The Universal Truth: AI and Data

There's a lot of buzz about AI these days. There's also a lot of buzz about data. However, in the conversations I'm having, the general feeling I have is that all of the conversations generally have too short of a horizon, in terms of building out the necessary architecture and infrastructure to power this.

Is this natural? Yes.

Like any new wave, will there be a motion of Build and Re-build? Yes.

Will there be a first-mover advantage for those who get it mostly right? Yes.

This is why I'm writing this post. To provide my two cents on Data and AI, and the way I'm currently being "schooled" to think about it. The general rule is this:

More Data = Better AI

This rule, to me, is not debatable. This does not mean there aren't nuances to consider. But let's start with the basics first.

Example # 1 - Product Interest Score or NBO

Classic marketing use case. Imagine that you have a customer browsing your website, looking at X, Y, Z products and buy a "Modern" sweater and a "Country" styled hat. Your data points are now:

• Web Behaviour: Products, Categories, Colors etc. browsed
• Order & Order Lines: Sweater and Hat. SKUs etc.

If we browsed 60+ modern items, and 5+ country items, our interest is skewed in one way of course. Natural assumption is you're into modern stuff mainly.

But what if you return the sweater, and don't buy a new one. Then one potential hypothesis, is that you wanted to see if the modern style fit you, as an experiment, hence why it took so long, and you might determine that it's not for you, leaning back into the safe choice of your country wardrobe.

The above is predicated on your AI model getting enriched with:

• Returns: I.e. Warehouse data and returned SKUs

However, more data = better AI. So let's play it further. There are at least two additional things that could be insightful that we could learn:

• Size of Items vs. Size of Orderer: Was this actually an item for you? or was it a gift to a friend, a partner, a child? How do we make the right inferences.
• Style Aspirations / "Celebrity Aspirations": Tell us somebody whose style you love and what parts of it. This can help tailor recommendations towards the person you're looking to become more like and move away from your current style

The more you learn about me, the more you know what to recommend. The age old marketing fallacy of people being "in market" for buying something vs. being "pulled into the market" still apply. Way easier to sell to those in market. The more you know, the more accurate your recommendation, the more you convert.

Example # 2 - GPT Content

Second example, will keep it brief. "GPT Generated Emails" are all the rave these days, or more generally, text generation is highly accelerated by GPT capabilities.

More Data = Better AI

• Scenario 1: Write a sales email for the CMO of Carlsberg to pitch a new Marketing Solution
• Scenario 2: Write a sales email for the CMO of Carlsberg to pitch a new Marketing Solution, enriched with industry references
• Scenario 3: Write a sales email for the CMO of Carlsberg to pitch a new Marketing Solution, enriched with industry references and enhanced with the key results from case studies in the industry
• Scenario 4: Write a sales email for the CMO of Carlsberg to pitch a new Marketing Solution, enriched with industry references and enhanced with the key results from case studies in the industry, making the style of the email and the top 3 bullets informed by the CMOs tweets and linkedin post

Now, I'm aware that this 4th scenario contains hard-to-come-by data, that needs to be mapped etc. - some things are doable, but the alternative could be to have the email be informed by the last 10 meetings you've had with other stakeholders at the client. Highly doable.

The point is this -> The more we know -> the more the AI model know -> the better the recommendations can be.

Why does it matter?

Back to my starting point. The horizons of the conversations I'm having with customers are generally not long enough. There's of course a lot of nuances in all of this. One key point to make is this:

Do not give all your data to a single AI model.

But, you do need to consider, for all of your AI use cases, what are the data sets that would presumably make the AI use case better?

• Product Interest Score -> Is it just "browsing" data? what about orders? returns? in-store talks? chatbot interactions? celebrity aspirations? etc.

And tying this back to the heart of AI -> It's powered by technology, this would be my closing considerations:

1. You need to consider making the technology investment that will unlock and un-trap the data from the various source systems from across your enterprise
2. You should consider having the architecture that caters for the lowest amount of data replication / integration work
3. You could consider all your AI use cases on as few platforms as possible, as this allows an accelerated time to market and greater return on investment

The balance of an AI powered solution of course includes balancing the amount of data and the amount of compute power needed, as this informs the cost of operating the solution. However, having too short of a time horizon, incurs switching costs and ultimately impacts time to market down the line. Hence, taking adequate care in the early phases, to me, is the superior strategy.

Keen to hear your thoughts on this, and if you are looking for a sounding board, please don't hesitate to reach out !