach in it, give it a login and authentication, and deploy it so anyone can open it at a normal URL.
Two practical notes from the build:
- Pick the mid-tier model at a high reasoning setting before reaching for the top of the range. It costs meaningfully less and finished this job fine.
- Name your deploy target up front, or the agent will stop and ask you which one you want.
The deploy conflict is the useful part
The first build finished in about eighteen minutes and passed its own production build check across the landing page, dashboard, and voice room. Then publishing stopped on a rule conflict: the work had started inside the company repo, and a throwaway demo had no business deploying from there.
The fix was a sentence, spoken rather than typed. This is just a test, deploy it to a side branch, nothing goes to main, I just need it reachable at a normal URL. Another eighteen minutes and it was live behind a demo sign-in.
That is worth sitting with, because it is the common case. The build worked. The deploy hit a guardrail that existed for a good reason. Describing the outcome you actually wanted was faster than debugging it.
AI is your thought partner, not your thought leader. You are the leader.
Where this fits in a service business
If you run a coaching, consulting, wealth, or any process-driven service business, you already have the raw material: a roadmap clients follow, trainings you have already recorded, and the same four conversations you walk clients through every month.
The portal moves that off your calendar and into their software. Clients get more access to your thinking without more of your hours, and they stop waiting on you to make progress.
The honest limit: the backend has to be set up correctly and secured, and the voice coach needs an API key wired in before it does anything at all. What the video shows is a demo of the thing, not a finished product for paying clients. But the gap between deciding to build a client portal and having a working one is now measured in an afternoon.
The short version
- A client portal is no longer a software project. This one went from plain-language prompt to live URL in about forty minutes of agent time.
- The roadmap and resources get clients oriented. The voice roleplay and its score are what bring them back.
- Try the mid-tier model at a high reasoning setting before the top of the range. It finished this build for less.
- Name the deploy target and the branch up front. The only failure in this build was a deploy rule conflict, not the code.
Questions this raises
What is actually in a client portal like this?
A login per client, a roadmap showing where they are in your process and what comes next, the resources for their current stage, document upload, and a live voice roleplay that scores the attempt. Sageridge's own client portals also carry the agreement, the invoice, and the proposal.
Do you need to be able to code to build one?
No. The build in the video was a plain-language prompt to a coding agent, and the one problem that came up was solved by talking through the outcome rather than editing code. You do need somewhere to deploy it and an API key for the voice model.
How long does it take?
This demo took about eighteen minutes to build and about eighteen more to clear a deploy conflict and go live, so call it under an hour. A production portal for real clients needs the backend and the security done properly, which is more than an afternoon.