RPA is a scripted bot on a predictable screen or API. An AI workflow reads intent in messy documents and events, then runs a deterministic step. In regulated work, neither should write to a system of record until a person has approved that write.
RPA
A recorded path: this control, this selector, this field. Fast when the screen never moves. Fragile when it does. Little judgment on an unstructured email or a PDF that is not the form the bot was taught.
AI workflow
Intent and judgment on the messy input, then one determined execution step. Still not a license to write. The sensitive step waits for a person.
Why RPA breaks
RPA stores the path, not the point. When a button moves, the selector misses and the bot stops, or it clicks the wrong control. Unstructured mail and PDFs fail the same way: the bot has no reading of what the document is asking, only a map of where it expected the pixels to be.
Where an AI workflow fits
The fit is friction across files, core systems, and the apps a team already uses. Watch the work, propose one execution, wait. Email can be one of those surfaces. It is not the product.
Approval before the write
Intent can be model-assisted. The write should not be a series of guessed clicks. Execution is a determined step: an API call, or an action on the accessibility tree, after a person approves the sensitive ones. The approval shows the proposed action and the destination. It is not a checkbox after the record has already changed. Where a step is designed to be reversible, the write can be undone. That reverse belongs to the step. It is not a claim that every system of record supports undo.
Hybrid, not a replacement
Keep RPA where the path is stable and the volume is high. Use a Flow-class layer where the input is a document or an event that needs judgment, and the close crosses more than one app. You do not retire every bot to add approval before a write.
Glance is a different product
Glance closes a clear personal intent from Gmail with one Do It, and it stays quiet when it is not sure. It is not a trial of Flow, and it is not how a firm samples this deployment.
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
Do we need to replace all RPA bots with AI agents?
No. Keep RPA on stable, high-volume paths. Add an AI workflow where the input is unstructured and a person must approve the write. The practical pattern is hybrid.
Why do RPA bots break when the UI changes?
They depend on selectors and recorded clicks. A layout change breaks the path. An AI workflow is aimed at intent, then a deterministic adapter (an API or an accessibility-tree action), so the judgment is not stored as a brittle click script.
Does Flow auto-write to client systems?
No. Sensitive writes wait for a human approval of the specific action and the destination. Flow does not treat a model guess as permission to change a system of record.
Is audit or SSO live today?
No. SSO, an immutable audit log, and Flow-Edge are on the roadmap or offered per engagement. Ask about them as deployment options. Do not treat them as switched on.
Related
- What real human-in-the-loop means (Flow)
- Running AI on your own infrastructure (Flow)
- Glance, a separate personal product: close from Gmail
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 planIf you already know the environment, use the contact form.