10+ Use cases & LLM Best Practices – TechToday

Generative AI, also called GenAI, presents novel opportunities for enterprises compared to middle-market companies or startups including:

However, generative AI is a new technology with unique challenges for enterprises:

• Valuable proprietary data can be exposed to competitors

• Generative AI tools will create new services and solutions. Technology leaders can leverage them to enter new markets gaining market share at the expense of incumbents.

• Generative AI models, also called generative models, will bring new automation opportunities with the potential to increase customer satisfaction or reduce costs. Competitors can leverage them to get ahead.

• Reputational or operational risks due to generative models’ bias or hallucinations

Executives wonder how their organizations can reap the benefits of generative AI while overcoming these challenges. Below, we outline:

• Generative AI use cases for large companies

• Guidelines to leverage the full potential of generative AI including how to build and manage your company’s generative models.

How should enterprises leverage generative AI?

We charted a detailed path for businesses to leverage generative AI.

While most firms may not need to build their models, most large enterprises (i.e. Forbes Global 2000) are expected to build one or more generative AI model specific to their business requirements within the next few years. This will allow them to customize model output in detail for their own domain, generating higher levels of accuracy. We have already seen firms like Bloomberg generating world-class performance by building its own generative AI tools leveraging internal data.

What are the guidelines for enterprise generative models?

At a minimum an enterprise generative AI model should be:

Consistent

Most current LLMs can provide different outputs for the same input. This limits the reproducibility of testing which can lead to releasing models that are not sufficiently tested.

Controlled

Be hosted at an environment (on-prem or cloud) where enterprise can control the model at a granular level. The alternative is using an API like OpenAI’s LLM APIs. The disadvantage of relying on APIs is that the user may need to expose confidential proprietary data to the API owner. This increases the attack surface for proprietary data. For example the API provider or bad actors working at the API provider:

Ethically trained

Model should be trained on ethically sourced data where Intellectual Property (IP) belongs to the enterprise or its supplier and personal data is used with consent.

• Generative AI IP issues, such as training data that includes copyrighted content where the copyright doesn’t belong to the model owner, can lead to unusable models and legal processes.

• Use of personal information in training models can lead to compliance issues. For example, OpenAI’s ChatGPT needed to be disclose its data collection policies and allow users to remove their data after the Italian Data Protection Authority (Garante)’s concerns.

Explainable

Unfortunately, most generative AI models are not capable of explaining why they provide certain outputs. This limits their use as enterprise users that would like to base important decision making on AI powered assistants would like to know the data that drove such decisions.

Fair

Bias in training data can impact model effectiveness

Licensed

The enterprise should have a commercial license to use the model. For example using models like Meta’s LLaMa have noncommercial license preventing their legal use in most use cases in a for-profit enterprise. Models with permissive licenses like Vicuna built on top of LLaMa also end up having noncommercial licenses since they leverage the LLaMa model.

Secure

Enterprise-wide models may have interfaces for external users. Bad actors can use techniques like prompt injection to have the model perform unintended actions or share confidential data.

Sustainable

Training generative AI models from scratch is expensive and consumes significant amounts of energy, contributing to carbon emissions. Business leaders should be aware of the full cost of generative AI and identify ways to minimize its ecological and financial costs.

Enterprises can strive towards most of these guidelines and they exist on a continuum except the issues of licensing, ethical concerns and control.

• It is clear how to achieve correct licensing and to avoid ethical concerns but these are hard goals to achieve

• Achieving control requires firms to build their own foundation models however most are not clear about how to achieve this:

How can enterprises build foundation models?

There are 3 approaches to build your firms’ LLM infrastructure on a controlled environment:

1. Build Your Own Model (BYOM): Allows world-class performance costing a few million $ including computing (1.3M GPU hours on 40GB A100 GPUs in case of BloombergGPT) and data science team costs.
2. Fine-tuning is a cheaper machine learning technique for improving the performance of pre-trained large language models (LLMs). Instruction fine-tuning was previously done with a large dataset but now it can be achieved with a small dataset (e.g. 1,000 curated prompts and responses in case of LIMA) and limited training costs compared to BYOM.
3. Reinforcement Learning from Human Feedback (RLHF): A fine-tuned model can be further improved by human in the loop assessment.

Which models should enterprises use to train cost-effective foundation models?

