Umbral

Inspiration

AI can automate a decision, but a company’s real operating judgment lives in the corrections made by experienced people. Today, those corrections usually disappear inside chats, emails, and meetings. Umbral starts at that exact moment and turns the correction into reusable company knowledge.

What it does

Umbral receives the original decision context and a free-form human correction. GPT-5.6 Sol transforms them into a structured operating criterion with four elements: scope, required data, exceptions, and escalation conditions.

A person reviews and validates the criterion before it can affect future decisions. When a new case appears, Umbral evaluates it against the validated criterion and returns either EXECUTE or ESCALATE, explaining the reason and identifying any missing information.

The prototype uses a fictional purchasing decision. It demonstrates the decision mechanism but does not connect to company systems or execute real purchases.

How we built it

Codex built the complete prototype: the interface, server-only API routes, structured schemas, tests, and deployment on OpenAI Sites.

The product uses two live GPT-5.6 Sol calls through the OpenAI Responses API. The first converts a correction into a strict criterion card. The second evaluates a new purchase against that validated criterion. Strict Structured Outputs keep both responses reviewable and predictable.

Challenges

The main challenge was avoiding two opposite risks: an AI that generalizes too freely and a governance system that creates more bureaucracy than value.

We reduced human effort to reviewing a criterion already drafted by the model. We limited AI autonomy through explicit scope, required evidence, boundary testing, and mandatory escalation whenever information is missing, stale, conflicting, or outside the validated criterion.

Accomplishments

We built a working end-to-end product with real model responses. One expert correction becomes a traceable operating criterion, and that criterion produces two visibly different outcomes: EXECUTE for a known case and ESCALATE for an incomplete or exceptional case.

What we learned

The most valuable company knowledge often appears during correction, not while writing a procedure manual. AI should not receive autonomy all at once. It should earn autonomy inside boundaries that people have explicitly taught, reviewed, and tested.

What’s next for Umbral

The next step is to test Umbral with real administrative processes such as purchasing approvals, supplier onboarding, invoice exceptions, and recurring operational reviews. Future versions would add criterion history, organizational roles, system integrations, usage limits, and metrics showing how many decisions safely moved from human review to controlled autonomy.

Built With

• codex
• gpt-5.6-sol
• openai-responses-api
• openai-sites
• structured-outputs
• typescript