01
AI Employees
Role-based multi-agentic workers coordinate through a four-loop harness: perceive, reason, act, learn.
Under the hood, an Auto is more than a chatbot with a few tools attached. It is a small self-driving operation.
Digital workers reason over a living map of the business, stay inside firm rules, run on a system built so it never loses its place, and record every step. This page opens the hood, in plain terms first, with the technical terms beside them.
A request comes in. The Auto looks it up in its map of the business, works it out using the cheapest capable AI for each step, checks it against your rules before doing anything that matters, carries it out on a system that never loses its place, and writes down every step. People handle only the exceptions. Here is that path.
Role-based multi-agentic workers coordinate through a four-loop harness: perceive, reason, act, learn.
A compact classifier routes each step to the cheapest capable model.
Authority, limits, and compliance are evaluated before consequential actions run.
Entities, relationships, policies, memory, and execution traces ground the reasoning.
Long-running work executes durably and resumes without losing state.
Every step is logged and replayable; exceptions surface with evidence and policy context.
Retrieval-augmented chatbots keep a cache of documents. An Auto keeps a graph of the enterprise. Entities, their relationships, the policies that bind them, the memory of prior decisions, and every execution trace live in one per-tenant substrate.
Every time an AI “thinks,” it costs money, measured in tokens. Many AI systems send every step to the largest, most expensive model, so the bill grows with every step in a process. That is the pattern Supervity calls tokenmaxing.
An Auto is smarter about it: most steps are simple sorting and extraction that a small, specialised model handles cheaply, and the large, expensive model is saved for the genuinely hard thinking. The result is roughly 60% fewer tokens per process.
A compact model classifies each step and routes it. Cheap, fast, and the reason the expensive models are rarely called.
A domain model carries the bulk of the work, specialised per tenant with lightweight adapters. The adapters are disposable, regenerable caches; enterprise knowledge lives in the Auto Graph.
The customer's own frontier-model contract is called only when the step genuinely needs it, which is where sovereign economics and margins hold.
The mechanism, in one line: deterministic code plus tight graph context plus a small model for the routine, with the large model held in reserve. The same economics let AIshore undercut outsourcing, and you can model it against your process.
Autonomy only works in production when authority is deterministic and execution survives the real world. Auto Policies and Auto Runtime work as complementary planes.
Guardrails written as prompts are probabilistic; they mostly hold. An Auto evaluates authority, limits, segregation of duties, and compliance in a deterministic decision model that runs before any consequential action. The boundary is not a suggestion to the model. It is a gate the action must pass.
Enterprise work is long-running, spans systems, and must survive failure. The Auto Runtime executes durably and resumably, so a process that takes days does not lose state. Agent changes ship through a versioned lifecycle with instant rollback, so improving an Auto never risks the operation.
Because policies and traces live in the Auto Graph next to the entities they govern, audit is not reconstructed after the fact. Every step an AI Employee takes, every policy decision, every human approval, and every model call is time-stamped and queryable. A regulator's question, “why did the system do this”, is answered by replaying the trace, not by interviewing the team.
Autonomous vehicles have SAE levels. Enterprise software needs the same shared language, so a CIO can say exactly how self-driving an operation is. Supervity proposes Software Autonomy Levels as that standard.
People operate the software. Every step is human.
A copilot suggests; the human still does the work and every decision.
Scripted tasks run alone, but break on exceptions and change.
AI Employees run whole processes and hand exceptions to a human. Where most Autos start.
Typical starting pointThe operation runs end to end under policy; humans govern by exception and objective. The Daikin Finance Auto operates here at 85% AI-first.
85% AI-first exampleSelf-optimising operations across functions on the Auto Graph, humans setting direction only.
The AutoPilot Bootcamp stands up a working Auto App on your data in three weeks. Bring your hardest process and your chief architect; leave with a running L3+ Auto and the trace to inspect.
Autonomous · Always Learning · Auditable