Deployments, automation audits and grounded concierges — built on the same platform that runs our own ten companies. Every external, financial or irreversible action waits for a named human approval. Facts first: assistants answer from your knowledge base at zero model tokens, and label everything else.
Site, portal, grounded assistant and CRM deployed from the group factory; KB built from your facts; gate policy configured so money and outbound actions need your keystroke. The group runs on this exact stack.
Your process mapped step by step: what agents take, what stays human (Type-1 gates), minutes saved per month, and the EU AI Act watch (as-of dated) attached. The audit is both the quote and the build spec.
KB-first routing: your facts answered at zero model tokens; everything else labeled by provider; research answers carry cited sources or refuse. Injection-battery tested with Sentinel before launch.
External, financial or irreversible actions are Type-1: staged, evidenced, and executed only on a named human's approval. The system has no code path around this.
Every answer carries its mode: knowledge-base, model (named provider), or template. Offline degrades honestly to templates — never a pretend AI.
Deliverables ship with reasoning traces, purchase tickets (when sibling evidence was used), version hashes and as-of dates. You can re-run and re-check.
| Service | Model | Price |
|---|---|---|
| Automation audit (process → agent design, minutes saved) | one-off, 48 h | CHF 0 |
| Full automation audit (on-site data, gate policy, AI Act watch) | per process | CHF 900 |
| Deployment setup (site + portal + assistant + KB + gates) | one-off | CHF 2,400 |
| Assistant retainer (hosting, KB updates, monitoring) | monthly, cancel monthly | CHF 500 |
| Gateway ops (provider chain, cost caps, diagnostics) | monthly | CHF 150 |
All prices in CHF, excl. VAT where applicable. Swiss QR-bill, net 14 days. Model token spend on your assistant is capped per month at a number you set — the cap is enforced in code.
A named person replies within 1 business day; your audit returns within 48 hours of the process description.
Your process mapped step by step with three columns: agent-takes / human-gate / reason, plus estimated minutes saved per month and the gate policy we'd configure. It is the same document we build from if you say yes — no second discovery call.
Internally and reversibly, yes (drafts, computations, research). Externally, financially or irreversibly: never without a named human's approval keystroke. That is architecture, not policy — there is no code path around the gate.
KB-first grounding (your facts, zero-token), provider labels on every other answer, injection-battery testing by CogOps Sentinel before launch, monthly token caps you set, and cited sources or refusal for research questions.
Ten CogOps companies run on it daily — five of them in regulated domains, linked in a complete graph of internal evidence trades. Your deployment is the same machine, one tenant down. Reference deployments are linked from every page footer.