We ran AI for 100 million users before the wave.

What did AI at scale look like before LLMs?

It looked like the Active Content Platform, first deployed to 20 million users in 2012, more than a decade before the wave had a name. ACP worked out, in real time, what was most relevant to a shopper: in email, on the web, in the moment. It weighed location, shopping history, even the weather, and it did it at a scale that still stands up: 100 million users, a million transactions a second.

What the numbers do not show is what they cost. Systems like that are not assembled from tutorials. They took grit, engineering excellence, and something rarer: clients trusting us enough to put their customer relationships through our platform, every second, for years. That trust was the hardest-won part, and the part we are proudest of.

That capability ran in production at national and global scale: real-time personalisation with privacy designed in, a worldwide loyalty programme taken from data seed to live in four weeks, multi-lingual and multi-currency rollouts across the US, EMEA and Asia, and one-to-one customer communication at every stage of a journey. We name few of the organisations involved, and that is deliberate. Much of our best work is for clients in regulated and high-stakes environments who reasonably do not want it discussed, and that was as true then as it is now. The thread was always the same: put the right thing in front of people at the right moment, and both the customer and the business are better for it.

What does production teach you that pilots cannot?

Four lessons, learned the expensive way so our clients do not have to.

First, data quality decides outcomes before any model gets a say. Second, integration is harder than intelligence: the clever part is rarely the part that takes the time. Third, governance is what allows scale; systems handling millions of customers inside regulated businesses do not run on enthusiasm. Fourth, adoption is won in workflows. If the new capability does not fit how people actually work, it does not matter how good it is.

And a fifth, quieter lesson: the destination of good technology is boredom. ACP stopped being exciting years ago. It became infrastructure, the kind nobody would give back. It is a legacy platform now, and that is precisely the point. Waves pass. Capability keeps working.

Why does this matter in the LLM era?

Because the gap every organisation is struggling with right now, the gap between a promising pilot and a production system that people rely on, is the same gap we have been crossing since before the vocabulary existed. The models are new and remarkable. The physics of production are unchanged.

When we tell a client which AI use case will survive contact with reality, that judgement was not learned from a webinar. It was learned running a million transactions a second for some of the most demanding brands in the world.

Common questions

What is the Active Content Platform?

Altaire's personalisation platform: AI-driven, real-time content decisioning at enterprise scale. Legacy now, and proudly so: it ran in production for years.

Is ACP available as a product today?

ACP is legacy. What we sell today is the judgement earned by building and running it.

What transfers to LLM projects?

Almost everything that matters: data quality, integration, governance, adoption. The models changed. The physics of production did not.