Adopting Generative AI Solutions in Enterprises
As we advance into the era of generative AI-led transformation, enterprises and business leaders are recognizing generative AI as a strategic imperative. The landscape of generative AI is rapidly evolving, with the various players locked in an intense competition for market share. Users and businesses have become excited about this technology’s value. However, the real adoption rate for enterprises has not matured yet given the concerns around data protection and transparency while using generative AI solutions.
Below, we evaluate the current state of the Generative AI landscape from the perspective of large enterprises and provide a framework for implementation, covering crucial considerations like validation, trust, access control, data protection, and compliance.
Is Generative AI Hype or an Opportunity?
Generative AI as a technology is in an embryonic stage, meaning there are both risks for deployment and benefits for early adopters. Many businesses are keen observers and only a few have gone on to implement, mostly in experimental areas like marketing or customer care that do not risk their core business.
2022 saw a number of generative AI tools such as ChatGPT explode on the scene with rapid user growth. The off-the-shelf usability for a general audience that ChatGPT, Midjourney, and Bard provided through prompts has been a great enabler. There have been quite a few use cases with excellent results for users, given that no specific training (either for the user or for a model) is required to achieve desired results. With generative AI, now you can create new data and artifacts, which have helped its popularity go through the roof. Some of the notable use cases are:
- Research augmentation — From finding new machine learning models to generating complex algorithms to validating existing research in material science, generative AI tools are accelerating and augmenting research. Multiple instances of Concept to Clinical trials being completed in record time have already been reported by pharma companies.
- Data and content generation — The technology can generate a vast amount of synthetic training data demanded by large models including neural networks from a sample set. It can also generate content including novels, articles, poems, images, animations, and all sorts of content that require basic initial inputs. These are used in marketing communications, the music industry, business document drafting, and graphic design. Many leading platforms such as LinkedIn currently offer AI drafts of various content right from your resume, including a summary and targeted marketing. Generative AI can also create synthetic bots and personas for customer service.
- Programming automation — The technology can generate code to achieve specific results across multiple languages. This has evolved well with the likes of Microsoft Copilot entering the scene. Various other scenarios including test automation are in use with accelerated productivity and quality benefits.
- Contextual analysis and insights — Generative AI is transforming financial analysis, targeted marketing, fraud detection, and business insights. The breadth and depth of text translations is evolving with the availability of generative AI solutions. Customer service and general Q&A scenarios are also well served with personalization being a winner.
A Framework for Adoption
As mentioned previously, generative AI is transformative across a wide range of use cases and industries. But when business leaders consider embracing this technology, they need a strategic approach.
A framework as provided below will help simplify and structure the risks, investments, and approach. It starts with the planning phase (vision and strategy), leading into execution and outcome management, enabling value delivery.
1. Setting the Strategic Vision and Prioritizing the Right Use Cases
Everyone is talking about generative AI. But that shouldn’t be your reason to want to adopt it. Other choices may be relevant and cost-competitive depending on your use case. While most of the leading Generative AI solutions can handle broad needs, you will see that most of your use cases may be solved by Narrow AI (e.g. a customer support bot) in a cost and time efficient manner.
The long-term success will be in ensuring an approach covering the following:
· Understanding the existing AI capabilities (~80% of the use cases may be solved through supervised learning models) vs solutions that really need generative AI capabilities. Generative AI tools are costly to acquire and run given the depth of models and processing needs.
· Once you are sure of the use cases that need generative AI, evaluate ready solutions in the market vs. effort and benefits of building the solution yourself based on existing LLMs. Building a custom solution from the ground up is a big commitment and requires longer term efforts, resources, and infrastructure. An understanding of the existing solutions will also enrich your understanding and improve your outcomes in the long run.
· There are Proprietary LLMs and open-source models. The proprietary ones from leading providers are much more evolved at various tasks; however, using open-source models gives you unlimited access to customize the model.
· Off-the-shelf approach vs. custom/fine-tuning your models. You will find that the generative AI tools have ready capabilities off the shelf to address your needs. However, for your enterprise to be able to use them, configurations around scale, constraints and security will need to be satisfied. Model fine-tuning is a great way to optimize the solution outcomes.
· Augmenting vs. replacing existing products/features. With this technology already garnering attention, there will be a lot of demand to use it to innovate existing products. A key aspect to keep in mind is to stage the transition. The best approach is to augment existing products with new capabilities from generative AI (e.g.: you can automate the workflows, dashboard, and insights in an existing CRM solution and keep it, providing a continuity in experience for users and the business while benefiting the business). This enables you to approach it from a Pareto principle, delivering the best value.
· Identifying and balancing the key risks will ensure value and safeguard business interests. These risks include data protection, compliance, expanding regulations, inherent biases from training data, hallucinations, and the lack of transparency around black box models.
Setting up a strong vision that builds on your organization’s larger strategic direction and balances people, processes, and technology aspects is a necessity. Familiarizing your employees to this technology and ensuring contributions across the board will provide the right direction to the organization.
How to Identify the Right Use Cases for Generative AI ?
