White paper · B2B outreach only
We gave one AI writer the same prospect and the same offer two ways: once on its own, once inside our engine. Then independent judges decided. This is what happened, how we tested it, and what we learned.
Scope. This study covers application 01, B2B outreach. Three prospects, one offer, AI judges. It does not test the engine’s other uses. See all four applications.
In one paragraph
The engine writes cold outreach the hard way. Before it writes a word, it assembles a verified evidence file on each prospect, confirms the material is actually about that person and not someone who shares their name, and works out what the person is genuinely under pressure about. The fair question from any buyer is whether that machinery beats simply handing the same job to a capable AI. We tested it head to head, with the same AI writer on both sides, judged blind by two independent AI systems from different companies. The engine won every evaluation. More important than the score is the reason it won, which was the same every time and is one anyone can inspect: the engine tells the reader who the offer is wrong for, out loud, and that honesty is what makes a skeptical executive believe the rest.
01 · The question
Every serious buyer of an AI marketing tool asks the same thing, and they are right to: couldn't I just ask ChatGPT to look this person up and write the emails? If the answer is no, it has to be demonstrated, not asserted.
The only honest way to demonstrate it is to hold the writer constant. If we compared our engine's output against a weaker model, we would be measuring the models, not the machine around them. So we used a single, capable writing model both ways:
On its own · given the prospect's name, title, and employer, and the offer, with the instruction to write the best outreach it could. This is the honest stand-in for "just ask the AI."
Inside the engine · the same model, writing from the engine's verified evidence file, its read of the person, and its plan for how to reach them.
Same writer, same prospect, same offer. The only variable is the machine. Throughout this paper, the seller is described generically as a venture-stage financial company offering a longevity benefit to large employers, and the executives who were written to are identified only by role and industry. See the note on anonymization at the end.
02 · The system under test
The engine is named for James Boswell, whose complete record of another man's life is the reason we still have the founding sentence of modern advertising. The premise is the same: the complete, verified record of what a person has said in public is the raw material of persuasion. Four things happen before a word is written.
The engine gathers what the prospect has said and written in public, and what others have said about them. Then it does the step most tools skip: it confirms the material is actually that person. Common names return strangers, and a file quietly filled with a namesake's words produces confident outreach to someone who does not exist. In this project the engine caught exactly that, and declined to write to a prospect whose public record turned out to belong mostly to other people sharing her name. The unaided model, given the same name, wrote to the stranger without hesitation.
From the verified file, the engine works out what the person is under pressure about and what they already care about, each finding tied back to a specific source. This is the difference between knowing how someone writes and knowing what would move them, and it is the part a generic "look them up and write" prompt never reaches.
A senior financial decision-maker weighing a considered, hard-to-reverse commitment reads nothing like a consumer browsing a lifestyle product. The engine adapts accordingly. As the results below show, this single adaptation is worth the entire margin of victory.
A leading AI writing model produces the copy, following an explicit copywriting method. A separate critic model, from a different company, then reads it against that method and sends back specific fixes. The writer never grades its own work.
03 · The experiment
For each prospect we produced two sets of outreach: the unaided version and the engine version. A third AI, from a company different from the writer, judged the two sets blind. Which version was labeled A and which B was randomized per prospect by a content hash, so neither position nor origin could sway the judge. It scored both on five dimensions:
Specificity · is it grounded in true, particular things about this person, or could it have gone to anyone with the same title?
Resonance · does it connect to something the reader already cares about?
Credibility · would a skeptical senior executive believe it?
Restraint · does it respect the reader and avoid hype and servility?
Reply-likelihood · all told, would a busy person actually respond?
To guard against a single judge's idiosyncrasies, the whole comparison was then re-run with a second judge from a third company. Neither judge was from the writer's company, so at no point did a model grade its own house style. Three executives were selected for the depth and cleanliness of their public records: a combined CFO and head of HR at a large private agribusiness, a workforce-strategy leader at a global software company, and a chief people officer at a global real-estate firm.
04 · The protocol
The honest part of this study is that the engine did not win at first. Our method was a loop: generate, judge blind, and when the engine lost, find the mechanical reason from an established body of persuasion research rather than fiddling with wording, fix that cause, and test again. Each fix was predicted to work before it was measured. The record of that loop is the real evidence, because a system that can be diagnosed and corrected is a system that is understood.
| Round | What changed | Result |
|---|---|---|
| 1 | First engine build | Engine loses, slight |
| 2 | Added a full persuasion-mechanics layer | Engine loses, clear |
| 3 | Matched the register to a high-scrutiny reader | Engine wins, clear |
| 4 | Ran the full three-prospect sweep | 2 of 3 won |
| 5 | Added "heavy emotion, never personal" rule | 2 of 3 (a different one flipped) |
| 6 | Reconciled the two register rules | 3 of 3 won, clear |
Round 2 is the instructive one. Adding more persuasion technique made the copy measurably more specific and more resonant, and it lost by more. That failure was predicted: the reader was a senior financial executive, and the research on high-scrutiny audiences says plainly that they punish copy that visibly performs emotion. The fix in round 3 was not better words. It was the right register.
