AI now rewrites each nurture email for its reader

KEY TAKEAWAYS

Nurture Agent's report compares opens and clicks only. Both move for reasons other than buyer intent.
In Mercari's test of over a million users, AI titles lifted taps 24% while overall buyer conversion did not move.
Split contacts at random when they enter the workflow, so each person sees one version for the whole sequence.
Price the agent per incremental meeting. At five emails and $0.10 each, a 0.5 point lift costs $100 a meeting.

HubSpot's Nurture Agent is now in public beta. HubSpot updated its help page on 16 September 2026, and MarTech's 16 October recap lists the agent as a public beta. It rewrites each automated email for the person about to receive it. Subject line, preview text, body and call to action can all change, drawn from CRM records and live web research. The question for a funnel team is not whether the copy reads better. It is whether more of the right people book a meeting, and what each extra meeting costs.

What exactly does the agent change?

HubSpot's help page, last updated 16 September, sets out the mechanics.

It works on published automated emails only. Transactional emails are excluded.
You choose the inputs: contact properties, other object types, activity history.
It rewrites subject lines, preview text, CTA text and body copy. Any module can be locked.
A slider sets the share of recipients who get the rewritten version. The default is 50%. The rest get the original.
An account can run it up to 20,000 times a month.
The sent version of each email stays visible on the contact record for 30 days.

HubSpot's product catalog prices it at 10 credits per personalized email per contact. At the pay-as-you-go rate of $0.010 a credit, that is $0.10 an email.

There is no approval step per email. A HubSpot partner writing in September put it plainly: control happens before sending, through brand voice, data quality and the percentage, not message by message. The same partner warned that contacts with little more than a name and address get something close to the original while still using credits.

Why will the built-in report make it look good?

The Nurture Agent performance report compares the two versions on four numbers: deliveries, open rate, click rate and click-through rate. Each has a problem for this particular test.

Opens. The agent rewrites the subject line. Any new subject line will open at a different rate, so an open lift says the subject changed, which you already knew. Opens are also partly machine-made. Apple states that Mail Privacy Protection downloads remote content in the background whether or not the reader engages. Litmus's September 2026 data, from over 1.5 billion opens, puts Apple at 76.56% of tracked opens and cites studies placing privacy-affected opens at roughly 55% to 60%.

Clicks. Security scanners click links in corporate mail. Their clicks land on both versions, which shrinks real differences and adds noise. We worked through the size of that effect in our piece on bot clicks in email.

The split itself. The slider works per email. HubSpot's help page does not say whether a contact who got the rewritten first email also gets rewritten later ones. In a five-email sequence, that matters. If people move between versions, nobody experiences a clean "personalized sequence" and the meeting it is meant to produce cannot be credited to either side.

None of the four numbers is a meeting, a reply that matters or pipeline.

What does controlled evidence say about AI-written email?

Vendor figures are easy to find and hard to trace. Stripo's Q3 2026 report lists a 41% click lift from AI personalization, and labels its own figures self-reported, directional and not from controlled trials. McKinsey's often-quoted 10% to 15% revenue lift from personalization appears without a sample or method. HubSpot's Fall 2026 Spotlight gives no performance figure for Nurture Agent at all. MarTech's 16 October recap reports that HubSpot's own testing showed the strongest lift on nurture emails that already perform fairly well, with no numbers attached.

Three pieces of research hold up better.

Study	Design	What moved	What did not
Sahni, Wheeler and Chintagunta, Marketing Science, 2018	Randomized field experiments, three companies, millions of emails; recipient's name added to the subject line	Opens 9.05% to 10.80%; leads 0.39% to 0.51%; unsubscribes 1.2% to 1.0%	All three reported measures moved
Mercari, arXiv, 2025	Over one million lapsed users split into equal groups for one week; GPT-4o-mini titles against template titles	Item taps up 23.63%, significant	Opens up 0.46%, not significant; overall buyer conversion down 0.42%, not significant
Matz and others, Scientific Reports, 2024	Four studies, 1,788 people; ChatGPT ads tailored to personality	20 of 33 tailored messages beat generic ones on stated ratings or willingness to pay	No behavioural outcome was measured

Read together, they say three things. Copy changes do move engagement. Engagement gains do not reliably carry through to buying: Mercari's purchases through the email rose 45%, not significant, while overall purchases did not move, which suggests the email took credit for buyers who would have come anyway. And the one study with a clear downstream gain used a simple name token and millions of recipients to see it.

