AI agents propose. People decide.
A case created by an AI assistant is born awaiting confirmation: it runs nothing and costs nothing until someone presses "Confirm start".
Included feature that applies to starts from AI agents: the MCP server is available from Optimize, Flow Agent from Automate.
- For
- Operator
- Compliance
- IT
The problem
Connecting an AI assistant to business processes is powerful and risky: an ambiguous instruction can start ten cases, trigger integrations with suppliers and consume AI credit before anyone notices. Blocking agents entirely removes value; leaving them free removes control. The temptation is to pick one of the two extremes, and either way the organization loses: it either gives up the benefits of the assistant or gets them at the price of side effects nobody authorized. A clear middle ground is needed, where the agent does the useful part, preparing and proposing, while the part with consequences, starting the work, stays in the hands of a named person.
Case list
- Filter: Awaiting confirmation (3)
Proposed case
- Proposed by AI assistant · on behalf of Laura B.
- [Confirm start] [Reject]
- Expires in 6 days
How it works
1Enable agent start on the flow
With the "Require confirmation" option, on by default for new flows. The control is therefore present from the start, without anyone having to remember to switch it on.
2The agent creates the case
It is born "awaiting confirmation": it executes no tasks, AI calls, rules or timeouts. Zero cost. The case exists and can be consulted, but nothing the flow provides for has started yet.
3A person decides
Confirm start launches the process. Reject requires a mandatory reason and closes the case without starting it, but leaves it available to consult. Either way the decision is recorded with the name of whoever took it.
4Nothing stays pending forever
A configurable expiry leads to automatic rejection or an alert. A reminder arrives before the deadline, so the request does not get lost in a forgotten queue.
Specifications
| Scope | external assistants via MCP and Flow Agent |
|---|---|
| Default | on for new flows |
| Expiry | from 1 to 2,160 hours (7 days preset), with automatic rejection or alert only |
| Reminder | 24 hours before expiry (at 75% of the time for short windows) |
| Rejection | mandatory reason; case remains viewable |
| Visibility | dedicated filter, badge on the flow with the count, "Start outcome" column in the export |
| Incoming events | deferred and applied on start, not lost |
An operational example: an assistant's proposal
Laura asks an AI assistant to start a case. With the Require confirmation option on the flow, the case does not start: it is born awaiting confirmation. In the case list, the Awaiting confirmation filter shows how many are waiting, three in the example. Opening the proposal, whoever reviews sees it was proposed by an AI assistant on behalf of Laura B., with the initial data filled in. There are two choices: Confirm start or Reject. The proposal also has an expiry, here in six days, after which it does not stay open indefinitely. The diagram beside it is illustrative and not an official screenshot. The point is simple: an agent can prepare the work, but starting a real process requires a person to approve it. The option is available from the edition indicated on the page.
How to verify human oversight
To verify that oversight works, check that the flow really has the confirmation option on and that someone is assigned to look at the Awaiting confirmation filter. A queue of proposals that expire unanswered shows the control exists on paper but not in practice. Each proposal shows the assistant and the person on whose behalf it acts, so it is always clear who originated the request. After confirmation or rejection, the history records the decision and who took it. This helps those who must demonstrate human oversight of agents: declaring that an agent does not act alone is not enough, you need to be able to show the cases where a person approved or rejected. Expired proposals remain a figure to watch, because they tell whether the confirmation workload is sustainable for those responsible.
Related features
- MCP server with 34 tools34 tools with minimum role and annotations, for MCP-compatible clients.
- Built-in Flow AgentA conversational assistant always at hand (⌘/Ctrl+J), always within the user's permissions.
- Agent attributionHistory shows the AI client and the person holding the key; the cost goes to the key.
- Human fallback in the same taskIf automation is uncertain or fails, the same task passes to a person: no extra task.
- Flexible flow activationManual start, from API, from an event, from an AI agent, or scheduled for a future date.
Use cases
Frequently asked questions
Does it also apply to APIs and the browser extension?
No: confirmation concerns starts from AI agents. API Trigger, extension, events and sub-processes follow their own rules.
Who can confirm?
Whoever has permission to start that flow, and always Admins.
What happens if nobody responds?
At expiry the case is automatically rejected or flagged as expired, depending on configuration.
Is there any cost for pending cases?
No: until there is confirmation, no tasks or AI calls start.