Resource Regions: Western Europe

Unless you’ve been living under a rock for the past six months, you’ll have heard of generative AI – technology that enables computers to create synthetic data or digital content based on previously created data or content. The launch of ChatGPT in late 2022 lit a fire under this emerging space and seemingly overnight, hundreds of millions of people became inspired by the results of work that had already been going on for years within academic and commercial technology vendor research departments.

Earlier in June we spent two days touring around investment banks and hedge funds in London to talk to investors about generative AI and answer their questions.

Download eBook: Generative AI in EMEA: Opportunities, Risks, and Futures

We had many great, in-depth discussions. Here are the questions that came up most frequently.

1. Where is the Value in Generative AI in the Short, Medium, and Long Term?

Today, most of the value is being captured by hardware vendors – most notably NVIDIA, which has seen its share price take off following a sharp upswing in its predicted revenues. As the market leading provider of GPUs with a strong enabling software story and emerging as-a-service play, too, NVIDIA is very well positioned to capitalise on the generative AI boom.

Of course, NVIDIA isn’t the only vendor that potentially stands to benefit; AMD and other semiconductor vendors (including start-ups like Graphcore, Cerebras & Moore Threads) are emerging as challengers, and generative AI platforms will drive storage and networking infrastructure investments too.

In the short to medium term, hyperscale public cloud providers can also expect to benefit significantly. With its early move investing in OpenAI and accelerated investments in generative AI across its software portfolio, Microsoft is in a particularly strong position; but AWS, Google, and Oracle are all also making significant moves in this space.

In the medium-term platform and application vendors also stand to benefit, although the value equation for them is less clear cut. There are significant question marks over which generative AI use cases can support direct monetization, and which will be important to implement from a defensive point of view. Many of the costs associated with managing generative AI models for scale, security, privacy and trust will also fall on their shoulders.

2. What Will Have to Be True to Make GenAI a Truly Broadly Adopted Technology?

Right now, we’re still in “year zero” for generative AI in a commercial context. There is still a lot of confusion around the technology and its applicability in practical real world use cases.

What is already clear, though, is that publicly shared foundation models delivered as a service (such as those hosted by OpenAI) will only be suitable for a subset of enterprise use cases. For many, enterprises will use fine-tuned, specialised domain-specific models that are made available directly to them on a private (or controlled) basis.

The current state-of-the-art in generative AI yields systems that are prone to accuracy problems, difficult to control and predict, and expensive to run. All of these issues need to be worked on.

3. Where Are the Implications for the Software Landscape?

Every software vendor that IDC is speaking to is updating or recreating their product roadmaps to incorporate their respective Generative AI strategies. Obviously, this will play out differently across infrastructure, platforms and applications – however there are certain common questions that are being asked:

• Should we develop our own large language models, or should partner with model providers like OpenAI, Anthropic, Cohere and AI21 and tune them for our software capabilities?
• How should we price our new Generative AI features?
• Should we include getting access to customer data to train models as part of a new set of licensing terms and conditions. What do we offer in return (if anything)?
• Do we need to evolve our support models to include service level agreements (SLAs) on accuracy on certain use cases that are being delivered?

Across all these questions, what is clear is that margin protection will be a major question for software vendors over time – especially those with questionable pricing power. In addition, there will be increased requirements for additional levels of support to deal with model, context and data drift. For the application players, there is an increasing likelihood that forms-based computing as a basis for applications will likely disappear over time and certain markets – for example, salesforce automation and human capital management could potentially be redrawn in the medium-term. 

As part of these changes, what is becoming clear is that the application vendors that are cloud laggards will be AI laggards, and that platforms will continue to dominate the software landscape.

More importantly, incorporating trusted and responsible AI principles into both product development and customer engagement will move from being a differentiator in the short term to table stakes in the medium term.

4. What Are the Implications for Developers?

There’s been a significant amount of excitement about the ability of generative AI services (such as GitHub CoPilot, Replit Ghostwriter and Warp AI) to generate code, documentation, test scripts, and more.

Today’s state-of-the-art models are not going to put developers out of work. Rather, for some specific types of development work, and for some particular types of software asset being created, generative AI services are very likely to help developers accelerate their efforts to deliver working software, acting side-by-side with human developers in a “CoPilot” arrangement.

