The actual problem
People assume AI outreach fails because the writing is bad. The writing is usually fine. It fails because the model has nothing to write about.
Give a model a name, a job title and a company and ask for a personalised opener, and it will produce something grammatical and utterly empty: “I noticed you’re driving growth at Acme, and I imagine scaling the team brings its own challenges.” That sentence contains no information. It could be sent to ten thousand people. The prospect deletes it in under a second, and they are right to.
Give the same model the person’s last three posts, their role history, their company’s recent announcements and what their job ads say they are hiring for, and it produces something only sendable to that one person. Same model, same prompt structure, completely different result.
The variable is input, not prompt engineering.
What to automate
Research. Reading a profile, a company page, recent posts and job ads takes a human five to ten minutes per prospect. A model does it in seconds and does not get bored on the fortieth one. This is the single biggest time saving available.
Qualification. Deciding whether a prospect actually fits your ICP before you spend a message on them. Around a third of most raw lists never should have been messaged. This is what Prospectio’s ICP filter does before anything is sent.
The opening line. One sentence, built from what the research found. This is the highest-value sentence in the message and the one that most obviously cannot be templated.
Reply classification. Sorting inbound replies into interested, question, objection, not now and not interested, so the right ones get a fast human response.
What not to automate
Who to sell to. A model can score against an ICP you define. It cannot decide what your ICP should be. That comes from your last ten customers, and there is a guide for it.
The offer. What you are asking for and what you are promising is a business decision. Models will happily invent a value proposition, and it will be vague, because vague is the safest output when the model does not know your business.
Anything factual about your product. Pricing, integrations, capabilities. A model that guesses here will eventually guess wrong in front of a prospect.
The relationship after the reply. Automate the first response for speed, then get a human in.
A prompt that works
The shape matters more than the wording:
Here is a prospect's LinkedIn profile, their last three posts,
their company's about page, and two of their current job ads.
Write ONE sentence I can open a connection request with.
It must reference something specific from the material above
that could not apply to anyone else in their role.
No compliments. No questions. Under 25 words.
If nothing in the material is specific enough
to build that sentence, say "no hook found" instead.The last line is the important one. Giving the model permission to fail stops it inventing a hook, and “no hook found” is a signal that the prospect probably should not be on your list.
The vocabulary to strip
Every model overuses the same words, and B2B buyers have learned them: leverage, streamline, elevate, unlock, seamless, robust, delve, landscape, testament, crucial, in today’s fast-paced world. Ban them outright in your prompt, then read the output and cut anything that still sounds like a brochure.
Also strip: “I hope this finds you well”, “I wanted to reach out”, “quick question” when it is not, and any sentence that begins with “As a {job title}”.
The part AI cannot fix
If your offer is not interesting to the person you are messaging, no amount of writing quality will save it. Well-written outreach to the wrong person is still outreach to the wrong person. That is why the qualification step matters more than the copy step, and why we built the product in that order.
Where the real lift is
In our campaign data across 389,890 prospects and 15 controlled tests, moving from templated text to AI-researched personalised text roughly doubles reply rates. Moving from text to a personalised voice note or video took reply rate from 28.8 percent to 40.4 percent and meetings up 57 percent in a controlled test. AI writing is worth doing. Changing the format is worth more.