June 7, 2023

Taking the First Steps Toward Enterprise AI

By Grant Henke

In Short

It is clear that AI is being used to power up businesses now more than ever, and companies that don’t use AI risk being outcompeted by those who do. The best AI-powered products are fueled by a diverse collection of high-quality data. The most critical and impactful step you can take towards enterprise AI today is ensuring you have a solid data foundation built on the modern data stack with mature operational pipelines, including all your most critical operational data.

phData can help with this foundational strategy and platform along with your AI building needs by utilizing our teams of data engineering, analytics, and AI experts.


In the face of an unprecedented surge in consumer Artificial Intelligence (AI) platforms – think ChatGPT, DALL-E, Whisper, Stable Diffusion, MidJourney, and Bard – and the tidal wave of products riding on similar models, such as AI-powered Bing, Adobe Firefly, GrammarlyGo, Github Copilot, and Amazon CodeWhisperer, it’s clear that we’re in the midst of an AI revolution. Add the bustling activity in community hubs like Hugging Face, and the situation becomes even more formidable.

It’s no surprise that every boardroom, small business, and ambitious garage startup recognizes the immense potential of AI and is trying to understand how to get started in this daunting and rapidly evolving space.

Once you consider that all of your competitors are also trying to leverage AI, the race to harness the power of AI is truly on! But how do you take the first steps toward implementing AI? This can be intimidating, especially for those new to this fascinating field, and we get it. Having guided hundreds of organizations through data modernization initiatives, we know what it takes to properly wield the power of AI for eye-opening results.

Whether you’re a seasoned professional or a curious beginner, this blog serves as a guide to kickstart the integration of AI into your enterprise.

Before we dive into how to take the first steps toward enterprise AI tangibly, it’s important to understand the history and terminology of this AI revolution.

A Brief History of AI

It may seem like AI came out of nowhere with recent advancements and attention. However, the history of AI dates back to the mid-20th century, with pioneers like Alan Turing, John McCarthy, and Marvin Minsky laying the groundwork for the field. Early AI research focused on rule-based systems, known as expert systems, which mimicked human decision-making in specific domains.

Research in artificial neural networks – now the foundation of modern AI – began around 1950, but progress stalled for decades due to challenges in training them to solve meaningful problems.

Modern AI, on the other hand, is built on machine learning and artificial neural networks – algorithms that can learn their behavior from examples in data. As computational power increased and data became more abundant, AI evolved to encompass machine learning and data analytics. This close relationship allowed AI to leverage vast amounts of data to develop more sophisticated models, giving rise to deep learning techniques.

Today, AI and data analytics are deeply intertwined, with AI playing a crucial role in extracting valuable insights from massive datasets and driving innovation across industries.

Large Language Models (LLMs) Timeline

Even LLMs like ChatGPT, creating all of the buzz in the market today, are built on the shoulders of giants. Years and years of iterative development from commercial and academic experts worldwide have edged them toward where they are today. The diagram below highlights the evolutionary tree and timeline of just this tiny part of the AI ecosystem.

Terminology & Acronyms

Over AI’s growth, a sprawling ecosystem with a wide range of applications and many different models to support them has been created. As a result, like many technology domains, it is also filled with complex terminology and acronyms. Recognizing some standard terms as you enter the AI world can help you confidently navigate the vast evolving landscape.

Note: This section is long but important. Feel free to quickly scan or skip ahead and use it as a reference when needed as you read on.

First Steps Towards Enterprise AI

Artificial Narrow Intelligence (ANI) or Weak AI: ANI is designed to perform specific tasks or solve particular problems within a limited domain. It is task-oriented and cannot understand or apply knowledge beyond its designated tasks. Examples include chatbots, recommendation systems, and image recognition systems.

Artificial General Intelligence (AGI) or Strong AI: AGI refers to AI systems with human-like intelligence, understanding, and problem-solving abilities across various domains. It can learn, reason, and adapt to new situations like humans. AGI is still a theoretical concept and has not yet been realized.