Can personalization backfire?

Some surveys say yes. Gartner surveyed 1,464 B2B buyers and consumers in late 2024. 53% reported a negative experience with personalized marketing, and those customers were 3.2 times as likely to regret a purchase. Validity's June 2026 survey of 1,000 US consumers found 40% would trust a retailer's emails less if they knew AI wrote them.

These are stated attitudes, not measured behaviour, and the Validity figure concerns retailers rather than B2B nurture. They do not cancel the field results. They do mean the sign of the effect on your list is an open question. Only a test with a true control answers it.

How do you set up a test that measures meetings?

Use the workflow, not the slider. HubSpot's workflow help page, updated 23 September 2026, describes a random split branch that assigns enrolled records at random and needs about 1,000 records for a roughly even split.

Clone the sequence. Keep every email, delay and condition identical in both copies.
Add a random split at enrollment. Branch A gets the copy with Nurture Agent set to 100% on every email. Branch B gets the untouched copy. Each contact now sees one version for the whole sequence.
Store the branch on the contact. A property such as "nurture test arm" lets you report on it in the CRM long after the 30-day view of sent versions expires.
Pick one outcome and one window before you start. For example, a meeting booked within 60 days of enrollment. Replies and opportunities can be secondary.
Size the test with the table below and run it to that size. Do not stop when the open rate looks good.
Review a sample of rewritten emails weekly. Check facts the agent pulled from web research. Lock modules that carry claims, pricing or legal wording.
Compute cost per incremental meeting and compare it with what a meeting is worth to you.

If you have the volume, add a third branch that gets no nurture emails at all. It tells you whether the sequence earns its place before you spend on rewriting it.

How many contacts does a fair test need?

These are our calculations, using the standard two-group formula at 95% confidence and 80% power.

Metric	Baseline	Rewritten	Contacts per group
Open rate	40%	44%	about 2,400
Meetings booked	2.0%	3.0%	about 3,800
Meetings booked	2.0%	2.5%	about 13,800
Leads, Sahni's numbers	0.39%	0.51%	about 48,800

The first row explains why open-rate tests reach significance so fast. The ratio of rows two and three explains why most B2B teams can only detect a large meeting lift. If your meeting rate is low and your flow is small, a test can only tell you whether the agent is dramatically better. Decide in advance that "no clear difference" means you keep the cheaper version.

There is also a ceiling on scale. At 20,000 runs a month and five emails per contact, an account can personalize a full sequence for at most 4,000 new contacts a month.

What is an extra meeting worth paying for?

This example is hypothetical. The figures are for illustration.

A software company runs a five-email nurture. About 2,000 contacts enter each month, and 2% book a meeting within 60 days. Rewriting every email costs 5 x $0.10, or $0.50 per contact.

The cost per incremental meeting is $0.50 divided by the gain in meeting rate.

Gain in meeting rate	Extra meetings per 1,000 contacts	Cost per extra meeting
1.0 point (2% to 3%)	10	$50
0.5 point (2% to 2.5%)	5	$100
0.1 point (2% to 2.1%)	1	$500
None	0	No meetings bought; $500 per 1,000 contacts spent

At this volume, detecting a one-point gain needs about 3,800 contacts per group, close to four months of enrollment. A half-point gain needs well over a year. The credit cost of the test arm in the first case is about $1,900, before anyone's review time.

The decision follows from the table, not from the open rate.

