Nox

All work GRADOS · Santa Fe, Argentina · Live in production

A support agent that knows when to stop.

Anyone can ship an agent that answers. The engineering is in the refusal.

Escalates human On low confidence, personal data, complaints and action requests
The problem

The domain where a confident wrong answer causes real damage

GRADOS supports an organisation covering around 82 schools. The same procedural questions arrive constantly, in Spanish, about systems and statutory filings with real legal consequences if answered wrong.

A support agent that answers those questions is easy to build and dangerous to deploy. Payroll and statutory filing is precisely the domain where a confident wrong answer costs someone money or puts them out of compliance.

The requirement wasn't an agent that answers everything. It was an agent that answers what it should and refuses what it shouldn't.

What I built

Grounded, classified, and designed to escalate

  • Grounded

    Answers come from a curated knowledge base of internal procedural documentation, not from model recall. Sources are returned with every response.

  • Classified

    Each question is classified before it is answered. The classification, the confidence and the outcome are all logged.

  • Escalating by design

    Hands off to a person on low confidence, personal data, complaints, or requests for an action rather than an answer. Certain trigger phrases escalate every time, regardless of what the model's confidence says.

  • Aware of context

    Business hours, holidays and out-of-hours fallbacks are handled explicitly, so nobody gets an instant answer at 2am implying a human is standing by.

  • Never talks over a human

    Once a conversation is handed to a person, the agent stops. It does not auto-reply into a live handoff.

QUESTION CLASSIFY + RETRIEVE TRIGGER? personal data complaint / action CONFIDENT? sources found HUMAN HANDOFF ANSWER triggers escalate regardless of confidence
The refusal path is the design
The part most chatbots skip

The handoff rules aren't a fallback. They're a boundary.

They are not there for when the model fails. They're a designed boundary around the class of questions that should never be automated at all, no matter how confident the model is.

Personal data requests. Complaints. Anything asking the system to do something rather than explain something. Those escalate on category, not on confidence.

Continuous improvement

It learns on a schedule

The agent runs a recurring review of its own work. On each cycle it reads back the period's conversations and reports on itself.

  • How many conversations it handled, and how many it escalated

  • Which answers ran below the confidence threshold

  • Which questions retrieved no usable source at all — the clearest signal of a genuine coverage gap

  • Performance by question classification, and coverage by source document

It also detects corrections. When an operator edits a draft before it goes out, the system flags it two ways: when the proportion of changed text crosses a threshold, or when a critical token changes. A small edit to an important value counts as much as a heavy rewrite, because a changed figure matters more than a reworded sentence.

Then it drafts the knowledge it thinks is missing, as proposed question-and-answer entries aimed at the gaps it found.

And then it stops. Drafts land in a review queue alongside the low-confidence cases. A person approves, edits or rejects each one. Nothing reaches the knowledge base without human sign-off.

Dated reports produced28
Period coveredMay – Aug 2026
Applied without review0
Why the constraint matters

A system that quietly rewrites its own knowledge is a system nobody can audit. This one improves continuously and still leaves a person in charge of what it is allowed to know. The agent proposes. It does not edit itself.

Reach

WhatsApp, properly

Deployed on WhatsApp and as an embeddable web widget. For Spanish-speaking users, WhatsApp is not an alternative channel. It's the primary one.

The integration is a proper Meta Graph API implementation with signature verification, and it carries the same handoff rules as every other channel.

ChannelsWhatsApp + web
LanguageSpanish, native
Loggingevery conversation
Transferable

Any support function where being wrong is expensive

Regulated industries, financial services, healthcare administration, compliance, HR. The differentiator is not answer quality. It's the documented, auditable boundary around what the system is permitted to answer at all.

Need an agent that refuses well?

Most of the value is in what it declines to answer. Let's talk about where your boundary sits.

Start a diagnostic More work