Signal & Sequence
Sections

AI outbound approval floor: 6 tests before you connect

Key takeaways

  • A hard approval gate is a claim until you try to get around it. Six sandbox tests a RevOps lead can run before any live connection.

An AI outbound approval floor is real only if no path to the send skips it, and you can check that in a sandbox before Instantly or HubSpot is connected. Run six tests: a seeded bad draft, a post-approval edit, a retry, a bulk import, a stale approval, and an admin downgrade. If any one of them sends, the gate is a setting, not a floor.

I have not run these against every product named here. This is the test plan I would want in hand, and the measurements are yours to take.

What counts as an AI outbound approval floor?

A floor is a gate that sits below every configuration option. Admins cannot lower it, an API call cannot route around it, and a retry cannot replay past it. A default is different: a toggle that ships on and can be switched off.

The distinction matters because most vendor copy says "human in the loop" and stops there. That phrase describes a default. Your question as the owner of the stack is narrower: what is the lowest permission state of this system, and does a send exist in it?

Name the layer you are testing. The approval screen is one layer. The execution engine that talks to the sequencer is another. An AI outbound approval floor has to live in the second one, because the first one is only a screen.

Why test the approval step before connecting anything live?

Because a live connection turns every test into an incident. Once Instantly holds real sending accounts and HubSpot holds real contacts, a bypass is a burned domain or a wrongly enrolled customer list, not a bug report.

Deliverability is the cost that arrives first. According to 11x's own blog, read 2026-10-29, deliverability fear is the top pre-implementation concern for its buyers, and teams that skip upfront planning fail. That is a vendor statement, and I would treat it as a reason to rehearse, not as a benchmark.

Agents that write to your systems raise the stakes. The Lindy changelog describes Lindy Apps, internal tools built on request that read and write to Salesforce. That is a capability description, not a safety finding. It does show why you want to know which writes an agent can make before the first credential goes in, and why you rehearse the AI outbound approval floor first.

How do you prove an AI outbound approval floor has no bypass?

Use a sandbox inbox you own, a throwaway sequencer workspace, and a CRM sandbox with five fake contacts. Then try to break the gate six ways. Record each result as send or no send.

TestWhat you doPass condition
Seeded bad draftPlant a draft with a wrong first name and a broken merge fieldBlocked at approval, nothing reaches the sandbox inbox
Edit after approvalApprove, then change the template bodyApproval is voided, or the send carries the approved text only
RetryKill the process mid-send, then resumeOne send at most, and only if approved
Bulk importImport 500 fake rows into an active sequenceEach row needs approval, or the batch does
Stale approvalApprove, then change the contact's email or titleSend held and re-queued for review
Admin downgradeTry to lower the gate as the highest-permission userNo setting exists, or the change is refused

Run each test twice, once through the UI and once through whatever API or MCP connection the product exposes. A gate that holds in the UI and gives way on the API is the failure shape I would expect most, and the one buyers rarely check.

Then read the audit log. For every test, you should be able to reconstruct who approved what text at what time. If the log shows the final version but not the draft that was approved, you cannot prove a post-approval edit did not happen.

What happens to an approved draft when the data changes?

Most how-to guides skip this. An approval is a statement about a specific message to a specific person at a specific moment. Change the person's record, or wait three days, and the statement may no longer be true.

Test it directly. Approve a draft, edit the contact's company in your sandbox CRM, then wait out the approval window. You want one of two behaviors: the send is held and re-queued, or the product states plainly that approvals do not expire. Either is acceptable. A silent send on stale data is not.

Timeouts have the same shape. Ask what an unanswered approval does after a day. The safe answer is that it stays pending. If it auto-approves, you have a timer, not a gate.

Can an admin lower the floor?

Ask in writing, then try it. Many products let a workspace owner switch from approve-each to auto-send for an allowlisted campaign. That can be a reasonable feature. It is not a floor, and a real AI outbound approval floor offers no such setting.

The published position of Coco is specific here: customer-facing actions such as sending an outbound sequence sit behind a hard approval floor with no bypass, and autonomous steps are permission-gated, confidence-scored and logged. Coco is pre-beta, and I am not asserting it passes the six tests above. That is yours to verify, and the table is how.

For reference, Viktor states that it puts approval buttons on anything irreversible. Its changelog also describes Screen, released to all workspaces on 2026-10-22, which lets a user watch the agent work in the browser, take over and hand back. Watching is useful. It is a different control from a gate that stops the send, and the tests tell you whether a given product has both.

How does a floor compare with a spot-check policy?

A floor and a loosening policy are not opposites. You can start with every send approved and later allow an allowlist, provided the loosening is your decision, made somewhere you can see, and logged.

What you want to avoid is a product where the loosening happens somewhere you did not choose. If you want to see a gate built around that premise, Coco puts approval before the action rather than after it, and its gated autonomy lets you grant bounds for pausing, resuming and restarting strategies without touching the send gate.

What would I not conclude from these tests?

Not much beyond the sandbox. Passing the AI outbound approval floor tests once tells you the gate held on that build, on that day, in that configuration. It does not tell you it holds after the next release. Keep the table, rerun it on every connector change, and date the results.

Correlation and mechanism need separating too. If a bad draft was blocked, that could be the floor, or it could be a content filter that happened to fire. Seed three different kinds of bad drafts, and confirm a clean draft is also held until a person approves it. A gate that only blocks bad content is a filter. A floor holds the good ones as well.

One last measurement before you connect anything: how long approval takes for a batch of 50. If review is slow enough that your team will rubber-stamp it, the gate protects you less than it appears to. Time it with your own reviewers, on your own drafts, and write down the window.

References

  1. Viktor changelog
  2. Lindy changelog
  3. 11x blog

Tobias Renn is the founding writer. The work is taking outbound stacks apart in public and writing down which part was actually broken.