An AI SDR Approval Workflow That Protects Your Domain
In the comments under "the dirty secret behind every 'AI SDR' demo", the useful line is not about which vendor is faking a dashboard. It is this: the bottleneck is the approval surface — per-message — and a tool is not doing the job if it just makes spray-and-pray faster. Chris, who has written about both camps being wrong, described the split: half of founders bought an AI SDR, and the other half still will not let AI write a single email.
Pramod Kumar, co-founder of Kingdom of Kumar. On X
·
I sit closer to the second camp on *sending*, and closer to a third camp on *drafting*: let a model help, then make a person sign the mail. This article is that approval workflow. How to build the wider copilot is in How to build an AI SDR workflow. The hire-versus-software question is in AI SDR vs a human SDR. First lines are in Cold email personalization that isn't fake. The surrounding work is on GTM engineer and GTM engineering.
If the offer is not ready to sell, do not build an approval queue so you can reject mail faster. Take the App Launch Score first.
"The bottleneck… is the approval surface"
That comment is right, and it is inconvenient for anyone selling autopilot.
If a person has to read each first email, the number you can send is capped by attention. That is not a bug. That is the thing that protects the domain, the brand, and the offer. The workflow should make approval *fast and strict*, not absent. A queue that takes three clicks and shows the evidence next to the draft will get real review. A queue that hides in a Slack firehose, or that auto-sends after thirty minutes, will not.
Follow-ups that add no new claim can sometimes be pre-approved as a template. First touches, and any line that mentions a fact about the person, should be per-message until your miss rate is honestly low — and even then, I would sample, not disappear.
"Half… bought an AI SDR… The other half still won't let AI write a single email"
Both reactions are usually about trust, not about philosophy.
The buyers often wanted a headcount substitute. The holdouts often saw a bad draft, or a scary demo, and banned the model from the building. I would not ban drafts. I would ban unattended sends, and I would ban drafts that do not have to show their evidence. A model that cannot point at the field it used should not get to propose a first line.
If you are in the "no AI at all" half, try a week of drafts-with-approval on twenty accounts you already understand. If you are in the "we bought one" half, turn off anything that can send without a person, and run the same week. You may meet in the middle: a copilot, not a hire, and not a ban.
The workflow, step by step
1. Only eligible records enter the queue. Fit rule passed, never-contact checked, email verified, not already in an open conversation. The model does not get to vote a record into eligibility.
2. Evidence is collected before the draft. A source URL, a quote, a hiring post, a field in your CRM — something a person can see. If evidence is missing, the draft must stay generic, or the row waits.
3. The model drafts in a constrained format. First line, plus a short body, plus a list of claims it thinks it made. No hidden "personalization" in a P.S. you might miss.
4. A person reviews per message, with the evidence beside the draft. Actions: approve, edit and approve, reject, or "this person should never have been in the queue." That last action should improve the fit rule, not just that row.
5. Only then may a sequencer or mailbox send. The tool gets the approved text, not a prompt to try again on the way out the door.
6. Replies leave this workflow and go to a human. Auto-replies are a different, sharper risk; see the replies article.
7. Friday, review the rejects. If you are rejecting the same invention repeatedly, fix the prompt or the evidence. If you are rejecting the accounts, fix the list. If you approved mail you later regretted, tighten the rule and say so in the note.
A queue layout that makes review honest
For each row, show on one screen:
- Who they are, and why they are in (the fit reason)
- The never-contact checks that already passed
- The evidence, as a quote or link, not a summary the model wrote about the evidence
- The draft
- The claims list
- Buttons: approve / edit / reject / should not be here
If your current "AI SDR" cannot show that, you can still do it in a sheet. The product is the approval, not the logo.
Set a daily cap that matches the time you actually have. A backlog of three hundred drafts is how people start clicking Approve all.
A week-one setup, if you already bought a tool
Day 1: turn off unattended send. If you cannot find that control, stop using the tool for live mail until you can.
Day 2: export or rebuild the eligibility rules outside the tool, on a page you own.
Day 3: put twenty records through the queue layout, even if you have to do it in a sheet beside the tool.
Day 4–5: send only what you approved. Track rejects by type: invented fact, wrong person, bad fit, dull but honest, other.
Friday: if invented facts showed up, the tool does not get more volume next week. It gets a tighter prompt, better evidence, or a smaller role.
You may find you do not need the product at all, only a model, a sheet, and a mailbox. That is an acceptable outcome.
What I would never pre-approve
- A first line that mentions a fact
- Any mention of funding, a job change, or a competitor
- Pricing, legal promises, or "we already work with companies like yours" unless you have a written, allowed list of public names
- A change the model made to an approved template without a new review
Templates for the generic parts of the body can be approved once. The first line still comes through the queue.
When you might loosen, slightly
After you have weeks of review and a stable, low miss rate, you can loosen generic follow-ups, or allow a teammate to approve instead of the founder. I still would not jump to unattended first touches because a vendor said the model is "on brand now." Brands slip in specific claims, not in tone.
FAQ
Will this slow us down too much?
It will slow sending. It should speed up learning, because you will see every miss. If you need more volume than you can approve, you have a staffing or a cap question, not an approval problem. Do not solve it by removing the bottleneck.
Can two people share the queue?
Yes, if they share the fit rule and compare rejects on Friday. If they do not, you will get two standards, and the looser one will win.
Where does Kingdom of Kumar sit?
We build and run a scoped system on Grok and Claude for $500 a month, three-month minimum. You own it, and you approve what goes out. The approval surface is the product, not an afterthought. That is not the same as buying an AI SDR, and it is not a $135,000 employee.