Identifying the specific scenarios and use cases that can benefit from generative AI is going to be the fulcrum for business value generation. You must leverage your team’s experience to brainstorm and identify the key pain points that require a technology like generative AI (ie, generative capabilities). However, the newly available computing power and broad training data are useful for a wide range of other scenarios too. Experimentation and research by subject matter experts in each business function is a great way to start. Given most generative AI tools work through prompts, the organization should attempt to ready Proof of Concepts for the most valuable use cases in each function.
Approaches including Pareto and RICE will help focus on value that each use case can deliver. An ROI mindset along with balancing various risks in play with generative AI will help choose the top areas you can invest in. This will set you free for playing to the strengths of generative AI tools. A service partner or advisor with relevant industry experience implementing similar solutions will be invaluable.
2. Strategic Execution
The core of the strategy and the execution will be the choice of the model and the generative AI provider.
Various players have been competing for leadership within the generative AI space. The truth is each solution has its positives and negatives. Some are great at content generation and artistic outcomes but fail when it comes to data analysis and facts. Others might be great at summarizing but find it difficult to change the output with context or hallucinate more than the others. The providers are constantly improving their LLMs and it helps to keep an eye on progress. There are various benchmarks that the LLMs are measured against on their efficacy. Combining factual eval, human eval, MT bench test, and MMLU scores can provide a good comparison of the models.
Note: With fast evolving landscape the above details may need updation
Most LLMs let you fine tune to yield optimum results. The best approach will always be to choose an LLM that’s right for the need and the solution. An MVP approach and iterative development is recommended to ensure ROI is always in focus. This is the way generative AI tools themselves have evolved.
3. Data Strategy
Over the past decade, organizations have moved closer to a cloud-native infrastructure and have matured data protection practices. This is going to serve as a base for how you will be able to leverage advances in generative AI. Ensure you understand the maturity level of your organization and enable the infrastructure policies accordingly.
To implement a robust data architecture and cloud-native approach requires a dedicated team of data architects and data engineers. A wholesome data strategy will encompass storage, access-authentication-authorization (AAA), encryption (in motion and at rest), scale, and performance. For applications involving generative AI, utilizing a full stack provider of the likes of AWS, Azure, or Google will help keep the exercise simple.
One key aspect to also seal is how LLMs use your data and the learnings from it. For enterprises that prioritize keeping data within their walls, such as banks, a privately hosted model within your data center or private cloud is the way to go.
The Full Stack Providers
Though your business may not have rigid constraints around data, it’s always good to use a service like Open AI Enterprise, Microsoft Azure, Google Vertex AI, or AWS Bedrock that can provide data management and scale. These full stack providers help you to roll out your generative AI solution, which can include performance and scale, global deployment, access to various LLMs, ready components/services (e.g.: search, chatbots), security, and compliance management over and above your data management requirements. The top-notch data architecture and compliance will come in handy sometime during the product lifecycle.
4. Change Management and Accelerating Adoption
AI and especially LLMs are expected to transform the workplace. Businesses adopting these technologies should ensure a capable workforce. Appropriate training and orientation, starting with low risk enhancements that demonstrate value and enable your employees to contribute is a good starting point. Akin to any large organizational change, these should be managed carefully and with a structured plan. Anything from how the organization generates and uses data to how individuals set their goals will be impacted. Hence, a cross-functional team will be valuable to plan and execute the change.
The key message is that businesses are not thinking about “if,” but “when.” Employees can be champions in helping accelerate the adoption and usage of this technology with the right motivation. Communication, training, and enablement can lead the way.
The human in the loop practice is a great way to reassure employees and clients, at the same time securing outcomes against errors and biases.
5. Business ROI and Value
Strategic business decisions should be based on ROI and implementation prioritized basis value delivery. The basic business sense should prevail for all projects including those for generative AI. This is an enormous, high-investment exercise that spreads across the organization and will transform your own operating model. It is to be noted that how you deal with clients and competition will see extensive change as well. Achieving success is hard given the unpredictability of technological developments and emerging regulatory risks.
A series of steps aimed at ensuring the accuracy of results will help your organization deliver value faster. This includes using capabilities from different models if necessary and intense testing along the following lines:
- Feedback loops (like Viable) for reinforcement learning
- Clear metrics and reporting for continued monitoring
- Dedicated QA resources
- Testing frameworks such as pytest (Python) and A/B testing
- Adversarial models
- Explainable AI methods including LIME, SHAP (limited applicability), ELK Stack/Splunk
- Sensitivity and red teaming (bias audits)
- Innovative approaches that use multiple models (e.g.: language translation done by GPT and audited by Bard)
The value will also evolve the overall operating model or specific functions of the business. A focus on ensuring improved productivity and quality of your employees by moving them up the value chain is an essential aspect of value delivery.
Conclusion
While Generative AI and LLMs present transformative potential, their adoption should be thoughtful and well-strategized. With the above approach you will be able to lead the way in harnessing generative AI’s power effectively. The framework can be expanded to your specific needs. An experienced technology partner specializing in your industry will be invaluable in successful outcomes.