But it’s important to keep things in perspective: when we zoom out to consider the broader software delivery lifecycle, pro-innovation developers happy to experiment with new tools tend to bump into deployment, operations and support professionals who are much more risk averse.

5. What Are the Implications for Services Providers?

Lastly, many of the investment teams we spoke to were very interested in discussing how professional services (particularly IT services) firms might be impacted by generative AI. Will it bring them major new opportunities? Or will its ability to drive automation of knowledge work mean that it forces providers to cannibalise their own businesses?

Our early research shows that more than 65% of early adopters of generative AI capabilities agree or strongly agree that their need for external services providers will be reduced in the future

The potential impact of generative AI on project delivery is, in some ways, analogous to the potential impact of low- and no-code development tools; if providers can embrace these tools effectively and also deliver trusted solutions to clients, they may find fewer hours are required to deliver projects – but outcomes will be improved for everyone.

Register for the Webcast: Generative AI in EMEA: Opportunities, Risks, and Futures

The arrival of Generative AI technologies has created what we believe to be a seminal moment for the industry: it will be so impactful that it will influence everything that comes after it. However, we believe it is just the starting point. We think that Generative AI will trigger a transition to AI Everywhere – moving us from the use of narrow AI for specific use cases to widening AI for a range of use cases simultaneously.

This means that it will impact every element of the technology stack, and also drive a rethink of all horizontal and vertical use cases. However, given the questions around risk and governance, it will also require every organization to develop and incorporate an AI ethics & governance framework to deal with the risks mentioned earlier.

The investors that we spoke to in London agreed that the tech industry needs to take balanced approach to commercializing the opportunity, while also ensure that policies and regulations continue to protect consumers, enterprises and society as a whole.

The first half of 2023 saw a surge of interest in generative AI (GenAI) that bordered on hysteria. For a few months, the world’s communications channels were abuzz with talk about its potential to impact almost every area of personal, social, and business life. Even industrial organizations started to examine if GenAI could add value to their operations.

GenAI opens access to a wealth of research that can be leveraged to generate a broad diversity of new content. Algorithms can be trained on existing large data sets and used to create content including text, video, images, even virtual environments.

We observe three ways that industrial users can get in touch with GenAI:

1. Publicly Available Tools: ChatGPT-like tools provide users with information, content generation, or codes. These publicly available tools and apps provide solid value to users. From a process area point of view, the great benefits come from gaining market and supply chain intelligence, procurement intelligence, and training. However, these applications are not ideal for industrial use. Some organizations have even banned using them to prevent sensitive data leakage.
2. Embedded Enterprise Solutions: GenAI can be embedded in enterprise solutions like enterprise resource planning (ERP), product life-cycle management (PLM), and customer relationship management (CRM) systems. They can be present as “copilots,” or an AI system designed to assist and support human users in generating or creating content using GenAI techniques. Most technology vendors are already implementing GenAI technology in their enterprise solutions, enabling organizations to benefit from it in areas like service management, supply chain planning, and product development.
3. Use Cases and Apps: Developers can use GenAI to create or empower use cases and to develop apps. My IDC colleague John Snow believes GenAI can bring real value to a wide variety of business areas, assuming it has been trained on relevant data. This means we will see the creation of GenAI solutions specific to areas of expertise (e.g., product design, manufacturing, service/support), industries (e.g., automotive, medical devices, consumer products, chemical processing), and individual companies. Such focused tools will augment — and in some cases challenge — human-generated knowledge and experience as we know it.

Download eBook: Generative AI in EMEA: Opportunities, Risks, and Futures

Be Ready — But Careful

In operations-intensive environments like process manufacturing, AI may provide a handful of beneficial use cases. These could include production planning models and the predictive maintenance of complex simulations through soft sensors.

Users have already learned to leverage the power of AI in daily operations in a safe way (i.e., in areas where the impact of a potential failure on the physical environment is minimal). Image recognition models, for example, can be trained on available data sets, enabling the model’s outputs to be verified against a standard.

AI is already part of countless aspects of manufacturing — but the reliability of AI-generated outputs remains unsettled. IATF 16949 is a great example. A global quality management standard developed for the automotive industry, it provides requirements for the design, development, production, and installation of automotive-related products. However, the standard does not explicitly cover AI or provide specific requirements for AI implementation.