Generative AI: Generative AI is a subfield of artificial intelligence that focuses on creating new content, data, or patterns by learning from existing data. It employs algorithms, often based on deep learning techniques, to generate outputs such as images, text, music, or even 3D models. Generative AI can be used to synthesize realistic and creative outputs, enabling applications like art generation, content creation, and data augmentation.

Data Science: Data science plays a crucial role in the development and application of AI, as it involves preprocessing, exploring, and transforming data to create high-quality datasets for training AI models. Data scientists use data-driven approaches to enable AI systems to make better predictions, optimize decision-making, and uncover hidden patterns that ultimately drive innovation and improve performance across various domains.

Machine Learning (ML): ML is a subset of AI that focuses on developing algorithms that allow computers to learn from and make predictions or decisions based on data. It includes techniques like supervised, unsupervised, and reinforcement learning.

Model: In the context of AI, a model is a computational representation of a system or process designed to predict outcomes or behaviors based on certain inputs. AI models, often based on machine learning algorithms, are trained using data to learn patterns, relationships, or decision rules. They adjust their internal parameters during training to minimize the difference between their predictions and the actual outcomes. Once trained, these models can make predictions, classify information, recognize patterns, or make decisions when presented with new, unseen data. AI models can range from simple linear regressions to complex deep neural networks.

Deep Learning (DL): DL is a subfield of ML that uses artificial neural networks, particularly deep neural networks, to model and solve complex problems. DL is particularly effective in processing large amounts of unstructured data, such as images, audio, and text.

Natural Language Processing (NLP): NLP is a branch of AI that deals with the interaction between computers and human languages. It focuses on enabling machines to understand, interpret, and generate human language in a meaningful and contextually relevant way.

Natural Language Understanding (NLU): NLU is a subset of NLP focused on algorithms that can interpret the meaning of a sentence or document in terms of syntax, grammar, or ontology. Examples of NLU tasks include parsing grammar or parts of speech to score a paper assigned as homework or Named Entity Recognition (NER) to identify specific concepts within a large glossary of terms.

Large Language Model (LLM): LLMs are AI models designed to understand and generate human-like text by learning from vast amounts of textual data. They are based on deep learning architectures, such as transformer networks, and trained on millions or billions of parameters to capture complex linguistic patterns, context, and semantic relationships. These models can perform a wide range of natural language processing tasks, including machine translation, sentiment analysis, text summarization, and conversational AI, demonstrating impressive fluency and comprehension capabilities.

Tuning/Training/Learning: In the context of AI, tuning, training, or learning refers to the process of developing an AI model by exposing it to a set of input data, often labeled with the desired outputs or outcomes. During this process, the model adjusts its internal parameters to minimize the difference between its predictions and the actual or target outputs. The training goal is to enable the AI model to recognize patterns, generalize from the training data, and make accurate predictions or decisions when presented with new, unseen data. This learning process can occur through various techniques, including supervised fine-tuning (SFT), supervised/unsupervised learning, and reinforcement learning (depending on the problem and available data).

Reinforcement Learning with Human Feedback (RLHF): RLHF is a method that combines reinforcement learning and human feedback to train AI models. In this approach, an AI agent learns to make decisions by interacting with an environment and receiving feedback from human experts or evaluators. Human feedback helps guide the learning process by providing valuable insights, preferences, or corrections, which the AI agent can use to improve its decision-making capabilities.

Generative Pre-trained Transformer (GPT): GPT is a type of LLM developed by OpenAI. It is based on the transformer architecture, which utilizes self-attention mechanisms to process and generate text in a parallelized and context-aware manner. The pre-training technique uses vast amounts of textual data from the internet to help models learn and capture complex language patterns, context, and semantics. A GPT model can perform various natural language processing tasks, such as text completion, summarization, translation, and question-answering. GPT models excel so strongly at answering questions that they have been widely adopted for Zero-Shot and Few-Shot learning, where well-engineered prompts can be used to create domain-specific AI without large volumes of training data.

Prompt Engineering: Prompt engineering refers to the practice of carefully crafting the input text (or “prompt”) to guide a model toward generating a desired output. The quality and nature of the input prompt can significantly influence the model’s response. Through skillful prompt engineering, users can maximize the utility of AI language models for various tasks, such as text completion, text generation, text classification, sentiment analysis, and more. It’s a way of indirectly “programming” these models to perform desired tasks without changing the underlying model or its parameters.

