The User Guide to Composable AI

Introduction

Xiatech are the providers of Xfuze - the world's first MACH-certified, composable Integration and Data Platform.

Xfuze enables three services:

1. System Integration
2. Real-time Single View of Data (Customers, Inventory, Product, Sales etc
3. Data Lake, Analytics, Insights, AI/ML

See: https://www.xiatech.io/

This paper (https://www.xiatech.io/wp-content/uploads/2024/03/Xiatech-The-User-Guide-to-Composable-AI.pdf) has been written to guide business, technology and data teams on how to plan and deploy artificial intelligence initiatives more effectively using Composable AI, a modular way to quickly generate tangible and incremental business value from AI investments, while helping those investments to scale for the long-term. Although some entry-level artificial Intelligence capabilities are now tantalisingly inexpensive, AI investments have historically represented substantial and ongoing investments. Consider that just the daily operating cost for ChatGPT has been pegged at $700,000. Even a lean approach to AI can cost hundreds of thousands of pounds and many businesses invest millions with no guarantee of success or value generation. Furthermore, deployed AI systems tend to work in an isolated environment with limited access to wider business processes – another example of how new technology often recreates silos. The clear competitive advantage that AI should provide is now spurring investments on a massive scale but even companies that are mature and experienced can have difficulty finding the value they aimed to achieve, and too many initiatives simply fall flat. In fact, according to a recent article in Harvard Business Review, the failure rate could be as high as 80% for AI projects, “almost double the rate of corporate IT project failures a decade ago.” An earlier article in IIOT World featured a similar lament and cited Gartner as pegging the failure rate at 85%. There are many reasons but three underlying problems contribute to most failures, namely the challenge of ensuring each AI model is working as efficiently as possible, each has access to the information it needs, and each is functioning effectively with the others. Few succeed in implementing that ideal, resulting in disappointments and cost overruns. One of the reasons for this comes down to the different goals in adopting AI usually involves adopting different kinds of AI technology and solutions such as machine learning, natural language processing, or Generative AI. Each of these can potentially magnify the complexity and management challenges your organisation is already facing. Xiatech is simplifying the adoption of AI through Composable AI, which this paper outlines so you can adopt AI more effectively than using traditional approaches

For the full paper, check out https://www.xiatech.io/wp-content/uploads/2024/03/Xiatech-The-User-Guide-to-Composable-AI.pdf

Composable AI is the best way to create value from your AI investments. It enhances speed to market and incrementally improves the value delivered by AI. This is so because once the first AI model is in place, all of the functionality and tasks that are common to all Composable AIs will not need to be implemented for subsequent Composable AIs (software, data and BI infrastructure, data connections, etc). Furthermore, because Composable AI maximises flexibility, it is much easier to shift and refine the capabilities of AI while in motion so that the organisation can seize opportunities and steer clear of risks: AIs become components that can be plugged in or out according to their performance and the latest business needs. It ensures that the organisation is fully prepared for changing market conditions and ready to seize new opportunities, while optimising the operational costs of AI. Because of all that, Composable AI also generates a quick payback. Once the foundation of Composable AI is in place, incrementally adding AI models not only provides new AI capabilities, the total cost of each AI quickly decreases as business value is generated from the initial and proceeding investments so organisations witness a faster overall AI payback than with Isolated AIs. Last but not least, Composable AI is aligned with modern IT thinking by being modular and flexible, able to integrate and build on the best available data and technology, while giving the business the agility to adopt AI that delivers results in an ever changing environment. An article in The AI Journal noted that 2023 was the first year that Composable AI surpassed platformbased AI as the preferred approach among a surveyed group of manufacturers. This implies a small but significant revolution in how deployment is conceived and implemented. Rather than being held hostage by a highly linear process, often signified by the “waterfall” development approach, Composable AI invites continued innovation and integration on an ongoing basis. That means, companies no longer have to risk an innovation becoming obsolete before it is implemented and it spells an improvement in time to value. Xiatech, building its approach to Composable AI on its modern all-in-one Xfuze Hyper-Integration Platform, gives customers the capability to integrate AI/ML resources with each other, and with any other IT resources. It is an approach that pulls back the veil to reveal AI as it should be: Adaptable, integrated, accessible and more productive from day one.

