Which AI BDR decisions should a person approve?
Estimated reading time: 5 minutes
Keep approval with a person when an AI BDR changes whom you contact, what you claim, or what the company commits to. Automate bounded work only after its inputs, checks, and exceptions are clear. Give someone authority to stop the workflow when those conditions fail.
“Human in the loop” leaves too much unspecified. A person may approve an account list, skim a sample email, or release every message. Those arrangements give different protection and require different amounts of work.
For a B2B team, the useful unit is a decision with an owner, evidence, and an expiry condition.
Approve the action and its allowed scope
An approval should identify the action it authorizes. A reviewer who accepts a research summary has not necessarily authorized an email. A reviewer who accepts a message template has not necessarily accepted every company selected for it.
OWASP's guidance on excessive agency recommends minimum tool permissions, human approval for high-impact actions, and authorization enforced in downstream systems. We apply that principle to outbound decision rights here; OWASP does not prescribe this sales matrix. OWASP's excessive-agency guidance.
Separate preparation from execution where you can. A research worker can prepare a draft without having permission to send it. A sending service can release an approved record without being allowed to change the offer. Check the actual product controls before assuming either separation exists.
Use a decision-rights matrix
This is a suggested starting matrix for a B2B campaign. The roles describe responsibilities, not required job titles. One person can hold several roles, but each row still needs an owner.
| Decision | What automation can prepare or execute | Who approves | When approval expires or work stops |
|---|---|---|---|
| Offer claims, customer proof, and exclusions | Retrieve approved material and flag missing support | Seller who owns the offer | A claim, permission, or offer changes |
| Account criteria and a new segment | Gather evidence and propose inclusion or exclusion | Campaign owner with market knowledge | The segment or inclusion rule changes |
| Buying-role selection | Propose contacts with evidence of responsibility | Account reviewer for ambiguous or sensitive cases | Identity, responsibility, or account ownership conflicts |
| Research-based message assertions | Attach source passages and draft within their scope | Research reviewer for unresolved or consequential assertions | Evidence is missing, contradictory, or no longer current |
| Sending within an approved campaign | Release records that pass the agreed checks | Campaign owner approves the operating envelope | A stop event occurs or an input leaves the envelope |
| New timing, channel, volume, or campaign overlap | Propose a change and show affected accounts | Sending owner and account owner | The change lacks approval or a restriction appears |
| A buyer reply or commercial commitment | Route the conversation and prepare context | Sales owner | Keep automated follow-up stopped until the owner decides |
| Restart after an incident | Collect diagnostics and propose a correction | Named incident owner | Correction and affected records remain unverified |
For each row, write the account scope, allowed action, source requirements, and time or event that ends approval. “Approved by sales” is insufficient if nobody can tell which sales decision it covers.
Review exceptions without letting them accumulate silently
Use three possible outcomes: approve within scope, return for correction, or hold for a named decision. Include a reason and the affected record. A reviewer should be able to reject the proposed action without editing it into an acceptable one.
Set an escalation path for unanswered reviews. When the owner is absent, the default should preserve the hold. Assign a substitute if continuity matters. Do not treat a review deadline as automatic permission to send.
Record the complete proposed message alongside its evidence. Reviewing only a subject line or a research snippet misses what the recipient will read. For a bulk approval, preserve the exact population and revision approved. Adding more accounts or changing the offer should require a new decision.
Illustrative exception: a placement provider targets manufacturers hiring maintenance engineers. A vacancy page verifies the role, but the research cannot determine whether the plant uses agencies. The reviewer can approve a question about external hiring support. The reviewer should reject a message claiming the company already uses agencies. That approval covers the narrower wording, not an invented procurement fact.
Match the review effort to the consequence
Choose a review scope the owner can sustain. If each message needs substantive judgment, plan for individual review. Reduce the account allocation when review capacity is the constraint.
At greater volume, distinguish checks that can be enforced from judgments that need review. A required source field or a blocked account ID can be checked mechanically. Whether an inference is fair to the buyer needs a substantive review. A completed field does not establish that its content is true.
Sample review can help monitor stable, lower-consequence work. It does not authorize individual records that still require approval under the matrix. Inspect both ordinary records and known difficult cases, such as similar company names, changed responsibilities, or missing sources.
Account for reviewer capacity. If releases depend on one person reading every draft, put that work in their schedule and in the campaign cost. A review queue that nobody can clear is a reason to reduce the campaign scope or change the process.
Test whether the approval changes the actual workflow
Run a dry test with an unapproved record. Confirm that it cannot reach the sending queue. Reject a record already awaiting release and inspect whether the rejection removes its authority to send. Change an approved message and check whether the old approval remains attached.
Test stop events separately from approvals. An opt-out, a reply, or a detected identity error should trigger the agreed hold without waiting for the next sample review. Preserve the stop event and identify every affected campaign.
For Quantonica's managed outreach, the client reviews the researched contact list and messages during preparation. When a buyer replies, Quantonica stops the sequence and the client handles the conversation. The client still needs a sales owner who can act on that handoff.
Choose one decision currently described as “reviewed by a human.” Name its owner, permitted action, evidence, and expiry condition. Then run a rejected record through the workflow and verify that the action stays blocked.