AI marketing automation, with worked examples

Lead scoring is the clearest example. A rule-based score adds points for a job title on a list and a company size in a range. An AI step can also read what the person wrote in the box that asks what they need, which is often the best signal on the form and the one a points system ignores.

Email flows that adapt are the second example. In Klaviyo, predictive analytics estimates when each customer is likely to order again, so a replenishment email can go out on that date instead of a fixed number of days after the last order. It needs history first: Klaviyo only shows predictions once at least 500 customers have placed an order, with at least 180 days of order history, orders in the last 30 days and some customers who have ordered three or more times.^[1]

The third is reporting. An automation pulls last week's numbers from GA4, the ad platforms and the CRM every Monday, works out the changes, and has a model write a short plain-English summary for Slack or email. Our page on an AI automation agency walks through that build step by step, and the one on AI marketing reporting covers what the report should contain.

The search data shows how new the phrase still is. Ordinary marketing automation draws several times the searches of its AI version in both markets.

The strongest systems use both. Rules handle whatever must happen the same way every time, and AI handles the steps where a person used to read something and decide. A basket flow might keep its timing on rules and let a model pick which of the customer's viewed products to show, and a lead flow might keep its routing on rules and let a model do the scoring.

The quiet failures are the reason for sample checks. Once a week, someone reads a handful of the model's scores or summaries against the source data. It takes minutes, and it catches drift before a sales team stops trusting the scores.

Whoever builds it, the running cost has four parts: the automation tool's plan, the model calls, any data tools such as an enrichment service, and the hours to look after it when something changes. The model is rarely the expensive part. At the rates on Anthropic's pricing page in September 2026, Claude Haiku 4.5 costs $1 per million tokens read and $5 per million written, so a weekly report that reads around 20,000 tokens of figures and writes a 1,000-token summary costs under three US cents a run in model fees.^[3]

The hours are where the money goes, which is why the build should be judged against the work it replaces. Our guides to what AI marketing automation costs and whether one workflow is worth it set out the sums in full.

Whoever you choose, get clear answers to these before the build starts:

We set up automations as part of running a client's marketing, and as an add-on to it. Before he founded the company, Jakub built one for a global technology platform whose paid campaigns were producing leads faster than the team could handle: each Meta form triggered an enrichment agent that researched the lead, checked it for duplicates, wrote a standard CRM record and alerted sales to the strongest fits.^[4][5]