The First-Touch Personalization Standard: Send, Rewrite, or Hold

This is a pass/fail standard for the claims a personalized first touch makes about the person who receives it. It is for founders doing their own outbound, account executives, SDR managers, and anyone who approves an AI SDR's drafts before they go out. It gives you six criteria, a Send, Rewrite, or Hold verdict that two reviewers should reach on their own, and a weekly sample audit that measures how often drafts fail.

What this standard grades, and what it leaves alone

This standard grades the factual claims a first-touch message makes about its recipient and their company, and nothing else. It does not judge tone, offer, subject line, or sequence design. A message can pass here and still be a weak email.

Three terms, defined once:

• First touch: the first message a named prospect gets from you, whether a rep wrote it or a tool drafted it.
• Claim: any statement about the recipient or their account that could be true or false. Examples: a title, a funding round, a hire, a post, a launch, a tool they use.
• AI SDR: software that researches a prospect and drafts or sends the outbound message for a rep.

You review claim by claim, but the verdict covers the whole email. A draft with five correct claims and one wrong one is a wrong email. That asymmetry is why this standard exists. In one study where two reviewers graded a language model's output, the hallucination rate was 2.1% per claim but 8.7% per response, because a long answer has many chances to be wrong somewhere. Outbound drafts work the same way. No primary source has published a measured fabrication rate for outbound drafts, so read those figures as a reason to audit, not a forecast of what you will find.

How much personalization actually moves replies

Less than vendor pages suggest. Randomised trials show a personal touch raises the odds of a response by about a quarter. Observational sales data shows the effect shrinking as seniority rises. The big vendor multiples come with no control groups.

An odds ratio (OR) compares the odds of a response with the treatment to the odds without it. An OR of 1.24 means the odds rose by about a quarter. These trials measured survey response, not sales replies. They don't forecast your reply rate, but they show how big a believable effect is. The vendor's "5x improvement" claim gives no control group and no sample definition. One audit of popular statistics traced a widely shared "17 to 18% versus 7 to 9%" personalization figure and found no source document at all. The likeliest explanation for big vendor multiples is selection: lists built from triggers are warmer before anyone writes a word.

Two findings change how you grade:

• Seniority changes what counts as relevant. Gong Labs found that one-to-one personalization more than doubles replies from non-managerial buyers but lifts director-and-above replies by only 50%. A company or account-based angle can triple executive replies. The report says it bluntly: "Executives don't care if you reference their alma mater." Gong also reports that C-level executives are 30.2% less likely than non-executives to reply to cold email. For senior recipients, the connective criterion below carries most of the weight.
• The baseline is low, so the downside dominates. Gong's engagement benchmarks put manual email at a 2.1% reply rate and automated email at 1.1%. A correct personal line lifts a small number a little. A wrong one sent to a whole segment can damage your sending domain (see the failure modes).

The six criteria every claim must pass

A first touch is ready to send only when every claim about the recipient meets all six criteria below. One failing claim fails the message. Whether the verdict is Rewrite or Hold depends on whether the fix needs new research.

What each criterion proves:

1. The claim exists outside the draft.
2. It is still recent enough to mention.
3. It is about this person.
4. A reasonable recipient would not object to you knowing it.
5. The detail earns its place in the email.
6. The generator left no traces behind.

The verdict rules

• Send: every claim passes all six criteria.
• Rewrite: identity is confirmed, and every failure can be fixed by deleting or rewording with material that is already verified. Typical causes are a stale trigger you can drop, a generic connective sentence, or a leftover placeholder.
• Hold: the fix needs new research, or a call the reviewer shouldn't make alone. Typical causes are an unconfirmed identity, a claim the primary source contradicts (so the underlying record is suspect), or a private detail that has to come out of the research record as well as the email.

The tiebreak is mechanical. If the reviewer can fix the draft using only what is already on the grading sheet, the verdict is Rewrite. If they would need to open a new tab and go looking, it is Hold. That rule is what makes two graders land on the same answer.

Freshness windows by trigger type

Nobody has published, for any trigger, the age at which mentioning it stops lifting replies. The windows below are practitioner convention. Adopt them as team policy, write the numbers down, and always date the event rather than the article.

How to apply them:

• Date the event. Use the date the round closed, the person started, or the post went up, taken from the first publication. A republished article can make a six-month-old round look like news.
• Where sources disagree, take the stricter window. One practitioner source says the job-change timing advantage is gone after week 12. Another allows 90 days for senior executives. Pick one, write it down, and grade against it.
• The 48-hour figure is about speed, not freshness. It tells a team how fast to act once it spots a trigger. It does not make a claim fresh if the event itself is old.
• For posts and speaking slots, set your own window and call it a guess. No public window exists. Write your choice on the policy page with the date you made it, and check it against your own reply data later.

For job changes, the hard part is getting a reliable start date before drafting. A search that filters on when someone started the role, such as asking Refolk for people who started their role in the last 90 days, puts the event date on the list before anyone writes a line.