AI can still be relevant in the automotive industry, however, and its applications may have implications for quality management. AI can be used in areas such as autonomous vehicles, predictive maintenance, quality control, and supply chain optimization.

Standards and regulations are continuously evolving — and new guidelines specific to AI or emerging technologies within the automotive industry may be developed in the future to address their unique considerations and challenges.

Output Challenges

Like any other methodology that serves industries, GenAI outputs must be 100% reliable. Most readers are probably familiar with the application of reproducibility and repeatability. Let me remind you that reproducibility allows for more accurate research, whereas repeatability measures that accuracy and confirms the results. Both are a means to evaluate the stability and reliability of an experiment and are key factors in uncertainty calculations of measurements.

GenAI-based tools might seem to be a black box for many potential industrial users. GenAI bias is a significant fear. This refers to the potential for biases to be present in the outputs or generated content produced by GenAI models. These biases can arise from various sources, including the training data used to train the models, the algorithms and techniques employed, and the inherent biases present in human-generated data used for training.

GenAI models learn patterns and structures from large data sets. If those data sets contain biases, the models can inadvertently learn and perpetuate those biases in their generated content. For example, if a GenAI model is trained on text data that contains biased language or stereotypes, it may generate text that reflects those biases.

GenAI bias can have several implications. It can perpetuate stereotypes, reinforce discriminatory practices, or generate content that is misleading or unfair. In some cases, GenAI bias can lead to the amplification of existing societal biases, as the generated content may reach a wide audience and influence perceptions and decision-making processes.

Addressing GenAI bias is a crucial aspect of using it properly — and mitigation of bias is a crucial stepping stone to increasing the technology’s reliability. Model creators and owners should ensure that the data used to train GenAI models is diverse, representative, and free from explicit biases.

If possible, mechanisms to detect and mitigate bias during the training and generation process should be implemented. Generated outputs should be continuously evaluated and monitored for biases. This includes the establishment of feedback loops with human reviewers or subject matter experts who can provide insights and flag potential biases.

We recommend striving for transparency and explainability. Make efforts to understand and interpret the internal workings of models to identify sources of bias and address them effectively. User feedback and iteration of GenAI models based on that feedback is encouraged.

Users must also be wary of GenAI “hallucinations,” or situations where a GenAI model produces outputs that appear to be realistic but are not based on real or accurate information. In other words, the AI system generates content that is plausible but may not be grounded in reality. For example, a generative AI model trained on images of defects may generate new images of defects that resemble those in an existing defect category but do not actually exist.

Avoiding AI hallucinations entirely is challenging, but there are several actions that can be taken to limit occurrence or minimize impact. Let’s touch on a few: It is crucial to ensure that your AI model is trained on a diverse and representative data set that covers a wide range of examples from the real world. To improve the quality and reliability of the model’s outputs, the training data should be preprocessed and cleaned to remove inaccuracies, outliers, or misleading information. The model’s outputs should also be continuously evaluated and monitored to identify instances of hallucination or generation of unrealistic content.

Register for the Webcast: Generative AI in EMEA: Opportunities, Risks, and Futures

Evolving Challenges

Because they involve generating new and original content without explicit programming, proving the reliability of GenAI models can be challenging. However, there are several approaches you can take to assess and provide evidence of the reliability of GenAI models.

Commonly used methods include defining and utilizing appropriate evaluation metrics to assess the quality and reliability of generated content. Evaluation by humans is useful, including subjective evaluations that involve assessing and rating the quality and reliability of generated content.

For some specific use cases (e.g., copilots), test set validation can be utilized. This includes creating a test set of specific scenarios or inputs representative of the desired output and evaluating the generated results against these inputs.

Adversarial testing can also be employed to deliberately introduce challenging or edge cases to the GenAI model to assess its robustness and reliability. As GenAI outputs evolve, it is recommended that long-term monitoring be used to continuously track and evaluate the performance and reliability of the model. This could be applicable, for example, in supply chain intelligence GenAI-powered applications.

The Sky is the Limit — For Now

In the industrial environment, we are still scratching the surface of what GenAI can do. Organizations should collaborate with tech vendors and service providers to understand the value of GenAI and turn it into a significant competitive advantage. Regulators may try to restrict or otherwise control GenAI technology, but the cat is already out of the bag. Development is inevitable.

