Structured · consent-gated
Labs Concierge
An AI front desk that would rather refuse than guess.
Live on both sites as the koi's front office. It routes real conversations; none has become a paid engagement yet
Verified 2026-07-26 · 3 sources
- The front office leads the studio's concierge page
- SSR fetch of koinophobialabs.com/concierge renders the concierge-page hero, checked 2026-07-26
- Shipped to production through reviewed PRs
- koinophobia-labs-site PRs #35–#39, merged 2026-07-20/21 with production deploys verified from those exact SHAs
- The ease-of-use audit and its fixes are on the record
- 2026-07-21 audit scored the first release 63/100 across ten visitor journeys; all nine defects closed in PR #39 (docs/FRONT_OFFICE.md)
The human problem
Every business front door forces a bad choice: a dead form that flattens a messy situation into dropdown fields, or a chat widget that cheerfully improvises answers it has no right to give. My studio needed a front door too, and I wasn't willing to ship either.
The thesis
An assistant at the front desk should extract, clarify, and route — never invent. Deterministic understanding first, one question at a time, and nothing leaves the page until the visitor says send.
Where it actually stands
- Runs on both domains as the koi companion's front office: a messy first message becomes clarifying questions, then a structured brief the visitor can edit, then an honest recommendation.
- On the studio site it fills the exact same intake pipeline as the form — no shadow schema, no second lead system. On this site it collects no contact information at all, and hands hire-intent to the studio with the context carried over.
- It refuses honestly. A budget that doesn't fit is told a smaller or outside solution is the better call, and a question the site's own data can't answer gets a clarification instead of a guess.
- Days after shipping it, I audited it like a stranger and scored it 63 out of 100 — the front door was hidden behind an unlabeled fish. Nine fixes later, every existing help CTA opens it directly and the koi wears a label.
Decisions and tradeoffs
Deterministic extraction before any conversation.
A front desk that misremembers your budget is worse than a form. The understanding layer is code I can test, not vibes I can prompt.
Zero network calls before consent.
Nothing is created, scored, or sent until the visitor reviews the brief and says so. An abandoned conversation leaves no trace — which is the point.
Typing first, suggestion chips second.
The audit caught the free-text box buried under seven chips, two hundred pixels below the fold on a small phone. People think in sentences; the machine adapts, not the person.
What building it taught me
Conversion surfaces rot faster than any other code. The release that felt finished scored 63 out of 100 once I walked through it as six different strangers — and every point it gained back came from fixes a visitor would actually feel, not from new features.
What isn't true yet
- No lead that arrived through it has become a paying engagement. It routes conversations; it hasn't closed one.
- The five-human benchmark hasn't run — every score so far is my own adversarial walkthrough, and I already know how that can fool me.
- It only answers from what the sites already publish. Ask it something the pages don't know and it tells you so.