Can the portal assistant approve on my behalf?

No. Approval is your click, on the card or in "Needs you". The assistant can only explain why it might be worth it.

Does my data reach an external AI model?

Answers are generated from your portal state; models run on Cloudflare infrastructure (Workers AI), with an optional alternative provider. They are not used for training and are not kept outside the portal.

How to use Florentin AI and Ask Valhalla

The situation

There are two assistants, with two different jobs. On florentinpurcea.com, Florentin AI is the front door: it answers about Universal, capabilities, pricing, and leads you to a first session. In the portal, the Valhalla button (or ⌘K) is the assistant that knows your state: queue, sessions, health, results. Both share the same golden rule: they do not invent.

Florentin AI, on the site

It opens as a panel on the right (not a "bubble" in the corner). It answers from approved knowledge — public facts about the product: €2,000/month, 4 sessions × 10 hours, no hourly billing, what is included and what passes through at cost, what month one looks like.

What it does well:

explains the product and the process without marketing;
runs a short discovery ("what slows you down?") and proposes the recommended first session, with what it would ship;
hands the conversation off: "Send this to Florentin" opens a short form; the whole conversation reaches him and he replies personally, usually within a day. Or directly on WhatsApp.

What it does when it does not know: it says so. Verbatim: "I don't have that confirmed, so I won't invent it. Tell me in two sentences what you want to achieve and I'll propose the first session — or talk to Florentin directly." No made-up figures, no promised outcomes.

Valhalla, in the portal

You open it with the Valhalla button in the bottom bar, with ⌘K, with "Ask Valhalla or request something…" in the menu, or in context: Ask Valhalla on any recommendation, Ask about this session on any session.

Under the title it says exactly what it is: "Answers from your real state. Executes nothing — any "do X" becomes a request in your plan."

What it knows: the plan and remaining capacity, the active session and its stage, what awaits approval, the approved queue, what you parked, what shipped and with what result, health by area. On a recommendation, it also has the evidence and the priority sentence; on a session, the log.

What it does not do: it shows no scores (it explains "why" in words: impact, urgency, confidence, fit, risk, effort), it invents no data or results, it executes nothing. If your state does not contain the answer, it says it does not have it.

Good questions

The portal suggests four when it opens, and they are not random:

"Why is the first recommendation first?" — you get the factors, not a number.
"What ships in the current session?" — goal, deliverables, stage.
"How many sessions do I have left this month?" — real capacity, with rollover.
"I want to change the pricing page" — that is not a question, it is a request; Valhalla recognises it and proposes a draft with Send request.

Other questions worth asking: "What is the risk if I park this?", "What did the last scan find?", "What was measured on session 3?", "What depends on the fixed tracking?"

How it reaches Florentin

On the site: the "Send this to Florentin" form or WhatsApp — with the conversation attached.
In the portal: any request sent from the assistant reaches him as "New request to scope"; any message in Discussion is his; emergencies alert him immediately by email.