Our AI Sales Agent Tanked Our Sender Reputation. It Wasn't a Copy Problem.

Tuesday, 8:47 AM.

"Your emails are going to spam. Fix it by Friday."

That was the Slack message from our CRO. No build-up, no context, just that. I stared at it for a moment, trying to figure out why he'd sent it to me. I'm the quality and brand compliance manager at a B2B SaaS company. I review outbound communications for a living. Email infrastructure isn't my lane.

Then I opened the deliverability dashboard he'd attached. We'd gone from a 98% inbox placement rate to 61% in three weeks. Our AI sales agent—live for about a month—was responsible for most of that outbound volume. An AI agent we'd set up to automate prospecting, personalize first-touch emails, and let our SDRs focus on actual conversations. Which it did, right up until every email it sent started landing in spam.

And I was the one who approved it.

I'm the Person Who Approves Everything

My job is straightforward most days. Every piece of outbound communication passes through my review before it reaches customers. Roughly 1,200 items a year—email campaigns, sales collateral, product updates, support announcements. In 2024, I rejected about 12% of first deliveries. Typos we can fix. But claims without evidence, off-brand messaging, compliance issues—those go back to the drawing board.

When our VP of Growth pitched the AI sales agent idea in early 2025, I wasn't skeptical. The logic was sound: an agent that researches prospects, enriches contact records, and sends personalized first-touch emails. Every vendor was saying "agent-native" at the time, and our team had a real use case. Manual prospecting was eating hours that SDRs should have spent talking to interested buyers.

My assignment was to review the agent's output quality. I spent two weeks on email templates, tone guardrails, objection handling, and claim structures. We made sure the agent wouldn't state anything we couldn't back up—which matters, because per FTC guidelines (ftc.gov), commercial claims have to be truthful and substantiated, not just plausible. The agent's copy was clean. I signed off. Launch day arrived.

The AI Agent I Signed Off On

The first month looked great, honestly.

The agent wrote well. It followed our messaging hierarchy, kept our brand voice fairly consistent, and handled basic objections. Week three, the reply rate hit 3.1%. Our human SDRs averaged around 2% on cold email. The agent was outperforming the team, sending 1,500 emails a week with almost no one paying attention to the underlying contact data.

That's the part I keep replaying. We obsessed over every word of the agent's emails. We did A/B testing on subject lines. We wrote a 27-page brand voice guide. But the agent was pulling contacts from a database that included a heap of scraped LinkedIn profile data. A contractor had built a sourcing script that went through LinkedIn and pulled profiles into a CSV. It was fast, free, and full of junk.

Stale emails. Guessed email formats. People who had changed jobs months or years ago. Duplicate records. We all assumed the data layer was handled. It wasn't.

It Was the Data All Along

Here's something vendors and sales engineers won't volunteer: scraped data isn't verified. It's a snapshot of a profile someone saw once. The email addresses are often inferred. Phone numbers, if they exist at all, haven't been checked in months. Sending to data like that is a lottery—most of it lands badly, and the people you do reach aren't happy about it.

The most frustrating part of the whole situation: nothing in my standard checklist caught it. You'd think quality control would include deliverability, but it's not covered by the tools I normally use. The copy was good. The agent was doing its job. The emails just never arrived.

By the time the CRO flagged it, we were at a 40% bounce rate and a 0.4% spam complaint rate. For context, bounce rates above 3% raise red flags. Above 5% is severe. We were at 40.

I spent the next week learning sender reputation from scratch.

What Sender Reputation Actually Is

Here's the simplified version: ISPs like Google and Microsoft score you on every email you send. Bounce rate, spam complaints, engagement, recipient actions. That score is your sender reputation. It attaches to your domain and your IP address, and it determines whether your messages land in the inbox, the promotions tab, or the spam folder.

The maddening thing about sender reputation is how asymmetric it is. It takes months to build and days to destroy. You can send a thousand excellent, personalized, compliant emails—then one bad batch from a scraped list flips the switch.

There's a parallel in my old print world: federal law, 18 U.S. Code § 1708, says only USPS-authorized mail goes in a residential mailbox. Violate that, and there are fines. Email providers have a similar rulebook, but the enforcement is quieter. Instead of a fine, they just route your messages to the spam folder where nobody sees them. Your sender reputation is the key, and we'd lost ours.

