AI Cold Outreach: Where the Model Helps and Where It Embarrasses You

The useful split in AI cold outreach is not between good and bad tools. It is between two kinds of work: reading, at which a model beats any human researcher on cost and patience, and writing one sentence a stranger will actually read, at which it does not. Most tools are sold for the second job.

This is for teams evaluating AI outreach tooling or building it into their own motion.

The Reading Job

Here is a task that is genuinely hard for humans and genuinely easy for a model: read ten years of an employer's public Form 5500 filings and surface anything unusual.

The output might be "participant count grew 46% across two filings while the same generalist P&C broker stayed on Schedule A, and the commission did not change." Nobody was going to do that by hand across 400 employers, and the finding is specific, verifiable, and non-obvious.

Same shape elsewhere. Read three months of someone's posts and extract what they keep returning to. Read a job posting and infer what motion they are building. Read an annual report and find the segment that grew. All bounded, all verifiable, all tedious.

This is where AI earns its cost in outbound, and it is the least-marketed capability in the category.

The Writing Job

Now the part that goes wrong. Ask a model to turn that finding into a first email and it produces something competent and structurally identical to every other AI-written opener: an observation about the company, a pivot, a soft pitch, a meeting request.

The prose is fine. The problem is that prospects have seen the structure hundreds of times, and recognising it is worse than receiving an obviously generic email, because it is a generic email that was pretending.

Writing the sentence yourself takes thirty seconds when the research is already done. That is the whole trade: let the model spend an hour reading so you can spend thirty seconds writing.

Where It Actively Costs You

Three failure modes worth knowing before you turn anything on.

Confident fabrication. A model asked to personalize will personalize, including for the 40% of prospects with no meaningful public signal. It will infer something plausible and wrong, and send it from your domain. The correct behaviour is falling back to a clean generic message, and most tools do not.

Deliverability drift. Thousands of senders producing near-identical structures from the same handful of models is a pattern filters can learn. This is a slow cost that shows up as declining placement rather than a single incident.

Complaint rate. Irrelevant mail gets reported, and Gmail's threshold is roughly one complaint per thousand delivered. Fabricated relevance is worse than no relevance here, because it prompts a reaction.

Keep a Human Gate on First Contact

The rule we run: the model can read anything and draft anything, and a human approves anything leaving the building for the first time.

Autonomous replies to known-shape inbound are a different risk profile and are fine, "can you send pricing," "who handles this," "we are in contract until March" all have correct answers a model produces instantly. That is what Unibox and Ezra do inside grobot: routine replies answered or drafted, anything unusual held for a person.

Autonomous first contact at scale is where the risk sits, and any vendor selling it is transferring that risk to you.

The Test Worth Running

Split 600 contacts three ways, 200 each. Arm A gets researched personalization built on a real trigger, written by a human. Arm B gets AI-generated first lines. Arm C gets a short, clean, plainly generic email with no personalization at all.

Measure positive reply rate, not opens. Arm A should win clearly. The result worth paying attention to is C against B, in our experience the plain generic email beats the AI-personalized one often enough to matter, which is an uncomfortable finding if you are paying for a personalization tool.

Run it before you buy, not after. Vendors will show you curated output; the variance across twenty consecutive generations is the actual product.

What the Realistic Gain Is

Not more pipeline from more sends. The ceiling is set by LinkedIn throttling and mailbox limits, and a model does not raise a rate limit.

What you get is research at a scale that was previously impossible, which lets you run trigger-based lists you could not have built by hand. That is a genuine and significant advantage, and it shows up as a better list rather than a bigger one.