Customer Onboarding Email: from Mail Merge to AI agents

Onboarding email started as onboarding mail. Word processors like WordStar shipped a MailMerge add-on around 1980 that combined a letter template with a list of names and addresses. A welcome packet was a merged letter, a printed brochure and a stamp. Personalization meant the customer's name in the salutation and nothing more.

When email arrived in business in the 1990s, the same model moved online. Teams ran list servers such as LISTSERV and Lyris on their own hardware, or used desktop newsletter tools. Constant Contact began life in 1995 as Roving Software, selling email newsletters to small businesses. Every recipient got the same message on the same day.

What broke was relevance and control. Lists lived in spreadsheets and CRM exports that went stale within weeks. There was no link between what a customer did in the product and what they received. Unsubscribe handling was manual, and spam filters were starting to punish anything that looked like bulk mail.

The workflow was batch and blast: write once, merge, send, and hope. Measuring whether the welcome letter changed anything was mostly guesswork.

Mailchimp launched in 2001 as a side project of an Atlanta web agency, and added a free tier in 2009. That free tier put a hosted email tool in front of millions of small businesses. ExactTarget, Constant Contact and Responsys sold the same idea upmarket: templates, list management and deliverability as a service, priced by contacts and sends.

The CAN-SPAM Act took effect in January 2004 and set the floor: honest headers, a physical address, and a working opt-out. GDPR applied from 25 May 2018 and made consent and data rights a design input for every signup form. Compliance became a platform feature rather than a legal memo.

The real shift was the trigger. Intercom (2011) and Customer.io (2012) let a product send events, such as signed up or created first project, and fire email based on what a user did. The calendar drip became a behavior-triggered drip. Salesforce paid about $2.5 billion for ExactTarget in 2013 to bring this into the CRM.

What broke was the writing load. Every branch in a journey needed its own copy, so most teams wrote one sequence and sent it to everyone. Rules multiplied, journeys turned into spaghetti, and nobody could say which email moved activation.

Machine learning arrived as features, not as a new way of working. Send-time optimization used past engagement to pick the hour each contact was most likely to open. Predictive scores flagged likely churners. Tools like Phrasee generated and tested subject lines. The marketer still drew the journey; the model tuned the dials.

Then measurement broke. Apple's Mail Privacy Protection, launched with iOS 15 in 2021, preloads remote content including tracking pixels, so Apple Mail opens stopped meaning anything. Teams that optimized for opens lost their main signal and had to move to clicks, conversions and product events.

Generative AI landed in 2023. Intuit Mailchimp launched Intuit Assist in September 2023 to draft and edit email content inside the builder, and every major platform followed with a copy assistant. Writing a branch went from hours to minutes.

What did not change: a human still decided the audience, the trigger, the timing and the send. AI made the marketer faster. It did not make the journey smarter, because the journey still ran on rules a person wrote.

Agents now do work that used to need a lifecycle marketer's afternoon. HubSpot launched Breeze agents at INBOUND in September 2024. Salesforce's Marketing Cloud Next, generally available from June 2025, lets Agentforce build the audience, draft email and SMS content and set up the journey from a campaign brief. Klaviyo's Marketing Agent, unveiled in September 2025, learns a brand from its website and returns a plan with campaign and flow drafts.

The decisioning layer moved too. Braze bought OfferFit in 2025 and sells it as BrazeAI Decisioning Studio, where reinforcement learning agents choose the message, timing and offer per customer. Customer.io's spring 2026 release added an in-platform AI Agent plus LLM Actions, a journey step that runs a model against each profile to write, classify or branch. Iterable expanded Nova into Nova Agent in April 2026, and Hightouch launched Lifecycle Studio in June 2026. Smaller tools moved fast: Loops shows an agent setting up a full email lifecycle through its API.

The real limits are worth naming. Customer.io's agent, LLM Actions and Goals launched in beta, and every edit the agent makes needs approval. Klaviyo says every campaign its agent generates is reviewed and launched by the marketer. Braze's Agent Console was private beta at launch. In production, agents draft and assemble; humans still press send.

The binding constraint is no longer writing. As I argued in AI changed SaaS onboarding email, it is product event data and the inbox itself. Gmail and Yahoo started enforcing bulk sender rules in February 2024, and Outlook.com followed in May 2025. A twelve-email drip to every signup is the pattern those rules punish.

My bet is that the onboarding journey disappears as an artifact. Instead of a flowchart a marketer draws, each new user gets a sequence an agent composes from their product events, their plan, and what worked for similar users. The marketer sets the goal, the guardrails and the voice. The agent picks the next message, or decides that no message is the right message.

The second shift is on the receiving end. Gmail's AI Inbox, launched in January 2026, already sits above the inbox and summarizes, ranks and suggests to-dos. Within five years a large share of onboarding email will be read first by the user's own agent, not the user. That agent will ignore anything that does not help its person finish a task. I already see this in my own inbox: cold email has become a positioning audit because senders' AI researches my company before writing.

That changes what good looks like. Onboarding email will need to be structured, specific and useful to a machine reader: the exact next step, the exact link, the exact deadline. Fluffy brand copy gets summarized away.

It also changes the trust model. A sending agent with access to customer records, and a reading agent with access to a user's mailbox, are both identities with scopes. AI agents do not have passwords, so the platforms will need delegated, auditable credentials for both sides.