What the test shows	What to do
Meeting rate up, cost per extra meeting below your value of a meeting	Raise the share gradually and keep a 10% to 20% holdout running
Opens or clicks up, meetings flat	Keep the original, or keep only the rewritten subject line in a simple A/B test
Meetings flat, unsubscribes or complaints up	Turn it off and audit the rewritten copy
Not enough volume to detect a one-point gain	Test on a higher-volume flow first, or improve CRM inputs before spending
Thin CRM data for most contacts	Fix enrichment first; the agent has little to work with

Where does this leave nurture?

AI rewriting is a copy change with a price per send. Treat it like any other copy change: test it against a control on the outcome you care about, at a size that can detect a plausible effect.

The meeting is still not the end of the funnel. In the piece on demo request follow-up we showed how many requests go unanswered, and better nurture copy cannot fix a booking step that leaks. Nurture email also has to reach the inbox before anyone reads it, which is the subject of our note on the Outlook Promotions folder.

If the contacts entering your nurture come from search and AI answers, the waitlist is open.

Questions

What does HubSpot's Nurture Agent change in an email?

It rewrites the subject line, preview text, call-to-action text and body copy for a share of recipients, using CRM data and web research. It works only on published automated emails. You can lock any module to keep it unchanged. Each personalized email costs 10 HubSpot credits, which is $0.10 at the pay-as-you-go rate.

Is the built-in Nurture Agent report enough to decide?

No. It compares the original and personalized versions on deliveries, open rate, click rate and click-through rate. Opens move when the subject line changes, and many are machine opens. Clicks carry scanner noise. None of the four shows whether more buyers booked a meeting, so build that comparison in your CRM.

How many contacts do I need for a fair test?

It depends on your meeting rate and the lift you want to detect. By our calculation, moving from 2% to 3% of contacts booking needs about 3,800 contacts per group. Detecting 2% to 2.5% needs about 13,800 per group. An open-rate change of 40% to 44% needs only about 2,400, which is why opens declare winners early.

Does AI personalization in email work at all?

The best controlled evidence is mixed. A name in the subject line raised leads 31% across millions of emails in a 2018 field study. A 2025 Mercari field test of AI-written titles lifted taps but not overall purchases. Lab studies of AI-tailored ads show effects on stated preference, not behaviour. Your own list decides.

Sources

HubSpot Knowledge Base, Personalize automated emails with nurture agent knowledge.hubspot.com
HubSpot, Product and Services Catalog legal.hubspot.com
HubSpot Knowledge Base, Use if/then branches in workflows knowledge.hubspot.com
HubSpot, Fall 2026 Spotlight hubspot.com
MarTech, HubSpot's Unbound 2026. The big announcements and updates martech.org
Orange Marketing, HubSpot Updates Monday Recap, September 21, 2026 blog.orangemarketing.com
Media Source, Nurture Agent. El agente que reescribe cada email mediasource.mx
Jobson and others, Using item recommendations and LLMs in marketing email titles arxiv.org
Stanford GSB, Personalization in Email Marketing. The Role of Non-Informative Advertising Content gsb.stanford.edu
Chicago Booth Review, Open this, [NAME]. Why personalized email subject lines work chicagobooth.edu
Matz and others, The potential of generative AI for personalized persuasion at scale pmc.ncbi.nlm.nih.gov
Apple, Mail Privacy Protection and Privacy apple.com
Litmus, Email client market share litmus.com
Gartner, Survey Reveals Personalization Can Triple the Likelihood of Customer Regret gartner.com
Validity, Research Reveals Consumers Let AI Curate Their Inboxes validity.com
Stripo Research, AI in email marketing 2026 research.stripo.email
McKinsey, The value of getting personalization right or wrong is multiplying mckinsey.com.br

Esh

Founder, Ethos Engine. Builds the Digital Authority Score and runs the placement marketplace. esh@ethosengine.io

NEXT STEP

See where your brand stands.

Your Digital Authority Score, who AI recommends for your target questions, and the placements that close each gap. Your first report is free.

Get your free report →

AI now rewrites each nurture email for its reader. Judge it on meetings, not opens