The current wave of "AI email" tools almost all share one shape: a large language model bolted onto a legacy, reactive ESP. The ESP still waits for a human to decide what to send and to start the work; the model just makes a step or two faster.
It is a copilot stapled to a tool built for a different era. That is the baseline we set out to beat, and beating it required rejecting both halves of it: the reactivity, and the thin-wrapper architecture.
What the industry builds today, and why it caps out
A reactive ESP, however much AI you sprinkle on it, is bounded by human initiation. Nothing happens until a marketer decides it should, so the volume and quality of a brand's email is capped by how much time a small team has. Bolting a general model on top raises the ceiling on individual tasks but not on the thing that actually limits output: someone still has to notice the opportunity and kick off the work.
And architecturally, a thin wrapper inherits the base model's ceiling. As foundation models improve, the wrapper adds less, because any competitor can call the same model tomorrow.
Why we think the returns live somewhere else
Two convictions drive the whole system. First, the revenue a brand is missing is not in the campaigns it sends, it is in the ones it never gets around to, the opportunities that die in the gap between a busy team and a full calendar. Removing human initiation as the bottleneck is where the outsized return is, and reactive tooling structurally cannot do that.
Second, the durable advantage in this category is not the model you call, it is the proprietary data you learn on and the system you build around it. Those two beliefs, proactivity and proprietary data, are why the architecture looks the way it does.
How we approach it: two intelligences
The hard problems in email are knowing problems, not writing problems: what is worth sending this week, to whom, when, at what offer, and whether it actually made money. A general model has never seen that a subject line underperformed for this brand last quarter or that a cohort is quietly disengaging; that is learned from behavioral data. So we split the work.
A decisioning layer of proprietary predictive models forecasts revenue and margin, scores every customer, and sets timing and audience, this is what knows. A generative layer of a set of AI agents writes, designs, and assembles the finished campaign, this is what builds. Machine learning is what knows; the agents are what build and act.
Brand DNA: the context both layers learn on
Neither layer runs blind. Both read from Brand DNA, a living per-brand context holding the brand's voice, competitors, past creative and how it performed, and every reviewer approval and rejection.
It is not a profile filled out once. Every send and decision writes back to it, so the system knows each brand better over time, on that brand specifically.
The evaluation stack: how autonomy stays accountable
Autonomy is only useful if it is trustworthy, so nothing generated reaches a human unchecked. Automated brand and quality evaluations screen the work. Guardrails mined from past rejections constrain it.
Performance priors bias it toward what has worked. A human approves or rejects.
And a holdout, a control group that received nothing, decides whether a campaign actually drove incremental revenue rather than revenue that would have happened anyway. Brands see up to 22% lift measured this way. Each layer is an evaluation; together they let a marketer hand over real work without handing over judgment.
How the data compounds
A single brand often sends too little to train stable models alone, so each brand starts from patterns learned across the whole base, then sharpens on its own data as it accumulates. Every approval, rejection, and outcome adds signal. Data here is not a static asset, it is a compounding one, and it is the part no competitor gets by calling an API.
The systems below go deep on each piece. LTV.ai is hiring engineers: see careers.
- How LTV.ai optimizes subject lines while the campaign is still sending
- How LTV.ai forecasts a campaign's revenue and margin before you send
- How LTV.ai turns every approval and rejection into training signal
- How LTV.ai learns a brand's visual identity
- How LTV.ai turns send performance into a creative playbook
- How LTV.ai picks products for a campaign from plain language
- The predictive models LTV.ai scores every customer with
- How LTV.ai shapes exactly who receives a campaign
- How LTV.ai builds audiences three ways
- How LTV.ai segments the audience for the campaign in front of you
- How LTV.ai finds the hour each customer actually reads email
Frequently asked questions
How is this different from an ESP with AI added? An ESP with AI added is still reactive and still a wrapper. LTV.ai is proactive and built on proprietary decisioning models, with the language models used for creative and execution.
How do you prove it works? A holdout. Lift is measured against a control that received nothing, so it is incremental, not attributed.