With Xiatech’s Xfuze Hyper-Integration Platform, customers can integrate multiple AI models enabling them to collaborate across a complex ecosystem of technologies, data and processes.

Frequently Asked Composable AI Questions

1. What are the common steps to deliver each Composable AI solution? We have provided a step by step journey of how each Composable AI model is deployed through an ecosystem. This can be found on page 6 of our report (https://www.xiatech.io/wp-content/uploads/2024/03/Xiatech-The-User-Guide-to-Composable-AI.pdf)

2. How do Composable AI models decide which data they will use when training and optimising themselves? Each Composable AI model contains a dictionary of all possible features it could use to function, with some of them being compulsory and others optional. Some of the optional features might be outputs from other Composable AIs. In this case, both the Composable AI outputting data as well as the Composable AI receiving contain mechanisms to measure how trustworthy that information is (what is the likelihood to be erroneous, how complicated was to produce that piece of information, etc), so the receiving Composable AI can make a decision on whether to use it or not. The overall aim is to re-use the good information being produced from Composable AIs, while avoiding errors to be propagated across the Composable AI ecosystem.

3. How do Composable AI models avoid errors to be propagated into other Composable AI models that will utilise their outputs? Composable AIs produce all types of outputs that are then published into the Composable AI ecosystem for other Composable AIs to reuse if they want to. One caveat of this is that, as all AI systems, Composable AIs might be producing erroneous results from time to time. In order for these errors not to be propagated across the Composable AI network, every piece of data being produced by Composable AIs has a confidence or reliability score. These scores will help other Composable AIs determine whether to use that piece of data or not.

4. Can Composable AI work if only a subset of data is available? Each Composable AI model contains a dictionary of all possible features it could use to function, with some of them being compulsory and others optional. As long as all compulsory features are available, the Composable AI will be able to run. However, the more optional data is plugged in, the better the performance of the Composable AI is expected to be.

5. What’s the ideal time range of historical data to be used in Composable AI models? A typical rule of thumb is that two years is the ideal amount of data needed to spot different behavioural traits elongated over time as well as seasonality and holiday patterns. However, that is not a hard requirement for any Composable AI to work. We just expect performance to increase as more historical data becomes available. That said, one important caveat is that sometimes past data might be misleading in periods of extraordinary uncertainty such as a global recession or pandemic, or a large-scale conflict. In those cases, it becomes necessary to either reduce the noise related to those extraordinary events, or neglect that time period entirely. 10

6. Can Composable AI work outside of Xfuze? Yes but it would cost a lot of money because most platforms in the market don’t have the full suite of capabilities such as found in Xfuze, which features system integration, data management, business intelligence, advanced analytics and process automation.

7. Can Xfuze deploy any AI model? Yes, we can deploy any AI model including Gen AI on the Xfuze platform. In addition, thanks to Composable AI, the Gen AI model becomes connected to other AI models, and, using Xfuze, is accessible across the organisation. By connected, we mean that any other model in your Composable AI ecosystem can utilise the outputs of your Gen AI implementation model. Likewise, Gen AI can benefit from the models currently deployed through the Composable AI ecosystem.

8. What are the investments needed to go live with Composable AI? Very little if you embrace Xiatech’s approach to Composable AI and use its Xfuze platform. Generally, with any AI project, there are a number of investments required to develop, test, deploy and maintain each Isolated AI model and these include a data lake, devops infrastructure, a training / optimisation pipeline and an inference pipeline. You also need people including a data scientist, a machine learning engineer, a data engineer, a devops engineer and a business analyst. By embracing Composable AI, and using Xiatech’s Xfuze platform, you don’t need to buy and maintain the technology investments, which in the future will be shared across all Composable AIs, and the only people cost might be a data scientist if you don’t want to use Xiatech’s experts. In other words, the fastest way - and the most cost effective way - to deploy Composable AI is to leverage a solution provider such as Xiatech.