Why "Verified" Needs a Definition

Once I understood deliverability, the data question became urgent. Where exactly was the agent getting its contacts?

Mostly from that scraped LinkedIn dataset. And as I dug into it, I learned something about the sales data industry that I didn't expect: "verified" doesn't mean the same thing everywhere.

What most people don't realize is that some vendors verify a contact once when they first collect it, then keep the "verified" label forever. Others actually re-verify at point of use—checking that the email still works, the phone number still connects, the person still works at the company. The methodology matters more than the label. If you're feeding an AI agent that will send 1,500 emails a week, you want data that's verified now, not data that was verified on some random Tuesday in 2023.

I also had to think about compliance. Scraping LinkedIn profiles at scale runs into GDPR concerns and platform terms of service. But there's a legitimate way to get good B2B data: providers that build datasets through compliant channels, with documented legal bases for processing. For a company with EU customers, that's not a nice-to-have. It's a requirement.

Evaluating Cognism Like a Procurement Spec

I compared three data providers. I won't name the other two—they weren't terrible, they just didn't pass the quality bar. And here's a small observation that mattered: two of them required a sales call just to see a price. Cognism had a public pricing page. Tiers, included features, clear terms.

Actually, that's not quite right—I'd heard the name Cognism in procurement conversations, but I'd never evaluated it as a quality person until this mess forced me to. In my print procurement days, that kind of pricing transparency was normal. I could quote 500 business cards from three online printers in fifteen minutes—budget tier around $20-35, premium with thick stock and coatings $60-120, all listed publicly as of January 2025. When a software vendor hides pricing behind a demo call, it tells me something about how they think about trust. Cognism's pricing page was a genuine differentiator.

I also focused on the contact discovery tool features that mattered for our agent:

• Verified phone numbers: direct dials and mobile numbers, not just main switchboards.
• Email verification: whether emails are checked at the point of use or just labeled "verified" once.
• GDPR compliance: documented legal basis for processing, especially important for EU prospects.
• Agent-friendly workflow: API access and integrations that let the agent pull enriched contacts without manual steps.
• Intent signals: useful for prioritizing which prospects the agent contacts first.

On all of those, Cognism passed. The phone verification piece stood out—Cognism didn't claim 100% accuracy, which I respected. Instead, it described a live verification process. In my line of work, vendors who acknowledge limits are more trustworthy than vendors who promise perfection.

The Recovery Was Boring

Switching the agent to Cognism data took about a week. Then came the unglamorous work: rebuilding sender reputation.

We parked the damaged domain and migrated to a new one. We warmed up the new domain slowly, gradually increasing volume. We throttled the agent's sending while the reputation score recovered. The CRO checked in daily. The sales team asked where their replies were. There was a stretch where I honestly wondered if I should update my resume.

But the numbers improved. Week 6, we were back above 94% inbox placement. Week 8, 97%. The agent's reply rate went from 1.4% at the low point to 3.6%—actually above the pre-incident baseline. Better data didn't just fix deliverability. It made the agent's emails more relevant, because the contacts were current and the context was accurate.

It wasn't a heroic fix. It was a boring, systematic correction of the data layer.

What I'd Tell My Past Self

If you're running an agent-native prospecting workflow, sender reputation isn't an email marketing concern. It's the system health metric for your entire pipeline. An AI agent can send thousands of personalized emails overnight. That's power, but it's also risk—because bad data doesn't just waste money on bad outreach. It actively destroys your deliverability, which destroys every future email you send from that domain.

In practice, that means data quality and sender reputation are part of the same quality spec as your email copy. You can't review the templates and ignore the contact data. You can't celebrate a strong reply rate while your bounce rate climbs in the background. Every layer of the system needs a quality check, and the contact data layer is the one you truly cannot skip.

Per FTC guidance, claims have to be substantiated. I'd extend that to your entire outreach stack: if a data vendor says "verified," ask what the verification process actually does. If they say "GDPR compliant," ask for the details. If a tool promises you'll reach more prospects, ask where those contacts are coming from—and whether sending to them will cost you more than just the subscription fee.

I still approve every email template before it ships. But now I also check the sender reputation score, the data verification methodology, and the compliance documentation with the same intensity I once reserved for subject lines and commas. It took an AI agent tanking our deliverability for me to learn that lesson. I'd rather you learn it from this story.