Section 1
What personalisation actually means here
There are three levels, and the value gap between them is large. Cosmetic: name, company, job title inserted into identical copy. Cheap, ignored, occasionally embarrassing when the data is wrong. Segmented: the message differs by group. New customers get a different sequence from customers in month eleven of a contract. Behavioural: the message responds to what this person did. They opened the pricing page three times, they stopped using the feature they signed up for, they downloaded a guide about a specific problem. AI helps most at the second and third levels, and only when the underlying data supports them. Related mechanics are covered in [Streamlining Customer Service with AI-Powered Chatbots](/blog/streamlining-customer-service-with-ai-powered-chatbots).
Section 2
Where the model earns its place
Four jobs, in rough order of return. Drafting variants, so one campaign becomes six segment-specific versions rather than one compromise. Summarising behaviour into a sentence a marketer can act on, turning raw event logs into this account has stopped using the reporting feature. Classifying replies, so intent-to-buy, unsubscribe requests and support questions are routed rather than sitting in a shared mailbox. Suggesting send timing per recipient, where you have enough history to support it. Notice what is missing: the model is not deciding strategy, list membership or offer. Those remain human decisions, and automating them is how brands end up sending a renewal push to someone who churned.
Section 3
Segmentation logic before copy
The common failure is generating five hundred personalised variants against a list nobody has cleaned. Volume of copy is not the bottleneck and never was. Fix the inputs first. One record per person, deduplicated. A reliable signal for lifecycle stage. Suppression rules that actually suppress: current customers do not receive acquisition offers, and anyone in an open support complaint does not receive a promotion. Then define segments by decision rather than demographics. What would you say differently to this group, and why. If you cannot answer that, the segment does not need to exist and splitting it will only dilute your measurement. Testing method matters here: [Automating A/B Testing in Marketing Campaigns with AI](/blog/automating-a-b-testing-in-marketing-campaigns-with-ai).
Section 4
Sending more is how you lose the channel
Cheap production creates a volume temptation, and email punishes volume more harshly than any other channel. Complaint rates and unsubscribes feed deliverability, and once inbox providers decide you are unwanted, your recovery period is measured in months. Set a frequency cap per person across all campaigns, not per campaign. Watch complaint rate as a hard limit rather than a curiosity. Keep the unsubscribe visible and honour it immediately across every list. The other volume risk is voice. Six variants generated separately drift apart, and a customer who receives three of them notices that your company sounds like three different companies. Keep a reference example and check variants against it.
Section 5
Consent, disclosure and the personal data question
Read against the NIST headings of trustworthiness, design, evaluation and use, lifecycle email has two sharp edges: consent and inference. Consent: sending because you can technically reach someone is not permission, and consent obtained for one purpose does not extend to another. Inference: a message that reveals how closely you are watching behaviour reads as surveillance even when it is lawful. Referencing a page someone visited is usually a mistake even where the data is yours. Strip personal fields from any prompt that does not need them. Log what was sent to whom. Keep a human approving anything that goes to the whole list, because an error in an automated send arrives everywhere at once. See [How Storytelling Sparked Viral Marketing Campaigns](/blog/how-storytelling-sparked-viral-marketing-campaigns) for the creative side.
Section 6
What the research says
For drafting work specifically, the evidence is encouraging. In a randomized experiment published in Science, professionals using generative AI finished mid-level writing tasks 40% faster and produced output rated 18% higher in quality (Noy & Zhang, 2023), which is close to the exact task email personalisation involves at scale. The data condition is where programmes fail. Gartner predicts that through 2026, organizations will abandon 60% of AI projects that are not supported by AI-ready data (Gartner, 2025), and a fragmented contact database is the most common version of that problem in a marketing team. The broader caution is consistent across sources. The share of companies abandoning most of their AI initiatives jumped from 17% to 42% in a single year (S&P Global, 2025), and an MIT-affiliated study found roughly 95% of generative AI pilots produced no measurable P&L impact, mostly because tools never integrated into real workflows (MIT NANDA via Fortune, 2025). Only 39% of organizations report any enterprise-level EBIT impact from AI (McKinsey, 2025). The practical reading for a founder: AI raises the speed and quality of email output, but only after list hygiene, segmentation logic and send governance are in place. Automate the campaign after the data, never before.
Section 7
The metrics that survive scrutiny
Open rate has been unreliable since mail clients started pre-fetching images, and it is the metric most often used to declare a personalisation win. Use revenue per recipient as the primary measure, since it captures both response and list quality in one number. Then hold three guardrails: unsubscribe rate, complaint rate and reply sentiment. A campaign that lifts clicks while raising complaints has borrowed from next quarter. Measure per segment rather than per send. Aggregate numbers hide the case where one segment carries the result and three others were quietly annoyed, which is the failure mode that ends channels.