Most outreach stacks fail quietly. A sequence goes out, replies trickle in below expectation, and nobody can say which of a dozen moving parts caused it. The AI-drafted copy, the enrichment data, the sending domain, the targeting filter, the follow-up cadence: any one of them can sink a campaign, and when several drift at once the damage compounds before anyone notices.
This checklist exists to catch those failures before they reach a prospect. It is built as a gate, not a wish list. Each item is something you can verify in an afternoon, and each carries a short justification because a check you do not understand becomes ritual rather than judgment. Treat any unchecked box as a launch blocker until you have a deliberate reason to waive it.
The items are ordered roughly by sequence, from confirming who you are targeting through to scheduling the review that keeps the system honest. Run the whole thing once against a new stack, then rerun the data and deliverability sections monthly.
Confirm the Target Before the Tooling
The fastest way to waste an AI outreach budget is to point a powerful machine at the wrong list. Personalization at scale only helps when the underlying audience is correct.
Verify the list maps to a real ICP
- Every contact should trace back to a documented ideal customer profile, not a scraped export of convenience.
- Spot-check twenty records by hand. If a quarter of them are clearly off-profile, your reply rate is capped before the AI writes a word.
Confirm enrichment data is current
- AI personalization is only as good as the fields it reads. A model that references a prospect's "recent" funding round from three years ago reads worse than no personalization at all.
- Check the freshness date on your enrichment source. Stale firmographic data produces confidently wrong copy.
Validate that every contact is reachable
- Run the list through verification before sending. A high bounce rate from invalid addresses signals stale data and damages your sending reputation on the first send.
- Remove role-based and catch-all addresses where your goal is a named human, since these inflate volume without producing real conversations and raise complaint risk.
Vet the AI Drafting Layer
The generative component is the part everyone focuses on and the part most likely to embarrass you. It needs explicit guardrails.
Read ten generated messages end to end
- Do not sample the first line and assume the rest holds. Hallucinated details tend to appear mid-message where they feel earned.
- Flag any claim the model could not have known from the provided data. Those are fabrications, and a prospect will spot them.
Lock the tone and length constraints
- Define a maximum word count, a banned-phrase list, and a required call to action in the prompt itself, not as a hope.
- If the tool cannot enforce constraints, treat its output as a first draft that a human approves, not as send-ready copy.
Confirm personalization degrades gracefully
- Feed the generator a contact with sparse data and watch what it does. A tool that invents a plausible-sounding detail to fill the gap will do so against real prospects too.
- The correct behavior is to drop the personalized line and fall back to honest generic copy. Verify this directly rather than assuming it. A confident fabrication is the failure that costs you the most credibility per send.
Protect Deliverability
The most personalized message in the world is worthless in a spam folder. Deliverability is the silent killer of AI outreach because volume tooling makes it easy to burn a domain fast.
Authenticate and warm the sending domain
- Confirm SPF, DKIM, and DMARC are configured and passing. Skipping this is the single most common cause of campaigns landing nowhere.
- Use a separate sending domain from your primary corporate one so a deliverability mistake cannot poison your real mailbox.
Cap volume per inbox
- Respect conservative daily limits per mailbox regardless of what the tool allows. The ability to send a thousand emails an hour is a liability, not a feature.
Build in a clean exit on every reply
- Confirm that a reply, an unsubscribe, or a bounce removes the contact from the sequence immediately. Nothing erodes trust faster than a follow-up that arrives after someone already said yes or no.
- Test this by replying to your own sequence from a seed address and confirming the next scheduled message does not send. Automation that cannot stop is a complaint generator.
Build the Measurement Loop
You cannot improve what you do not instrument. The checklist closes with the systems that let you read results honestly.
Define a primary metric before launch
- Decide in advance whether you are optimizing for positive reply rate, meetings booked, or pipeline created. Each implies a different sequence design.
- For a deeper treatment of which signals to trust, see Reading the Signal Behind Every Outreach Sequence.
Schedule a recurring review
- Put a fortnightly review on the calendar with an owner. Unowned dashboards rot.
- If you are still assembling the stack, Choosing the AI Outreach Stack That Fits Your Motion covers selection criteria in depth.
Hold the Checklist as Judgment, Not Ritual
A list invites box-ticking. The temptation is to race to the bottom so you can declare yourself finished, which produces a complete-looking checklist and a campaign that still fails. The justifications under each item exist precisely to resist that pull.
Decide whether each item applies
- Not every check carries equal weight for every campaign. A low-volume, account-based push to twenty named contacts cares less about per-inbox volume caps and more about the truth of each personalized claim.
- Read the justification, decide whether the risk it addresses applies to your stakes, and only then mark the box. A waived item with a documented reason is fine; a skipped item nobody thought about is not.
Treat clearing it as a gate, not a trophy
- The goal is not a tidy list. It is outreach you would be comfortable putting in front of a stranger who is actively looking for a reason to mark you as spam.
- When in doubt, imagine defending each unchecked box to that stranger. If you cannot, do not send. For a structured way to reason about the whole pipeline behind these checks, the SIGNAL model maps each item to the stage it protects.
Frequently Asked Questions
How long should running this checklist take?
The first full pass takes an afternoon for a single sequence. Later passes are faster because your banned-phrase list, enrichment checks, and authentication records already exist. The data and deliverability sections deserve a monthly rerun even after the rest stabilizes.
Can I skip the deliverability section if I use a reputable tool?
No. Reputable tools make it easier to authenticate correctly, but they do not do it for you, and they happily let you send at volumes that burn a domain. Deliverability is your responsibility regardless of vendor.
What if the AI drafting tool will not let me enforce constraints?
Then route its output through human approval before sending. A tool that cannot enforce a length cap or a banned-phrase list is a drafting assistant, not an autonomous sender, and treating it as the latter is how fabricated claims reach prospects.
How many generated messages should I actually read?
At least ten per template variant, end to end. Sampling first lines hides the mid-message hallucinations that do the most reputational damage. Reading the full body is tedious and non-negotiable.
Does this checklist apply to LinkedIn outreach too?
The targeting, drafting, and measurement sections apply directly. The deliverability section is email-specific, but LinkedIn has its own equivalent: connection-request limits and account-restriction risk that you should treat with the same caution as a sending domain.
Should I run this for every campaign or just new stacks?
Run the full checklist for every new stack and every major template overhaul. For routine new campaigns on an established stack, the targeting and drafting sections suffice, since deliverability and instrumentation carry over.
Key Takeaways
- Targeting comes first; AI personalization cannot rescue a list that does not match your ICP.
- Read full generated messages, not first lines, to catch the hallucinations that surface mid-body.
- Authenticate and warm your sending domain on a separate domain before any volume sending.
- Treat conservative per-inbox volume caps as a rule, not a suggestion your tool can override.
- Define your primary metric and schedule an owned recurring review before launch, not after.