Vector Database:vector database is a specialized database designed to efficiently store, manage, and retrieve high-dimensional vectors, also known as vector embeddings. These vectors often represent complex data such as images, videos, audio, and text in a numerical format suitable for machine learning and AI applications. Vector databases support similarity search operations, allowing users to find vectors most similar to a given query vector. This capability is essential for many AI applications, including recommendation systems, image recognition, and natural language processing.

Deeper Learning (pun intended)

Equipped with this terminology, you can consume the vast expanse of content published on AI with a basic understanding of what is being discussed. If you want to go one level deeper, this blog titled “What Is ChatGPT Doing … and Why Does It Work?” by Stephen Wolfram is a great “starting from square one” walkthrough of how AI like ChatGPT works.

Chip Huyen also wrote an outstanding blog on RLHF, a key technological reason ChatGPT and AI like it has suddenly improved dramatically. Both are very detailed but worthwhile reads for anyone interested in equipping themselves with background knowledge of how AI models and products work.

AI Today

ChatGPT was released on November 30, 2022, and has since been coined the fastest-growing app in internet history. Since then, its meteoric rise has pushed tech giants like Google, Microsoft, and Facebook, along with a collection of tech startups, to accelerate their own AI projects lying in wait.

The result is overwhelming change as companies race to harness AI’s transformative force. So much so that the cutting-edge technology of last week often feels outdated, and it can seem impossible to keep up.

It isn’t just the tech giants and startups accelerating their AI focus. In a recent survey, 65% of 225 U.S. C-suite and senior executives believe generative AI will greatly impact their organization in the next three to five years, far above every other emerging technology. However, fewer than half of respondents say they have the right technology, talent, and governance to successfully implement generative AI today.

In times of significant technological change like this, and dare I say “hype,” it can be overwhelming and may feel more manageable to disengage than dive in headfirst. Instead, a middle-ground approach for handling times like this is to step back, observe the overall patterns, and “see the forest for the trees.”

Acknowledge that picking industry and technology winners will be hard or impossible, but that it is essential to engage, prepare, and experiment at an appropriate pace as the technology and market evolve.

First Steps to Enterprise AI

The message is clear. People who don’t use AI will be outcompeted by those who do. So, what are the concrete steps you can take toward enterprise AI? We can now answer the question with the background knowledge of the AI ecosystem, phData’s experience, and the context of today’s landscape. Below are the key steps we recommend.

1. Build a Strong Data Foundation

You can’t build a powerful AI application without a strong data foundation and a modern data platform. Every successful AI product has made it clear that the majority of the work and a key factor for their success is the highly curated and diverse collection of high-quality data that feeds it.

For example, Sam Altman, the CEO of OpenAI, highlighted on the Lex Fridman Podcast, “A lot of our work is building a great dataset.” He goes on to say how a wide array of data is used and combined to build the training data set required. That act of data collection and enrichment is enabled and accelerated by a modern data stack.

You may think that AI is only for tech giants with massive budgets, colossal data sets, and a collection of proprietary technology. However, a leaked internal Google document highlights that anyone can succeed in AI and that the moats that prevented competition before no longer exist.

You don’t need massive data sets because “data quality scales better than data size.” Small models with good data are better than massive models because “in the long run, the best models are the ones which can be iterated upon quickly.” And you don’t need to depend on expensive proprietary models because even Google can’t compete when “open source has some significant advantages that they cannot replicate.”

The key moats that remain in the long term are proprietary data combined with third-party data, understanding the use case deeply by “being your own customer,” and the ability to develop quickly, sustainably, and scalably. That is where the modern data stack, including the Snowflake Data Cloud, is critical.

