Approval gates for an AI co-worker work when three things are fixed before the first run: a short list of actions it may take alone, a hard stop in front of anything a prospect can see, and a log that lets a RevOps lead undo what was done. The sections below show the working. Every number is a round illustration unless a source is named.
What do approval gates for an AI co-worker actually gate?
An approval gate is a point in a run where the agent stops and waits for a named role to accept, edit or reject the next action. A permission bound is a limit on what the agent can touch at all, gate or no gate. Gates decide when. Bounds decide whether.
Sort every action by one question: can it be taken back? Reading a CRM record can. Editing a field can, if the old value was kept. A delivered email cannot. According to its homepage, read 2026-10-29, Viktor says it asks for approval before irreversible actions, and that dividing line is the one to use for outbound. The condition: it holds only while "irreversible" is defined by the RevOps lead and not left to the agent's judgment.
Which permissions should a RevOps lead grant first?
Start narrow and widen on evidence. Approval gates for an AI co-worker only mean something once the grants underneath them are small. The grants, in order:
- Read access to accounts, contacts and activity, limited to the fields the strategy uses.
- Draft-only access to outreach: messages are saved for review, nothing is sent.
- CRM writes on a short list of fields, such as enrichment fields. Owner, stage and amount stay read-only.
- Send caps set as hard bounds, not guidelines.
- A suppression check that runs before every send, covering opted-out contacts and any regional rule your counsel requires.
- A named approver role with a named backup, so a queue never waits on one inbox.
Here is the arithmetic for step 4. Three mailboxes at 30 sends a day gives a cap of 90 sends a day. A strategy that targets 120 contacts needs 120 first-step sends. 120 divided by 90 is 1.33, so the first step takes two days even if every draft is approved at once. The cap, not the approver, sets the pace. If your sending domains have different warm-up histories, set the cap per mailbox and not as a pooled total.
What should the execution log record?
A log is useful for rollback only if it holds the state before the action. Record these per step:
- The step, the strategy version and the timestamp.
- The data read and the data changed, with the before value and the after value.
- The confidence score the agent assigned, and the threshold that applied.
- Who approved, edited or rejected, and what the edit was.
The before value is the line most teams skip. Without it, a CRM edit is a one-way door in practice, even though it was reversible in principle.
Which actions can be rolled back, and which cannot?
Most checklists say "add a rollback plan" and stop. The plan depends on the action, and approval gates for an AI co-worker should be placed according to this table, not according to habit.
| Action | Reversible | How to reverse |
|---|---|---|
| CRM field edit | Yes | Restore the logged before value |
| Enrichment write | Mostly | Restore the before value, if one was logged |
| Sequence enrollment | Before send only | Pause or unenroll; delivered steps stay delivered |
| Delivered email | No | Only a follow-up; gate it beforehand |
| Strategy change | Yes | Restore the prior version |
The conclusion: put a hard approval gate on every row marked "No" and let the rest run under bounds. What would change it: a CRM with no field history turns the first row into a "No", and CRM writes then need a gate too.
How do approval gates for an AI co-worker avoid review fatigue?
Gate the plan once, then gate the sends in batches. Take a strategy with 120 contacts and 4 steps, which is 480 touches. Approving each touch at 20 seconds apiece costs 480 × 20 = 9,600 seconds, or 160 minutes. Now approve the plan once (audience, template, caps, 20 minutes), then approve sends in batches of 30 at about 2 minutes each. That is 16 batches, 32 minutes, 52 minutes in total.
The saving is real only when messages share a template with personalized fields. If every message is written from scratch, per-message review stays, and the 160 minutes is the honest figure. Batch approval also needs a sample: read at least a handful of drafts in each batch, not only the summary.
When should an AI co-worker earn more autonomy?
Promotion needs a threshold written before the evidence arrives. As an assumption for illustration, not a benchmark: after 50 consecutive drafts, the approver edits no more than 10% of them, which is 5 drafts, and the bounce rate stays under 2% of sends. Any complaint or opt-out violation resets the count. Promote internal actions first, such as enrichment and field cleanup. Customer-facing sends stay behind the gate regardless of the score.
Demotion should be as mechanical as promotion. One breach moves the action back to approval-required until the count restarts.
How does Coco implement approval gates for an AI co-worker?
Coco is built as an approval-first AI co-worker for go-to-market teams. According to its stated design, it asks before risky steps, and customer-facing actions such as sending an outbound sequence sit behind a hard approval floor with no bypass. That removes one line from the checklist above: the send gate is a fixed property, not a setting someone can loosen later.
Strategies on Coco run continuously and can pause, resume or restart within the bounds the user grants. Each autonomous step is permission-gated and confidence-scored, and each leaves a logged execution trail. Strategies and skills are authored behind a safety gate and versioned with rollback, which maps to the last row of the table. Listed integrations include HubSpot, Apollo, Instantly, Smartlead and Lemlist. The product is web-first and pre-beta, so this describes control and auditability, not a track record.
Other vendors publish their own controls. The Viktor changelog, read 2026-10-29, describes a Screen feature in which a user can watch the agent work, take over and hand back. Ask any vendor for the same three things before granting access: the fixed floor, the log fields, and the rollback path per action. Approval gates for an AI co-worker are only as good as those three answers.
An approval queue is the one meeting in RevOps that gets shorter when it works.