AI Marketing & MarTech
AI for Marketing: Real Use Cases and Workflows
When I started bringing AI into real marketing operations at Xister Reply, the first thing I learned is that the value is not in the model but in where you place it in the process. I implemented multilingual video dubbing with AI for Enel, launched an AI-generated podcast service for Toyota and Intesa Sanpaolo, and took hypothesis-driven experimentation on Optimizely to a +57% lift in conversion on the Whirlpool EMEA e-commerce. In this guide I collect the use cases I have actually seen work, the tools I built them with, and the workflows that hold them up in production, without the hype that crowds every conversation about AI in marketing.
Where AI creates value in marketing (and where it is just noise)
The question I ask before every project is not "can we use AI here?" but "is this task repetitive, high-volume and well-specified?". AI makes almost free what used to cost time and people when the task has clear rules: translating a script, generating variants of an asset, syncing audio and lip movement, summarising a report. It does not work as well where judgement is needed: defining a brand's positioning, deciding which story to tell, understanding whether a campaign makes sense for that market at that moment.
I sharpened this distinction studying for the Microsoft AI Product Manager Specialization and the Microsoft Program Manager Specialization, but above all by applying it in the field: every time a client asked "can we automate everything?", the useful answer was "let's automate the mechanical part, not the part that decides."
The Enel case: multilingual video dubbing and −75% on localisation costs
The most concrete project I have run on this front is multilingual video dubbing with AI for Enel. The starting problem was simple to describe and expensive to solve: corporate and training videos to adapt into multiple languages, with the traditional process running through a recording studio, professional voice actors and review cycles for each target language. Every additional language multiplied time and cost almost linearly.
We built a pipeline where AI handles the mechanical part (transcription, translation, voice and lip-sync) while people remain responsible for the linguistic review and the final tone check before publication. The result was a 75% reduction in localisation costs, keeping the quality a brand like Enel requires on content aimed at different markets.
AI-generated podcasts: the service for Toyota and Intesa Sanpaolo
The second case I often bring up is the AI-generated podcast service I launched for Toyota and Intesa Sanpaolo. Here the constraint was not the production cost itself, but the cadence: a corporate podcast that has to ship regularly needs a script, a voice consistent with the brand, and editing that previously required a dedicated audio studio for every episode.
With AI we could generate the audio base at a volume a traditional studio would not have sustained on the same budget, but the project only worked because we kept a human step on three points: writing or reviewing the script, the consistency of the voice over time, and the final listen before publication. A podcast generated entirely without supervision, for two brands like Toyota and Intesa Sanpaolo, would have been a risk neither client would have accepted.
Integrating generative AI into the content workflow without losing control
The two cases above share a structure I now apply to any AI content project: you automate the production of the first version (the draft, the translation, the variant) and leave people the decision on what actually ships. This means designing the workflow with explicit checkpoints, not trusting that "the model is good" and skipping the review.
In practice, for every content process I bring into AI I ask three questions: who is responsible for the final output (not the prompt, the output), what human checkpoint exists before publication, and what happens when the AI gets it wrong. If I cannot answer all three, the process is not ready for production, it stays an experiment. I go deeper into this in the guide on AI program management, where I cover the distance between a pilot that works in a demo and a system that actually holds up in production.
Personalisation and experimentation: every AI-generated variant is a hypothesis to test
One of the most promised and most often mishandled uses of AI in marketing is personalisation: generating different content, offers or messages for different user segments. The problem I see most often is treating it as an act of faith, you launch personalisation, assume it works because it is "AI," and never measure whether it converts more than the single version.
In my work I apply to personalisation the same discipline I use in Optimizely experimentation programs: every AI-generated variant is a hypothesis, not a certainty, and must be measured as such. The +57% lift in conversion we achieved on the Whirlpool EMEA e-commerce did not come from launching random variants, but from hypotheses built on a specific behavioural insight and verified with an A/B test. The same logic holds when the variant is written by a model instead of a copywriter: what changes is who generates the content, not the need to measure its effect.
How to choose AI marketing tools without chasing the hype
The AI marketing tool market changes every few weeks, and I have learned to distrust demos. The criteria I use to evaluate a tool before introducing it to a client are always the same four: it integrates with the stack the client already uses (CRM, CMS, analytics platform) without requiring a second parallel system; the cost scales predictably with real volume, not just in the demo plan; the workflow includes a human step where needed, not just "automatic generation"; and the tool lets you measure impact, not just produce output.
This approach also comes from direct experience with the martech stack: from the Salesforce Marketing Cloud Administrator certification to the Optimizely Experimentation Certified Core Strategist role I use every day in the field. A tool that looks powerful in isolation often turns out to be a bottleneck if it does not talk to the rest of the system.
From pilot to scale: governance and KPIs so you do not stay stuck in the PoC
Most of the AI marketing projects I see fail do not fail because of the chosen model, they fail because no one decided in advance who owns the output, which KPIs to track to tell whether the project works in production, and what to do when something goes wrong. A PoC that impresses in an internal demo and a system that holds thousands of pieces of content a month are two different things, and the difference almost always lies in governance and instrumentation, not in technology.
On the projects I have run, from the Enel dubbing to the podcast for Toyota and Intesa Sanpaolo, to the experimentation programs on Whirlpool EMEA, the constant has been defining before launch who approves what, which metric declares the project a success, and with what cadence the result is reviewed. Without this, even the most promising use case stays stuck in the pilot phase.
Where to start: the first process to automate
If I had to recommend a single first step, it would be this: do not start from "where can we use AI in marketing," but from "what is the most repetitive, high-volume and best-documented process we have today." Measure what it costs today in time and budget, then assess whether AI can handle its mechanical part while leaving a clear human checkpoint. If the answer is yes, you have a pilot with a real business case behind it, not a technology experiment with no objective. The details of how I structure these projects, from initial assessment to production, are on my services page.
Talk to me about your AI marketing project
The thread linking Enel, Toyota, Intesa Sanpaolo and Whirlpool EMEA is not the technology used, but the discipline with which we introduced it: automate what is repetitive, measure what changes, and leave people the decisions that matter. It is the same approach I bring to every new AI marketing project, regardless of which tool is fashionable that month.
Let's talk about your project
Found this useful? Let's see how it applies to your context.
Related articles
Evaluating and Choosing AI Marketing Tools
How to evaluate and choose AI marketing tools without following the hype: the criteria I use to decide what to adopt and what to leave on the market.
AI Multilingual Localization: Cutting Costs
How to cut multilingual localisation costs with AI: the Enel video dubbing case, with a 75% cost reduction while keeping quality.
AI Video and Podcasts: Two Real Enterprise Cases
Two real enterprise cases of AI-generated content: video dubbing for Enel and a podcast service for Toyota and Intesa Sanpaolo. How I built them.