AI Sales Automation: What to Automate and What to Never Automate

AI sales automation delivers most of its value in research, triage and drafting, and almost none in the place everyone points it: writing the first cold email. That inversion is the most useful thing to understand about the category, because it is the opposite of how the tools are marketed.

This is for revenue teams evaluating AI SDR tools or deciding what to build into their own motion.

Why AI-Written Cold Email Underperforms

A model writing a personalized opener from a LinkedIn profile produces something structurally recognizable: a compliment about a recent post or a company milestone, a pivot, then the pitch. It is grammatical, it is on-brief, and prospects have now seen it several hundred times.

The problem is not quality. It is that the personalization is about information that is public and cheap, which communicates the opposite of what personalization is supposed to communicate. "I noticed your company raised a Series B" tells the reader you ran a filter.

There is also a deliverability cost. Thousands of senders producing near-identical structures from the same handful of models is a pattern that filters can learn, and at scale that is how an entire sending style degrades.

What works instead is narrower: use the model to find the reason to write, and write the sentence yourself. A model reading ten years of an employer's Form 5500 filings and surfacing "participant count grew 46% while the same generalist broker stayed on" has produced something genuinely non-obvious. The sentence built on that is specific because the finding is specific, not because the prose was generated.

Where AI Clearly Wins

Research and summarization. Reading a filing, an annual report, a job posting, three months of someone's posts, and extracting what matters. This is unbounded, tedious work that scales badly with humans and well with models.

Reply triage. Sorting inbound into interested, not now, wrong person, and unsubscribe, then routing accordingly. Cheap, accurate, and it fixes the leak that actually costs teams pipeline, replies sitting unread for four days.

Drafting replies to routine messages. "Can you send pricing," "who handles this," "we are in a contract until March." These have correct answers and a model produces them instantly, which is what Unibox and Ezra do inside grobot: replies across LinkedIn, email and chat land in one inbox and the routine ones get answered autonomously or drafted for approval.

Data hygiene. Deduplication, normalization, inferring seniority from inconsistent title strings. Unglamorous and genuinely well suited.

Note the pattern: AI wins where the work is bounded, verifiable, and currently done badly because it is boring. It loses where the work is a judgment about what will matter to one specific person.

Keep a Human Gate on Anything Outbound

The rule we run: AI can read anything and can draft anything, but a human approves anything that leaves the building for the first time.

Not out of caution for its own sake. The failure mode of an unattended generative step is not a slightly worse email. It is a confidently wrong claim about the prospect's business, sent from your domain, at volume, before anyone notices. One of those costs more than a quarter of the efficiency it was meant to buy.

Autonomous replies to known-shape inbound are a different risk profile and are fine. Autonomous first contact at scale is not, and any vendor selling it is transferring the risk to you.

Evaluating an AI SDR Tool

• Ask what it does when it has nothing to say. A tool that always produces a personalized line will produce a bad one for the 40% of prospects with no public signal. The correct behaviour is to fall back to a clean generic message, and most tools do not.
• Ask to see twenty consecutive outputs, not a curated three. The variance is the product.
• Ask where the research comes from. A model inferring from a job title is guessing; a model reading a filing is not.
• Ask what a human approves and whether that can be turned off. Then ask what happens when it is.
• Ask how it is metered. Per-send pricing on AI-generated outbound aligns the vendor with volume, which is the opposite of what you want. We meter AI actions, verifications and enrichments; sends are included in the license.

The Realistic Gain

AI does not multiply your outbound capacity, because the capacity ceilings are set by LinkedIn throttling and mailbox limits, not by how fast anyone types.

What it does is move hours from research and triage into selling, and raise the floor on the quality of what goes out when the rep is tired. That is a real gain and a boring one. Anyone promising a 10x pipeline increase from the same number of sends is describing a rate limit they do not control.