Flow

Flow

Solutions

What real Human-in-the-Loop means for enterprise AI agents

Real human-in-the-loop means a person can stop or change the action before a system of record or a client communication is changed. A rubber stamp is an approval after the fact, or an approval of a bundle you cannot inspect. If the record has already moved, the person was not in the loop.

What approve has to show

The proposed actions, in language a reviewer can read. The destination systems. What will be written. Where the design allows a reverse, a way to undo that write. An opaque "approve all" on a hidden plan is a rubber stamp with a cleaner button.

The Flow pattern

Watch the work. When a close is clear, surface one Do It. Run the step only after that approval. The model is for intent. Execution is deterministic, not a spray of guessed clicks. If the signal is uncertain, the right behavior is silence, not a nudge to approve something half-seen. The same idea, for a personal Gmail close, lives in Glance. Glance is a separate product. It is not this enterprise loop and not a way to pilot it.

EU AI Act, lightly

High-risk contexts in the EU AI Act discuss human oversight. What that requires of a given system is for counsel. This page describes a product pattern: the person sees the write and approves it before it runs. A screen is not, by itself, compliance with the Act.

On the workflow, not in a side chat

Oversight that gets used sits on the work the person is already doing. Opening a separate chat and writing a prompt is a different shape. Flow's shape is: the work is visible, the proposed write is visible, the person approves, then the step runs. See also why that is not the same as RPA, and where the processing is allowed to run.

Roadmap, not live. SSO, an immutable audit log, Action Graph, Malpractice Shield, and Flow-Edge are on the roadmap or offered per engagement. They are not a live runtime today. This page is product framing, not legal advice.

Common questions

Does the EU AI Act require human oversight?

High-risk contexts discuss human oversight. Whether a given system is in that category, and what evidence a regulator expects, is a question for counsel. The product design on this page is narrower: a person sees the write and approves it before it runs. This is not legal advice.

How do you prevent rubber stamps?

Show the write. Name the destination. Approve before execute. If the signal is uncertain, do not push an approval. A checkbox under a record that has already changed is not human-in-the-loop.

Is there a full audit trail today?

No. An immutable audit log, and anything described as a Malpractice Shield, is on the roadmap or offered per engagement. This page is not a shipped audit product.

Related

Scope a design-partner deployment

The homepage waitlist asks for a work email and comes back with a deployment plan. Same capture the rest of the site already uses.

Get my deployment plan

If you already know the environment, use the contact form.