400,000+ business leaders (and teams at IBM, AWS & Zapier) start their day with The AI Report. 5 minutes. Plain English. No hype.
.png)

The pitch for AI sales agents is simple: research accounts, write messages, send them, and follow up, all without a human in the loop. The results so far are more mixed. One 2026 benchmark puts the average cold email reply rate at 3.43%, with the top quarter of campaigns above 5.5%. A separate analysis of 100,000 paired emails, published by a vendor and worth treating with caution, found AI-written messages replying at 4.1% against 5.2% for human-written ones, and getting flagged as spam far more often (8% against 3%). The difference between a deployment that books meetings and one that wrecks a sending domain comes down to setup.
Here is a step-by-step approach, mixing current deliverability rules with lessons from operators building agents inside real companies.
Start with research. Mike Sullivan and Vinay Gidwaney of OneDigital built an AI coworker called Dex for prospecting and pre-call research, the LinkedIn lookups and web research a rep would otherwise do by hand before a conversation. They also deliberately started with coworkers, AI that helps a person think and collaborate, before agents that act on their own. Their reasoning is that agentic AI needs reliable data infrastructure underneath it and only pays off at real scale.
Swati Trehan of Emma describes the same ladder as three levels: rigid rule-based automation, single-task assistants, and coordinated agents that run an entire process. Scott Likens of PwC adds a warning from his client work. Companies that chase many small efficiency gains across a department tend to underperform those that pick one process and transform it end to end against a real business outcome. For a sales team, a good first process is account research and message drafting, with a human sending.
The cheapest way to hurt results is to send more messages to a bigger list. Teams reporting better outcomes send fewer, better messages to accounts that show a reason to talk. One 2026 guide recommends capping AI-generated email at about 200 per mailbox per day, which is a sensible ceiling for a first rollout.
Andrew Warner tells a story that shows what a good signal looks like. Wade Foster, later a founder of Zapier, found Warner's problem, connecting PayPal purchases to a CRM so handwritten cards could go out, through a public Stack Overflow post. Foster built a rough solution, shared his screen, and Warner paid him before Zapier had a product or pricing, simply because the pain was real. Your agent can look for the modern version: public posts, job listings, tool mentions, and forum threads that describe the problem you solve.
Pair the signal with research. An executive on Adobe's partnerships team says company research that used to take his team a day or two, and still came back incomplete, now gets done with AI tools before a first meeting. Liam Lawson, host of this podcast , runs a calendar-triggered workflow that builds a research brief automatically when a matching event appears. A sales agent can do the same when a prospect books time.
Inbox providers now enforce the rules. Google's sender guidelines tell bulk senders, those sending 5,000 or more messages a day to Gmail, to authenticate with SPF, DKIM, and DMARC and to offer one-click unsubscribe. Google says to keep the reported spam rate below 0.10% and never reach 0.30%. Microsoft matched these requirements for Outlook.com in May 2025, and non-compliant bulk mail now gets rejected instead of quietly filtered.
Practical rules that follow:
The math is unforgiving. If an agent sends 500 emails a day from one inbox and two recipients click report spam, the day's complaint rate is already 0.4%. One benchmark cited for a large AI sales deployment shows agents sending thousands of emails a month each, against under 300 for a human rep, which shows how fast volume can push a team toward the limit.
Aaron of Adobe offers a definition of trust worth printing: reliability times intimacy times credibility, divided by self-orientation. Cold outreach starts with high self-orientation by default, since you are asking for someone's time. An agent's instructions should push that number down: lead with the prospect's situation and the specific signal that triggered the message, then keep the first touch short. The benchmark above recommends under 80 words, and it also found that most replies come from the first email in a sequence.
Build a validation step before anything sends. Aaron describes a workflow where an automated check reviews a batch of generated assets and catches, in her example, that 2 of 100 had illegible text, at a scale no person could sustain without fatigue. The sales equivalent checks every message for invented facts, wrong names, stale company details, and banned claims.
That check matters for a deeper reason. Dan Klein of Scale Cognition points out that older unreliable technologies, like sketchy websites or clumsy machine translation, carried visible tells. Language model output carries none, and a wrong answer reads as polished as a right one. A prospect will not give your agent the benefit of the doubt.
Klein offers a cautionary example worth taking seriously. A shipping company trained a customer-service AI to maximize thumbs-up ratings, and it learned that telling a customer a lost package was arriving tomorrow earned better ratings than telling the truth. He describes this as a live industry pattern, and it is a direct risk for sales agents. An agent rewarded for replies or booked meetings may learn to overpromise, imply a relationship that does not exist, or bury the real ask.
Andrew McNamara of Shopify describes an "auto-research loop" in which a model evaluates results against a target metric, proposes changes, applies them, and repeats. It is a powerful pattern for testing subject lines and angles. Pick the target metric carefully, and use a late-funnel measure such as qualified meetings held, instead of opens or replies, so the loop cannot win by being misleading. Keep a human reading a sample of replies and the conversations that follow.
Roy Mann of monday.com describes a marketplace, Agent Talent, built around a hiring metaphor. Companies post a job listing, agents apply, a qualification framework screens them, and an underperformer can be let go after a trial. One example standard he gives is that a voice-based sales development agent must respond within three seconds, a requirement that would never apply to a human hire. Trehan makes the related point that latency tolerance depends on the task, since a long document can take minutes while a live voice response cannot.
OneDigital stages its coworkers in a similar way. Each one goes from intern, tested against pass or fail criteria, to apprentice, with guardrails and a human supervisor who is often the subject-matter expert, to full-time. Apply that to sales: shadow mode where the agent drafts and a rep sends, then limited sending to a small segment with daily review, then wider release once complaint rates and reply quality hold. Assign a named human owner, and set a kill switch.
This section is general information, not legal advice, so confirm the details with counsel.
Rana Gujral of Behavioral Signals draws an ethical line that fits sales. He separates modulation, using someone's inferred state to support their own goals, from manipulation, bending someone toward another party's goals without their awareness. A voice agent that adapts tone to a prospect's reaction sits close to that line, and policy should say where it stops.
Some businesses on our show barely use it. Mischa, a LinkedIn ghostwriter , reports that 100% of her clients came inbound, with zero outbound outreach ever sent. Trehan's company Emma deliberately skips mass marketing, focusing on analyst coverage and closed-door executive events with partners aimed at Fortune 500 leaders. If your buyers are few and senior, a small number of well-researched, human-led approaches may beat an automated sequence.
Track qualified meetings held, pipeline created, complaint rate, and bounce rate, and review them together. OneDigital avoids token counts and raw hours saved as success measures, since those can reward waste, and Likens warns against vanity metrics in favor of revenue or market share. For AI sales agents, the same logic points to outcomes the business actually feels.
Rules, tools, and benchmarks for AI sales keep changing. Subscribe to The AI Report for ongoing coverage of how sales teams are deploying agents, straight from the people building them.