Personalisation at segment-of-one
Decide between a few segments and one message per person by pricing the lift against the cost of making and checking every variant.
On this page
Generation made one message per person cheap: 80,000 personalised emails cost well under $1,000 of model time. Whether they beat a handful of well-chosen segments is a sum, and the sum has a term the production price hides: checking, the cost that grows with every variant.
The sum, and the usual mistake
The proposal usually arrives as "one email per person for $800, and personal beats generic". Both halves flatter it. The $800 leaves out checking, and "beats generic" picks the wrong rival: the alternative worth beating is a few good segments, which capture most of the gain for a fraction of the cost.
Written out, one-per-person wins when:
audience x segment conversion rate x extra lift x value per conversion
> variants x generation cost + checking + setup - cost of the segments
Every term except the lift can be priced in advance; the lift is an assumption that, at most audience sizes, nobody can confirm.
Worked: 8 segments against 80,000 variants
A self-serve product has 80,000 free users and sends them a monthly upgrade email. At the course's email rates from the arithmetic a marketer carries, 40% open, 5% of opens click and 10% of clicks convert, so 0.2% of recipients become paying customers: 160 a send. Each is worth the default customer's $800 of expected margin.
One message. Generated and reviewed for $51. This is the baseline.
8 segments, cut by plan and by the last feature used. Eight variants at $51 each is $408, and defining the segments takes four hours at $100, so $808 in all. Assume the segments lift conversion 10%: 16 more customers, $12,800, net about $12,000 a send.
One per person. 80,000 variants at about $0.01 each is $800, built from the same segment data plus each user's history, with the same $400 of setup. Assume a further 3% on top of the segments: 176 × 3% ≈ 5.3 more customers, 21 in all over one message, $17,000. Then checking, priced in the next section at $5,000 a send. Cost $6,200, net about $10,800.
The 8 segments win by about $1,200 a send. Moving from 8 to 80,000 variants adds about 5.3 customers, $4,200, and about $5,400 of cost. One-per-person breaks even when its extra lift reaches ($6,200 − $808) ÷ $800 ≈ 6.7 customers, about 3.8% on top of the segments. That is close to the assumed 3%, which is the uncomfortable part: the decision turns on whether the truth is 3% or 4%, and the next two sections show why nobody can afford to check.
That segments take most of the lift is an assumption with a reason behind it: most of the difference in how people respond to an upgrade email follows a few attributes, such as plan, tenure and the feature last used. How much individual history adds, in your product, is whatever you measure.
Checking is the cost that scales
With 8 variants, reading 8 emails checks everything any recipient will see. Nobody reads 80,000: at 30 seconds each that is 667 hours a send. A random sample only bounds the error rate. If none of n emails reviewed is wrong, the 95% upper bound on the true error rate is roughly 3 ÷ n. To show that fewer than 1 in 1,000 emails are wrong, 80 people a send, a reviewer reads 3 ÷ 0.001 = 3,000 emails, at a minute each 50 hours, $5,000. Even then up to 80 people a month may receive a wrong price, a feature they do not have, or a claim the product cannot support.
That sample is an evaluation set in all but name, and evals as the unit test of LLM work applies: fixed criteria, a grader, a run on every prompt change. An automated check calibrated against a reviewer's labels could do much of the reading and change the sum; until it is calibrated, the $5,000 stands.
Three ways it fails
Brand drift. Each variant is a fresh draw from the model, and across 80,000 draws the voice wanders and claims creep in that nobody approved. Eight variants can be held to the standard by reading them; 80,000 only statistically, by the sampling and monthly reference set of brand stewardship.
Untestable output. Each variant meets one reader, so only the policy can be tested: the generator against the segments, recipients split at random. At the course's test settings, worked in experiments: holdouts, geo tests and power, detecting the segments' 10% lift on the 2% click rate takes about 81,000 recipients an arm, two sends of this list. Detecting the further 3% on the same click rate takes about 867,000 an arm, and on the 0.2% conversion rate about 8.8 million. At 40,000 an arm a send, the click lift needs about 22 sends and the conversion lift is out of reach. The vendor's figure is believed, never checked.
The reader noticing. Personalisation that shows what was collected, such as "we saw you looked at pricing three times last night", reads as surveillance. Measure unsubscribes and complaints per send against the segment version, and keep to what users expect their data to be used for, the rule in ethics and disclosure.
Where one per person does pay
The sum favours individual messages when the audience is small, each conversion is large, and a person reads every message anyway, so checking is already paid for. A two-person sales team writes to 200 named target accounts, each worth $30,000 of margin if won. Assume 5% book a meeting and 20% of meetings close: 10 meetings and 2 deals, $60,000. A generated draft built from each account's own filings and job posts, edited by the rep in 5 minutes, lifts meetings 20%: 2 more meetings, 0.4 more deals, about $12,000. The cost is 200 × 5 minutes ≈ 17 hours, about $1,250 at a rep's loaded $75 an hour, plus about $2 of generation. Net about $10,750, with every message read by a person before it goes.
The difference is the ratio of value to variants: $30,000 a conversion across 200 variants, against $800 across 80,000. Where it is large, one per person is a sales-assist tool worth building; where it is small, segments win, and building them is segmentation applied to users instead of markets.
In practice
To a vendor selling one-to-one generation: "What lift over our best segmentation, not over one generic message, and how would we confirm it at our list size?" A lift quoted against a single generic message credits the personalisation with the segments' gain as well. Ask for the checking plan too: what share is reviewed, and what error rate it bounds.
To the team: build the segments first and measure them, which takes two sends. Revisit one-per-person when an automated check has been calibrated, or for the audiences where a person reads every message anyway. Creative testing at volume covers the same problem from the other side: many variants, each with too little traffic to judge.
The mistake that costs money is pricing one-per-person at generation cost and comparing it with one generic message. $800 against a generic send says obvious yes; against eight segments, with $5,000 of checking counted, it loses $1,200 a send on an extra lift no one can measure.
Check yourself
1. A vendor offers personalised onboarding emails for your 20,000 new sign-ups a month at $0.02 each and claims activation rises from 30% to 34.5%, a 15% lift. Do you sign a year's contract?
Not before a test, and here a test is cheap. At a 30% base, a 15% lift needs about 1,700 users an arm at the course's settings, so a 50/50 split fills in under a week and reads once the seven-day activation window closes. A high base rate makes the claim checkable: sign only if the lift holds against your segmented onboarding, not a generic one.
2. In the worked case, the brand lead agrees to tolerate 1 wrong email in 500 instead of 1 in 1,000. Does one per person now win?
The sample falls to 3 ÷ 0.002 = 1,500 emails, 25 hours, $2,500. Cost becomes $800 + $400 + $2,500 = $3,700 and net about $17,000 − $3,700 = $13,300, against $12,000 for segments. On paper it wins by $1,300, but the gain is bought by accepting up to 160 wrong emails a send and rests on a 3% lift that would take years of sends to confirm. Keep the segments; the arithmetic flipped on a tolerance, not on evidence.
3. A sales team wants generated, individually researched emails for 50 strategic accounts, each worth $100,000 of margin, with every email edited by a rep. Approve or decline?
Approve. If a quarter of meetings close, one extra meeting is worth a quarter of $100,000, $25,000 of expected margin. The cost is 50 × 5 minutes ≈ 4 hours of rep time and cents of generation, and checking is complete because a person reads every email.