Freshness also predicts where AI drafts go wrong. General language-model research finds that hallucinations cluster on rare facts about specific people and small companies. It also finds that topics where facts change often hallucinate at three to four times the base rate. Funding rounds and job titles are exactly that kind of fact. The classic outbound example is "Congrats on the Series B" sent to a company that raised a Series A. Expect your failures to cluster on criteria 1 and 2 for those two trigger types, and sample with that in mind.

How strong the person-match must be

The source must match the recipient record on name, current employer, and current title. A name alone, or a name plus a title keyword, means Hold, because matching on title keywords is noisier than it looks.

If a title search over clean, structured profiles gets about one in six wrong, a draft built from a scraped list and a name lookup will get more wrong. The same sampling in Refolk's index found 96% in-title precision for US SDR managers but 80% for the UK equivalent, so precision also varies by market. Nobody has publicly documented a person-match threshold for outbound. The rule here is my proposed standard, and the sample above is why it is strict.

Confirm identity before you source any claim. Public sources don't settle the order, but the logic is clear. Sourcing three claims about the wrong person wastes the work. And a reviewer who has already found a matching funding announcement is primed to accept a shaky identity. If the recipient record came from a keyword list, look the person up again by name and company in Refolk and compare their current employer and title with the record before you grade anything else.

A confirmed match looks like this:

• The name matches, allowing for nicknames and transliteration.
• The current employer matches in both record and source, under the same legal or trading name.
• The current title matches in function and level, not just by keyword.
• If the email mentions the role itself, the source showing all three is dated inside the job-change window.

The grading procedure for one draft

Run these eight steps in order on every draft in the review queue. Pulling out the claims takes about a minute, and sourcing takes one to three minutes per claim. The output is a numbered claim log and one verdict.

Once identity is confirmed, two things decide between Send, Rewrite, and Hold: how good the evidence is and how specific the tie to your reason for writing is. Senior recipients raise the bar on the vertical axis, because for executives a true detail with no link to their mandate is close to worthless.

FIRST-TOUCH GRADING SHEET
Draft ID:
Recipient on record (name / employer / title):
Reviewer:                 Date graded:

IDENTITY (criterion 3)
Source URL:               Source date:
Name match: Y/N    Employer match: Y/N    Title match: Y/N
Identity: CONFIRMED / HOLD (stop here if HOLD)

CLAIMS (criteria 1, 2, 4)
#  | Claim as written | Primary URL | Says exactly this? Y/N | Event date | Window | Inside? Y/N | Professional? Y/N
1  |
2  |
3  |

CONNECTIVE SENTENCE (criterion 5)
Sentence:
Still true with another prospect's name? Y/N   (Y = FAIL)

ARTIFACT SWEEP (criterion 6)
Placeholders / merged-field errors / wrong names / unlisted specifics:

VERDICT: SEND / REWRITE / HOLD
Failing criteria:
Fix needs new research? Y/N   (Y = HOLD)

How this goes wrong: failure modes and false positives

Most failures are false passes: a claim looks verified, fresh, or relevant because the reviewer checked the wrong thing. Each failure mode below describes what the false pass looks like and the check that catches it.

Verified against an aggregator, not the primary source

The false pass is a confident, plausible fact that is slightly off, such as Series B for a company that raised a Series A. Aggregators restate, round off, and lag behind. Check: criterion 1 passes only on the press release, the company's own announcement, or a filing.

Source date read as event date

A republished or updated article carries a recent date while the round it covers is six months old. The reviewer sees a fresh page and ticks criterion 2. Check: find when the event was first published and grade that date against the window.

Identity merge

Two people share a name, or one person's old job gets merged with their new one. Keyword matching is noisy: in Refolk's index, 4 of 25 sampled US "SDR" profiles were not SDR roles, including a CEO and a "Sales Agent". Check: employer and title must both match, not just the name.

Private detail passes because it is public

The reviewer reasons that anything findable is fair game. The ICO says that using publicly available personal data to market to someone, even in a business context, must comply with UK GDPR. Its legitimate-interests guidance accepts uses that are proportionate, have minimal privacy impact, and are not something people "would be surprised by or likely to object to". No regulator publishes a list of private versus professional detail types, so apply that test directly. One documented example that fails it: "How was your trip last week away with your girlfriend?" Check: strike anything about family, relationships, health, home, or personal travel. Hold the draft if the research record itself contains that detail. This is a grading test, not legal advice; your lawful basis is a separate question.

Connective sentence present but generic

A compliment followed by a pivot is a pattern recipients spot instantly. The sentence exists, so a hurried reviewer ticks criterion 5. Check: the substitution test. A sentence that names the recipient's problem as a result of the specific detail passes. The 30MPC report lays out the shape to look for: Personalization > Problem > Solution > CTA.

If a sentence is still true with another prospect's name in it, it is decoration, not personalization.

