In brief
- Grant read and draft bounds freely, keep every write that moves credit behind approval, and refuse comp data until the execution log can reconstruct a payout.
Grant an AI co-worker read access and drafting freely, grant bounded writes to non-comp fields with approval, and grant nothing that touches commission or quota data until its execution log can reconstruct a payout line by line. Those are the three positions a RevOps lead should hold when reading AI co-worker autonomy clauses, and each follows from what a single field write can pay.
What do AI co-worker autonomy clauses actually pay for?
Read a clause the way a seller reads a plan clause: ask what it permits, then ask who gains when it fires. "The agent may update CRM records within granted bounds" sounds like hygiene. In a comp context it is a payout instruction.
Take a worked example. The figures are assumptions, not data. Quota is $600,000, paid at 8% up to quota and 12% above it, monthly in arrears. Seller A closed $660,000. Seller B closed $540,000.
- Seller A earns $48,000 on the first $600,000 and $7,200 on the $60,000 above quota: $55,200.
- Seller B earns 8% of $540,000: $43,200, still below target.
Now an agent runs a duplicate-account cleanup. The merge keeps the surviving record's owner, which is Seller A, and a $60,000 closed-won deal that belonged to Seller B moves across. Seller A is now at $720,000: $48,000 plus $14,400, or $62,400. Seller B drops to $480,000 and $38,400.
One field write moved $7,200 to A and removed $4,800 from B. Company cost rose by $2,400, because the moved dollars left the 8% band for the 12% accelerator. Nobody intended it and no clause was breached. That is why "within granted bounds" needs a definition before the agent touches an owner, amount or close-date field.
Which bounds are safe to grant?
Grant the bounds whose worst case is a bad draft, not a changed payout. Account research, enrichment suggestions, outreach drafts and read-only pipeline summaries all fit. A wrong output costs a reviewer some minutes.
Grant bounded writes next, but only to fields the comp plan never reads. Job titles, industry tags and contact phone numbers qualify. Owner, amount, stage, close date and any crediting split do not. A plan pays on those fields, so they are comp data even when they sit in a general-purpose CRM.
The test is short: if changing the field could change a seller's number on the month-end statement, it sits on the gated side.
Which steps stay behind approval?
Anything that moves credit, and anything customer-facing. The published design of Coco draws the second line explicitly: sending an outbound sequence sits behind a hard approval floor with no bypass, and the product asks before risky steps. Viktor states that it puts approval buttons on anything irreversible. Both are vendor statements of intent, and a RevOps lead should read them as clause language to test, not as proof.
Continuous autonomy is the harder case. Coco strategies can pause, resume or restart within bounds the user grants, and each autonomous step is permission-gated and confidence-scored. The grant is the clause that matters. Write the AI co-worker autonomy clauses narrowly: which strategies, which objects, which fields, which actions. A bound that says "CRM records" is a blank cheque. A bound that says "contact enrichment fields, no owner or amount writes" is a plan clause you can audit.
Two vendor pages show why wording matters. The Artisan homepage pitches an AI BDR that runs in your stack, alongside your team, with a mode to run alongside the stack or replace it. The Relevance AI blog describes an agent builder that tests the agents it builds and keeps watching after go-live. In both cases the comp owner's question is the same: after go-live, who can change what the agent may write, and does that change need your sign-off?
How do you read AI co-worker autonomy clauses line by line?
Ask four questions of every clause and write the answers down before signing.
- Who holds the grant? A bound granted by an individual rep is not a bound granted by RevOps.
- Is the bound a field list or a category? Categories widen quietly.
- Does the clause say what happens when the agent is unsure? Confidence scoring only helps if low confidence routes to a human.
- Can the grant be narrowed without a vendor ticket?
Metering clauses deserve the same reading. The Clay pricing page shows two separate meters, actions and data credits. A clause that pauses work when credits run out is an autonomy clause too: a cleanup that stops halfway leaves a half-merged table at month end.
What must an execution log show before comp data is touched?
Most guidance stops at "keep an audit log". Finance needs more than that to rebuild a payout. For every write, the log should hold:
- The record and field changed, with the value before and the value after.
- The acting identity, and the human who granted the authority it used.
- The strategy and step that triggered the write, with its version.
- The confidence score and whether a human approved it, with approver and time.
- A link to the rollback action, and whether it was used.
Retention should outlast the longest clawback window in the plan, plus the time finance needs to close disputes. Set that number from your own plan, not from a vendor default. The facts published for Coco include a logged execution trail and versioned strategies with rollback. A RevOps lead should ask whether those entries carry the fields above. The product is pre-beta, so that is a question for a demo, not an assumption.
How do you test an agent before it sees quota data?
Run it in shadow mode first. Give it a read-only copy of the CRM, let it produce the writes it would have made, and replay them against last quarter's plan. For each proposed write, compute the payout delta for every seller affected. In the example above, one merge shows a $7,200 swing and a $2,400 net cost. A shadow run with zero deltas on owner, amount and close date is the sign-off evidence. One with three deltas is a bound to tighten.
Ownership of that sign-off belongs with whoever signs the plan: RevOps proposes, finance approves, and the agent never approves a step that pays someone.
Which approach fits an approval-first rollout?
An approval-first design fits this sequence, because it starts from gates and widens bounds, rather than starting wide and adding gates after an incident. A short look at Coco's onboarding shows context elicitation at quick, standard or thorough depth, which is where a RevOps lead can state which fields are off limits. That conclusion rests on stated design only. No reliability or results claim is made, and none is available.
The practical order is simple: grant reads, gate writes, shadow-test, and only then discuss comp fields, with the log fields written into the contract. Held to that order, AI co-worker autonomy clauses stop being boilerplate and become plan language you can audit.
Sources
- Artisan homepage — Artisan, 2026-11-08
- Relevance AI blog — Relevance AI, 2026-11-08
- Clay pricing — Clay, 2026-11-08