Philosophy
What a model can do, and what stays with a person.
Organizations already know how to do most of their work. The open question is which parts can be written down, which parts can be handed to a machine, and which parts must remain answerable to a human. These are the lines Functionary is built around.
Not all judgment is delegable to AI
Accountability is not transferable. When a consequential decision turns out to be wrong, someone is answerable for it, and that responsibility cannot be handed to a model. The model produces an output; it does not absorb the consequence.
Functionary routes the non-delegable judgments to a person rather than pretending they were automated. A flow does not quietly fold a decision that someone will have to answer for into a step labelled as machine work.
Process and discretion are both real
Institutions run on standardized process and on delegated discretion. The standard part — the repeatable, specifiable part — can be formalized. The discretion part runs on human trust networks: reputations, relationships, the knowledge of who is accountable to whom.
An AI cannot join those networks. A tool that ignores them models only half of how an organization actually works, and the missing half is the half that decides what happens when the standard case does not apply.
Trust is checking placed on purpose
Being able to check an output after the fact does not, by itself, make a process trustworthy. Trust also requires knowing the process is stable, owned, and the same tomorrow as it was today.
Each step does a narrow job with only the information it needs, which is what makes checking it quick rather than a second full reading. Trust need not rest on any one step; it rests on the whole, laid out where you can read it.
The point of automation is not to check less for its own sake, but to check where checking earns its keep. A system you must audit output by output has not automated the work — it has moved the labor from doing to reviewing. A system you never check is not trustworthy either; it is just unexamined. The goal is to put checking where it belongs and remove it where it no longer does.
A visible flow lets you place checking deliberately. You pause where a judgment is yours to make, inspect a wrong result in the context of the step that produced it, and monitor nondeterministic work in proportion to its risk. Captured agent runs matter here: they show what happened, expose repeated choices, and provide evidence for what to formalize next. They are an on-ramp to a durable process, not a substitute for one.
Data control begins with explicit boundaries
Local files and an operator-selected data root give the process a home you control. They do not, by themselves, make every execution local. Model providers and connectors are egress choices, and the interface should make those choices visible before information crosses a boundary.
Local models can keep inference on-device. Remote providers receive the inputs sent to them. Configured PII checks can block a step, and output policies can redact or hash matching fields, but automatic pre-model redaction is not a promise the current system makes. Honest control starts by naming the boundary precisely.
Errors should fix the process, not patch the case
Organizations that function well do not blame individuals. They fix the process so the same mistake is less likely and easier to catch. The repair outlives the incident.
When the process is a declared flow, the fix is a visible, testable, versioned change — not an unversioned prompt edit with no shared record. The model does not have to get better for the work to. The process does — and the process is the part you own.
Ad-hoc AI can become organizational process
Pasting content into a private chat encodes an individual interaction, not a shared process. But freestyle itself is not the problem. The problem is work that is never captured, compared, or turned into an artifact the organization can improve.
Functionary captures that first discretionary attempt inside a project. Crossing from repeated freestyle to a stated process the organization can read, Test, evaluate, and gradually narrow is the gap Functionary exists to close.