5 min Analytics

Pega details ‘model mechanics’ for real world AI

Pega details ‘model mechanics’ for real world AI

There’s a lot of AI talk. We now exist in a technology industry obsessed with the development, implementation and extension of predictive, reactive, generative and (insert next big AI trend here) intelligence functions. So much is this the case that we seem to have forgotten about core basics such as system observability, application performance management and let’s add in hyper-convergence for good measure. We also appear to concentrating so intensely on the underlying Large Language Model (LLM) structures, their supporting vector databases and the essential guardrails that we need to put in place around AI that, in many ways, we’ve missed talking about the application of intelligence in real world marketplaces. 

Low-code ‘decisioning’ platform company Pegasystems has provided something of an antidote to the current AI reality by making some (arguably) worthy statements about how AI works in commerce that are not driven by any sort of contrived news agenda… so then, what’s selling?

Matt Nolan, senior director for 1:1 customer engagement at Pegasystems (often just known as Pega) might have a snazzy job title, but he’s also all about reality. Reminding us that generative AI gets a lot of press because it can generate content, he says it’s also the perfect tool to help human beings actually understand what’s happening in their data and then act on it. 

“In the past, commercial business teams needed a lot of analytics and technical subject matter expertise to analyse a data set, interpret it… let alone convert that information into an actionable form,” said Nolan. “That challenge is compounded because those analytic resources were typically scarce, so a lot of those insights went to waste – it either took too long to process them, or there was nobody available with the skills and experience to do it.”

Live AI model mechanics

As we know, nowadays generative AI can handle a lot of that heavy lifting – taking complex data and generating an understandable summary of what’s happening, along with a series of recommendations on how to act on what was found. Nolan says that those recommendations could be things like ‘AI model x isn’t performing well, you should look at replacing it’, ‘model attribute y is highly predictive and would add 6% lift if you added it to this model’, or ‘you have 14,346 clients without any relevant offers available, you should create a new retail or sales offer to optimise their value’ and so on. 

“When we look at simulation to understand how a particular sales offer is likely to perform – it’s a relatively straightforward process that uses predictive models to score through samples of production data, so analysts can do ‘what if’ analysis to determine which campaign configuration would give them the best offer redemption,” explained Nolan. “But with AI, now organisations can do that kind of simulation at scale, across all the different offers and messages they have in the market at one time… and across many different parameters of performance.”

For example, he says, if a company has 20 different campaigns running at once, they can use AI to test what would happen if a new offer was introduced into their marketing mix – and, crucially, start to look at whether it would it cannibalise the other offers. Would it appeal to different types of customers? Would it have a negative impact on response rates, or retention, or profitability… and so on.

Big picture, macro-decisions

The Pegasystems customer engagement and enablement man explains that AI gives data (and indeed business) analysts the ability to look at the bigger picture and make macro-level decisions to optimise for big-picture enterprise goals. Subsequently, AI can run these simulations constantly in the background, projecting different scenarios before human beings even consider what questions they’re going to ask… letting it be much more timely with the information it provides.

Concurring with her colleague’s stance on these issues is Tara DeZao in her capacity as director for 1:1 customer engagement at Pegasystems. “Having an always-on brain, AI at the centre of all channels that are working to ingest and analyse data constantly (both internal and external), that can leverage both predictive and adaptive models together to analyze data in real-time is a differentiator,” enthused DeZao.

She suggests that predictive analytics is not a new concept, but adding adaptive capabilities means that the business can react to extremely fresh customer data, data from interactions that happen in milliseconds. We know that AI this powerful is capable of analysing data and understanding the context around the insights in ways that humans were never capable of beforehand.

Moving past predictive-only

“Predictive-only capabilities means that the tech is working solely with historical data, it’s missing that layer of context that allows the business to understand customer behaviour holistically,” said DeZao. “For example, the geographical location [of a store or shop] is an example of context and a brand might engage with a customer differently if they are in a location which is outside of normal patterns – like on an international trip. Brand-to-consumer interactions will be much, much more effective if a business can analyse data insights and real-time and also understand relevant information that provides a broader understanding of the customer’s behaviour.”

The nuanced application of AI at the coalface of real world (often real-time) working live production environments is what will be interesting to track from this point onwards. That process will be made even more compelling if we can understand how models are tweaked, tuned and trained to the nuances of every marketplace they are applied to.