05 · Results
With the engine converged, both judges scored all three prospects. Scores are the sum of the five dimensions, five points each, so fifteen is a perfect card and any number above the control's is a win.
| Prospect (by role) | Judge A (Company 1) | Judge B (Company 2) |
|---|---|---|
| Agribusiness CFO & HR head | 25 vs 16 clear | 25 vs 13 decisive |
| Software workforce lead | 22 vs 12 clear | 25 vs 8 decisive |
| Real-estate people chief | 19 vs 13 clear | 25 vs 8 decisive |
The engine won all six. The second judge, from a different company, was not only in agreement but more emphatic than the first, scoring the unaided model lower and calling every engine win decisive. Whatever "good" means to these two systems, it is not one company's house style, and both named the same reason for the result.
The engine version openly admits the product is wrong for much of the company's frontline workforce, which is exactly the kind of uncomfortable truth a real tailored note includes and a template avoids.Judge B, on the winning copy
A win could still be luck, or simply the engine writing nicer sentences. So we ran one more test. We took the winning prospect and forced the engine to use the wrong register, the performing, emotional one meant for a low-scrutiny reader, changing nothing else. Same prospect, same judge, same control.
Correct register: engine 25, control 16 → engine wins, clear
Wrong register: engine 17, control 19 → engine loses, slight
An eight-point swing from changing one variable. The engine's advantage is not that it writes prettier copy; it is that it reads the person and chooses how to speak to them. Remove that judgment and the advantage disappears.
06 · What we learned
The value of the experiment is not the scoreboard. It is three lessons, each earned by a loss and confirmed by a fix.
Early on, the engine's copy was more specific and more resonant than the unaided model's and still lost. It had learned a great deal about each person and spent it too early, as a wall of researched facts, before earning the reader's belief. Research is an asset only when it arrives in the right order. Deployed too soon, it reads as bragging.
A senior decision-maker weighing a serious financial commitment does not want to be moved; they want to be convinced it is real and governable. For that reader, a clear mechanism is the emotional experience, and any visible attempt to make them feel something makes them audit you instead. This is not a stylistic preference. Removing this single adaptation cost the engine a clear win and turned it into a loss.
Difficult feelings are the engine of persuasion, the tension that creates the pull toward resolution. The engine may use them freely, with one hard limit: the difficult feeling must never be personal to the reader. One prospect's copy opened on her own company's recent layoffs, a real and heavy tension, and it lost, because a hard fact about the reader's own world makes them defensive rather than moved. Moved to a tension about the situation instead of about her, the same copy won.
07 · A worked example
Here is the opening of the outreach to the agribusiness CFO, in the engine's first draft and its final one. Identifying details have been altered for anonymity; the persuasion structure is unchanged.
"You invested a nine-figure sum in a rural company town: clinics, mental-health support, free care for a community most employers would overlook. That standard is why I am writing you and not a hundred other CFOs.
Our fund is a longevity benefit offered through payroll. The employer pays nothing, matches nothing, screens nobody, and can be live in a day…"
"Our fund locks contributions until age 80. No early access, no exceptions. For a working-class employee who needs cash this year, that is the wrong product, and I will not pretend otherwise.
Which leaves a fair question: why write to you at all. Because I am not offering this to everyone…"
Both were written by the same AI model. The difference is entirely in what the engine told it to do: puncture the reader's assumption instead of flattering them, concede the limit before making the claim, and speak to a serious buyer in the register a serious buyer respects.
08 · What we are careful to claim
A result is only as trustworthy as its stated boundaries. These are ours.
Small sample. Three prospects, one offer, one seller. This is a strong, consistent signal, not a statistical proof. A larger study across more offers and industries is the natural next step.
AI judges, not humans yet. Two independent AI systems agreed decisively, which is meaningful, but the judges' tastes may themselves favor restraint. Human evaluation by experienced buyers is the necessary next measurement.
The baseline is the honest one, and it is not a human expert. We compared against the same AI writing unaided, which is the right test for "does the machinery help." We did not compare against a seasoned human copywriter with full research time.
Run-to-run variation is real. AI writers do not produce the identical draft twice. Individual scores moved across rounds; the direction and consistency of the result, across two judges and three prospects, is what carries weight.
What the study does establish is narrow and real: given the same capable writer, the engine's evidence, judgment, and register-matching produce outreach that two independent judges prefer, decisively and for a reason that can be inspected and reproduced.
09 · The bottom line
"Just ask the AI" produces competent, generic outreach that a senior reader recognizes instantly as a template. The engine, using the same AI, produces outreach that names the reader's real situation, concedes what the offer cannot do, and speaks in the register that reader respects. Two independent judges preferred it every time.
Underneath the copy is the part that compounds: the engine writes only from a verified record, and it refuses to write when it cannot confirm who the person is. In this study it declined a prospect whose public trail belonged to strangers who shared her name, while the unaided model wrote to them confidently. In outreach, the letter you do not send to the wrong person is worth as much as the one you send to the right one.
The engine will keep improving: broader evidence, deeper models of how each person thinks, and human evaluation to sit alongside the machine judges. But the core question this study set out to answer is answered. The machinery is not decoration on a prompt. It is the difference between being read and being deleted.