Machine learning platforms released foundation models with commercial licenses relying mostly on text on the internet as the primary data source. These models can be used as base models to build enterprise large language models:

• BLOOM by Huggingface with RAIL license.
• Dolly 2.0 instruction-tuned by Databricks based on EleutherAI’s pythia model family.
• Open source RWKV-4 “Raven” models
• Eleuther AI Models

What is the right tech stack for building large language models?

Generative AI is an artificial intelligence technology and large businesses have been building AI solutions for the past decade. Experience has shown that leveraging Machine Learning Operations (MLOps) platforms significantly accelerate model development efforts.

In addition to their MLOps platforms, enterprise organizations can rely on a growing list of Large Language Model Operations (LLMOps) tools and frameworks like Langchain, Semantic Kernel or watsonx.ai to customize and build their models.

In early days of new technologies, we recommend executives to prioritize open platforms to build future-proof systems. In emerging technologies, vendor lock-in is an important risk. Businesses can get stuck with outdated systems as rapid and seismic technology changes take place.

Finally, data infrastructure of a firm is among the most important underlying technologies for generative AI:

1. Vast amounts of internal data need to be organized, formatted.

2. Data quality and observability efforts should ensure that firms have access to high quality, unique, easily-usable datasets with clear metadata.

3. Synthetic data capabilities may be necessary for model training

What are alternatives to controlling models?

Enterprise organizations can leverage pre-trained and fine-tuned models from tech giants or AI companies (e.g. OpenAI) in cases where they are:

• Not concerned about increasing attack surface of the input data or

• Confident that their inputs will not be intercepted by 3rd parties or stored or

• Confident that even if their inputs are stored, they are stored for a limited time and will not be leaked while stored

In such cases, technology teams can use APIs to access models at affordable costs per API call. They can use these approaches:

Zero-shot learning, also called prompt engineering, involves structuring the prompt to help improve the LLM output

Few-shot learning, also called in-context learning, involves adding examples before the prompt to improve response quality.

This can also include chain-of-thought prompting.

Fine-tuning is also available to further improve model performance.

What should enterprises do about generative AI before building their foundation models?

Building your enterprise model can take months since the steps below need to be completed. Each of these steps can take weeks to months and they can not be fully parallelized:

• Data collection can take weeks to months. Sponsored: Data collection providers like Clickworker can accelerate this process helping companies generate balanced, high quality instruction datasets and other data for building or fine-tuning models.

• Hiring data scientists with LLM expertise or hiring consultants can take weeks to months.

• Training and deployment

• Integrating models to business systems

Therefore, we recommend business leaders to encourage experimentation. GenAI requires a paradigm shift: Our understanding of machines need to evolve from senseless robots to co-creators. Organizations need to start working with GenAI to start this mindset shift.

Teams can leverage existing APIs to automate processes in domains where value of confidential data is lower and system integration is easier. Example domains where teams can leverage GenAI to improve productivity and increase teams’ familiarity with generative AI without building own models:

• New content creation and optimizing generated content for marketing campaigns

• Code generation for front-end software

• Conversational AI for customer engagement and support

• There are tens of more generative AI applications

What are enterprise generative AI use cases?

The web is full of B2C use cases such as writing emails with generative AI support that don’t require deep integration or specialized models. However, enterprise value of generative AI comes from enterprise AI applications listed below:

Common use cases

Enterprise Knowledge Management (EKM): While SMEs and mid-market firms do not have challenges in organizing their limited data, Fortune 500 or Global Forbes 2000 need enterprise knowledge management tools for numerous use cases. Generative AI can serve them. Applications include:

1. Insight extraction by tagging unstructured data like documents

2. Summarization of unstructured data

3. Enterprise search

Part of enterprise includes answering employee questions about:

1. Company’s practices (e.g. HR policies)

2. Internal company data like sales forecasts

3. A combination of internal and external data. For instance: How would potential future sanctions targeting MLOps systems sales to our 3rd largest geographic market affect our corporate performance?

Larger organizations serve global customers and machine translation ability of LLMs are valuable in use cases like:

1. Website localization

2. Creating documentation like technical manuals at scale for all geographies

3. Multilingual customer service

4. Social media listening targeting a global audience

5. Multilingual sentiment analysis

Industry specific applications

Most enterprise value is likely to come from using generative AI technologies for innovation in companies’ specific industries: This could be in the form of new products and services or new ways of working (e.g. process improvement with GenAI). Our lists of generative AI applications can serve as starting points:

What is generative AI?

Clarifications: Generative AI includes text, image and audio output of artificial intelligence models which are also called large language models LLMs, language models, foundation models or generative AI models.

External Links