Fabricated detail that reads fluently

Placeholders are easy to catch. Invented specifics are hard, because they read fluently and sound like research. One practitioner blog claims that 85-95% of "personalized" output from five tools was templates with 3-5 fields swapped, and that the research-based rest hallucinated about 15% of the time. It links no primary study, so treat those figures as unverified. Check: every specific noun in the draft is either on the claim list with a URL or deleted.

Kappa looks high, or undefined, because nearly everything passes

When almost every draft passes, two graders agree most of the time by default. The statistics literature notes that high percent agreement can sit alongside a low or undefined kappa when base rates are extreme. Check: report percent agreement next to kappa, and add known-fail drafts to the calibration sample.

The complaint signal arrives late

Google's sender guidelines say to keep the spam rate reported in Postmaster Tools below 0.10% and never reach 0.30%. At 0.30%, three complaints per thousand messages cross the line. A secondary source says a sender over the line stays ineligible for mitigation until the rate holds below it for 7 consecutive days. Because the rate is calculated daily, a week of bad sends shows up only after the damage is done. Check: treat the pre-send gate as the control, and complaint data only as confirmation.

The weekly calibration and sample audit

Once a week, two graders score the same sample of drafts independently. Then you compute percent agreement, Cohen's kappa, and the failure rate by criterion. Agreement tells you the standard is applied consistently. The failure rate tells you how good the drafts are.

Cohen's kappa measures agreement between two graders after correcting for the agreement chance alone would produce. A score of 1 is perfect agreement and 0 is no better than chance. On the standard Landis and Koch bands, 0.61-0.80 is substantial and 0.81-1.00 is almost perfect. For reference, human screeners in systematic reviews average 0.82 on abstract screening and 0.77 on full text, and those are trained people reading structured documents. Treat 0.61 as your floor and aim for the high 0.7s. If you land below the floor, the written criteria are ambiguous. Fix the wording before you retrain the graders.

How big the sample needs to be

No outbound review program has published a weekly sample size, so these are statistical floors, not benchmarks:

• Failure rate: estimating it within 10 points either way at 95% confidence takes about 96 graded drafts. That comes from the standard proportion formula at its worst case, where the true rate is 50%. The number is derived, not measured.
• Agreement: one power analysis needed an effective sample of about 173 items to test kappa against a threshold. Pool double-graded drafts across weeks until you reach it, rather than double-grading 173 every week.

Who has the time to grade

Review capacity is the real constraint, and Refolk's index shows its scale.

A manager with 17 reps can't personally read every first touch, and grading 96 drafts a week is real work. The time is usually already being spent somewhere else: a Gong survey of 600+ sellers and analysis of 30,000+ emails found reps spend an average of 12 hours a week personalizing. Moving part of that time from writing personal lines to checking them is more realistic than adding manager hours.

Report failures by criterion, not just in total, because the fixes differ. One industry survey of AI SDR rollouts found generic or robotic output (37.5%) and poor data quality (31.3%) were the most common problems. Generic output shows up as criterion 5 failures and points to the prompt or template. Poor data shows up in criteria 1 to 3 and points to the list source.

WEEKLY FIRST-TOUCH AUDIT
Week of:                  Graders: A = ____   B = ____
Drafts in queue this week:
Drafts sampled:           Known-fail drafts seeded:

VERDICTS (grader A)
Send: __   Rewrite: __   Hold: __   Failure rate (Rewrite + Hold / sampled): __%

FAILURES BY CRITERION (count of drafts)
1 Source: __  2 Freshness: __  3 Identity: __  4 Private: __  5 Connective: __  6 Artifact: __

AGREEMENT (double-graded subset)
Drafts double-graded: __   Cumulative toward 173: __
Percent agreement: __%     Cohen's kappa: __
Seeded fails caught: A __ of __   B __ of __

DISAGREEMENTS RECONCILED
Draft ID | A verdict | B verdict | Criterion at issue | Wording change made

ACTIONS
Prompt or template change:
List source change:
Policy page edits (with date):

Before you call a draft done, and keeping the standard current

A draft is done when every item below is true and logged. The standard stays current only if you re-check the external thresholds it depends on and update the windows from your own data.

What to re-check, and how:

• Mailbox-provider thresholds. Read Google's email sender guidelines directly and confirm the current spam-rate lines and how they are enforced. A secondary source reports that Gmail has moved from warnings to outright rejection of non-compliant bulk mail, so enforcement changes over time. Write the date you last checked on the policy page.
• Privacy guidance. The ICO revised its legitimate-interests guidance after the Data (Use and Access) Act. Re-read its business-to-business marketing and legitimate-interests pages whenever your team adds a new research source or detail type.
• Freshness windows. Every window in this standard is convention. Once you have a few months of audited sends, compare reply rates for claims early and late in each window. Shorten any window your own data says is too generous, and record the change with its date.
• The criteria wording. Every disagreement you reconcile in the weekly loop is a candidate edit. A standard that two graders apply identically is worth more than one that sounds complete.