AI should draft the email, but it should never invent the reason for sending it. At Understory, AI combines research, drafts copy, and personalizes messages at scale, but only from verified data a human could point to. A human owns every claim before a campaign goes live. If the signal isn't real, the sentence doesn't ship.
We support automation and build AI-assisted outbound workflows in Clay and Claude for SaaS clients every week. These are the safeguards we use in live campaigns.
What AI-generated outbound actually looks like today
The tooling matured fast. A typical AI outbound stack now handles the full sequence from prospect identification to send. Data sourcing pulls contacts from LinkedIn, job boards, and funding databases. Enrichment layers add verified emails, titles, and firmographics. An AI research agent reads the prospect's company site, recent posts, and relevant news. A drafting layer writes a personalized opener referencing what the research found. Automated sequencing handles follow-ups, and inbox warm-up keeps sender reputation stable.
The result is that any team with a credit card can launch thousands of AI-written emails in an afternoon. Two years ago, running personalized outbound at that volume required a team of SDRs and weeks of manual research. A single operator can now match that output in hours. The barrier to entry dropped from headcount and months of ramp time to a software subscription and a weekend of setup.
Where teams differ is how much of this pipeline they actually verify. On one end, AI handles research and drafting while a human reviews every send. On the other, autonomous AI SDR tools run the entire workflow from prospect identification through delivery, with no human checking individual messages before they ship. Most teams fall somewhere between these poles, and the gap between "AI-assisted" and "AI-autonomous" is where the risk concentrates.
That gap raises a practical question for any growth leader running outbound: where does useful automation end and fabrication begin?
Where AI outbound goes wrong
The failure cases are public. TechCrunch investigations have documented AI SDR hallucinations, product failures, and customers discontinuing deployments.
Buyers feel the effect even when they can't identify individual AI emails. A 2025 Gartner survey found that 73% of B2B buyers actively avoid suppliers that send irrelevant outreach. The pattern is clear: prospects penalize irrelevant messages whether or not they can tell a machine wrote them.
Then there is fabricated evidence of attention: the "I noticed your recent post" that no one noticed, generated thousands of times a day. Name and company data, even paired with a job title, do not constitute research. They constitute a mail merge.
The problem compounds at scale. When thousands of fabricated openers hit the same SaaS vertical, recipients compare notes. In tight communities where VPs of Marketing and Heads of Growth talk to each other, one exposed fabrication can circulate and damage the sender's credibility across an entire target segment before the first reply comes in. Brand damage in a close-knit vertical travels faster than any outbound sequence.
The line we draw
We run AI-assisted outbound at Understory on Clay enrichment and Claude drafting. The machine does the research and writes the first draft; a human owns the send.
Research depth matters, which is why we build around full prospect research. AI drafts against verified enrichment fields: a scraped job posting, a funding round with a date, a tech-stack detection, or a CRO hire. A human owns the targeting and the offer. Every factual claim in the send belongs to that human. When the data column is empty, the AI leaves the sentence blank.
We catch this in our own builds by reading a sample of drafted rows in Clay before a campaign goes live. On each sampled row, the reviewer confirms that the signal column is populated and includes a source URL or date, then checks the drafted opener against the signal the opener references. If you're a founder with nobody free to read every draft, the hard SKIP gate is what makes the workflow safe. Trust-based review doesn't survive volume.
Each example below pairs an assisted version with a fabricated version of the same signal:
- AI-assisted: Clay scrapes an open RevOps job posting from the company's careers page, and Claude drafts one line referencing that specific role and posting date. AI-fabricated: "I saw you're scaling your SDR team" when no posting was found.
- AI-assisted: an enrichment provider returns a funding round announced within the last 90 days, and the opener references it. AI-fabricated: congratulating a prospect on a Series B that never happened.
- AI-assisted: job-change monitoring flags that a CRM contact moved companies, an email waterfall checks multiple data providers to find the new address, and outreach references the actual move. AI-fabricated: pitching a prospect off stale database records, wrong sector, wrong level.
- AI-assisted: a web research agent pulls a real, current LinkedIn post and the draft quotes what it actually says. AI-fabricated: hallucinating a post and pitching off it.
The difference in each case is whether the signal existed before the AI wrote the sentence. That's the line.
In our Clay workflows, we use conditional rules so an AI column only runs when the enrichment column has a result. Our prompts follow one rule: reference the signal specifically, and "If the signal is vague or missing, return exactly: SKIP." Signal-gated campaigns skip rows, which means fewer bad sends and a healthier domain.
Broader research on AI text generation identifies mass manipulation, low-quality content, and unclear accountability as structural risks. Heavier AI assistance can also put sender trustworthiness at risk.
This approach costs volume. A list gated on real signals will always be shorter than a blast. We think that trade is correct.
The ethics line and the deliverability line are the same line
The guardrails that prevent fabricated outreach also protect sender reputation. Spam complaints leave little room for error at scale, and generic AI-generated outbound can exhaust that margin fast.
Signal-triggered campaigns earn relevance through verified data. Fabricated or generic personalization can undermine the campaigns you actually researched.
Email providers track complaint rates, bounce rates, and engagement at the domain level. A cold outbound campaign that generates spam reports doesn't just hurt that campaign; it reduces inbox placement for every email the business sends, including customer onboarding, renewal sequences, and support communications. A growth leader who owns nurture and lifecycle email should treat sender reputation as a shared asset across the business; one team's volume experiment becomes a wider operational problem when domain health drops.
Tight ICP segmentation before a single line gets drafted is where we see the biggest lifts. That protects both reply rate and reputation, and it keeps messaging consistent across paid media, outbound, and creative.
Build signal-gated outbound with Understory
We build allbound programs for B2B SaaS teams that run outbound on real signals: a recent CRO hire, a fresh funding round, a tech-stack change detected in Clay. Claude drafts only when the row contains a verified signal. Missing signals trigger a skip.
Paid media, outbound, and creative sit under one team, so prospects who receive researched emails see consistent messaging across LinkedIn ads, retargeting, and sales collateral. We built Yofi's outbound system from scratch using signal-gated Clay workflows; the pipeline generated so many qualified leads they had to pause campaigns to keep up with sales capacity.
If your outbound is high-volume and low-signal, or you're evaluating AI SDR tools, book a consultation with Understory to build a signal-gated program for your ICP.
FAQs
Is it ethical to use AI to write cold outreach emails? Yes, when AI uses verified data and a human remains accountable for the claims. The ethical breach is fabricating signals: inventing posts, funding rounds, or hiring activity that never happened.
Can prospects tell if an email was written by AI? Trust penalties show up even when prospects cannot identify individual AI-written emails. Decision-makers react to messages that feel generic or disconnected from their actual situation, and medium-to-high levels of AI assistance can put sender trustworthiness at risk.
Does AI-generated outbound hurt deliverability? Deliverability risk comes from scaling generic sends that recipients reject. AI makes that risk cheap to produce at scale, which is why strict signal gates and human accountability matter.