To get first-hand information about the development of GenAI, organizations should follow well-known AI technology specialists, as well as start-ups and hyperscalers. Hyperscalers like Google, Microsoft, and Amazon are at the forefront of AI research and development. They invest significant resources in exploring and advancing AI techniques, including GenAI. Hyperscalers often offer cloud-based AI services and platforms that include GenAI capabilities. Keeping up with their offerings can help you understand the latest tools and services available for developing GenAI applications.

Managers traditionally expect to start seeing ROI for tech like GenAI within 1.5 years — but with the right IT infrastructure in place to deliver scalability of GenAI tools, an ROI target could be reached within months. Improved customer service, for example, brings additional revenues almost immediately. And process optimization using data intelligence can provide improved productivity while reducing costs incurred due to poor quality.

Beware the Competition!

GenAI is poised to revolutionize the manufacturing industry, enabling manufacturers to unlock new levels of efficiency and innovation. From product design to supply chain optimization, GenAI can have a significant impact on KPIs.

But beware: Do not allow the competition outrun you in terms of GenAI adoption. Stay on top of developments and act before competitors use GenAI to threaten your business.

At the same time, do not underestimate the risk of intellectual property (IP) leakage, or the unauthorized use, disclosure, or exposure of valuable intellectual property through the utilization of generative AI models. Embed an IP leakage prevention mechanism in your general AI and data governance. This should include removal or anonymization of sensitive or proprietary information from training data sets.

As always, stay busy with what works — but keep an eye focused on the future. Embracing this transformative technology is a crucial step toward more efficient and innovative prospects for businesses of any size.

You often see it on television: programs about people who are struggling financially. They run out of money at the end of the month, they can’t sell their house, they have a problematic debt burden, and so on. A common denominator is often the lack of insight into their own situation, and while coming up with ways to save money may not be very difficult, actually implementing and sticking to them is much harder.

I mean, it’s easy for an outsider to suggest that someone should get rid of their dog, but if that pet is their only source of comfort, it will take some effort.

The same goes for cloud costs: saving money is easier said than done. There are all sorts of great tools available from both cloud providers and third parties to help you understand your costs.

These tools provide various reports and dashboards, and even recommendations on which instances to remove or resize (rightsizing). With the right knowledge, you can also determine how to use discount options (reserved instances, savings plans, reserved capacity, etc.), how to manage licenses intelligently, and what you can do in your application architecture to save costs. And, of course, you can always turn off instances when you’re not using them.

All of this insight is great, but then comes the second part. Just as people have a hard time saying goodbye to their pets, users and administrators have a hard time shedding their old habits and ways of thinking. And that’s something cloud providers never talk about.

For example, consider turning off instances outside of working hours. In theory, this is an excellent way to save money, but instances are part of applications, which in turn are part of chains. It can happen that data exchange takes place in a chain outside of working hours.

Testing teams that are under a deadline may also need their environment outside of the predetermined working hours. And if environments are used in the management chain, they must also be available after working hours in case of an emergency. So savings are theoretically simple, but practice is more complicated. It can be done, but it takes a lot of effort.

Rightsizing is also less straightforward than it seems. Users and administrators are often hesitant to remove capacity: users see their performance decrease, and administrators see the risk of more outages because there is less excess capacity to handle issues. In the latter case, you need to analyze where these issues are coming from: a poor application can benefit from more capacity, but that is not a long-term solution.

If the roof is leaking, you can replace the bucket you use to catch the water with a mortar tub, but even that will eventually fill up. Ultimately, you’ll have to repair the roof.

So, objections can be raised for all types of savings. Eventually, you’ll need to adopt an approach that not only makes costs visible but also involves users and administrators, and leads to the right considerations on where to save on your cloud costs and where not to.

Don’t know where to start? Can’t figure it out quickly enough? IDC Metri has helped several organizations get started. Our specialists can help kickstart your cost-saving efforts in the cloud. Because understanding costs is one thing, but it’s only useful if they actually decrease.

Want to learn more? Subscribe to IDC Metri’s monthly newsletter full of actionable insights on IT benchmarking, intelligence, sourcing and more.