Why Snowflake is Good for AI

Snowflake’s key cloud-native features provide significant operational capabilities for companies to gain an AI advantage. The list of advantages is long, but below are a few of the most important:

  • Data Sharing & Marketplace: Because high-quality data is a key to good AI, the value of a broad marketplace of data sets that can be easily used to enrich your own proprietary data seamlessly can not be understated. Additionally, the value of your internal data as a product for others to use in that marketplace itself could be an interesting and lucrative opportunity. 
  • Native Cloud & Web API integrations: Snowflake users can use cloud and web-hosted LLM APIs using external functions and Streamlit as an interactive front end for LLM-powered apps. Some published examples include a GPT-3 Dataset GeneratorAI assistant, and MathGPT. We have also published blogs on using Snowflake to power AI-generated marketing emails and talking to your data with AI
  • Elastic Scalability, Performance, Cost, and Operations: Because Snowflake can scale vertically and horizontally to fit all of your needs without impacting other users and be scaled down to zero easily when the work is done, you can experiment using all of your data and resources safely without a long term commitment and at a predictable and reasonable expense. When it comes time for production, the benefits continue as managing pipelines on Snowflake is far simpler than a legacy data platform.
  • Security and Governance: Snowflake intensely focuses on the security and governance of data. It enables users to bring the power of LLMs to the data without compromising security or governance. This is particularly important when dealing with sensitive or proprietary data such as source code, personally identifiable information (PII), internal documents, wikis, code bases, and other sensitive data sets.
  • Key Acquisitions & Innovation: Snowflake is focusing deeply on AI and has recently acquired Applica, the Document Visual Question Answering Challenge winner, with a multi-modal LLM for document intelligence and Neeva to accelerate search in the Data Cloud through generative AI. And I have no doubt there will be a slew of exciting feature announcements and the Snowflake Summit event at the end of June as their pace of innovation has continued to impress.

The best part of this step is that focusing on building a strong data foundation and operational maturity around data pipelines will not only help prepare you for AI success but is also a critical step for more traditional analytics maturity and becoming a more data-driven organization. This foundational step will provide tremendous value to the organization even if AI has yet to become a part of your long-term strategy.

2. Review & Define Internal Policies

Like all extremely powerful tools, where there is a tremendous opportunity, there is also a potential for immense risk and misuse. As a result, protecting your business, employees, and customers is essential as AI use and tools become more prominent. This is particularly true regarding data privacy, security, and ethical guidelines. 

An AI policy can help your company prepare for these AI advancements, manage potential risks, and maximize the technology’s benefits.

There is an abundance of articles on the importance of AI policies, overviews of the security risks to consider (like those highlighted by Microsoft’s Chief Security Advisor), and public examples of terms of use policies that can be helpful.

Each business is unique, and the exact contents and policies that should be defined will undoubtedly vary. That said, below are some key factors and risks you may want to consider when drafting an internal policy:

  • Security Risks: AI poses a whole new class of risks, including more powerful phishing attacks, even going so far as to clone voices, a new class of injection attacks called prompt injection, and new and complex types of malware that can be generated more quickly than ever and avoid conventional protection measures. Additional employee training and guidance may also be needed to help protect against these new, more powerful attack vectors. 
  • Incorrect Information: Generative AI is known to “hallucinate,” always producing an answer and potentially even claiming that the answer is true and well-sourced when it is entirely made up. Mindlessly trusting the output without doing your research and due diligence is a recipe for disaster. Sam Altman has said, “That is the current biggest single problem with language models…they are very convincing bullshitters”. Torsten Grabs, Senior Director of Product Management at Snowflake, also provided the guidance “We want to make sure that we have oversight in the loop. We should use this [AI] as a tool to make people more productive and accelerate their work. We should be very careful to completely automate the human away.” For that reason, you should consider encouraging users to verify the output of these models because when things go wrong, you won’t be able to blame the AI.
  • Legal and Regulatory Compliance: Building on the Incorrect Information section above, if you let it, AI could inadvertently do things that land the company in legal or regulatory compliance trouble. Ensuring a policy provides the necessary guardrails and reminds users of the critical compliance requirements that must be followed is very important. 
  • Data Ownership: The law around the ownership of data input to and output from AI models is not well defined today, but the law is moving quickly to catch up. Addressing how you expect to maintain rights over the content or proprietary information you create using AI and protecting your intellectual property and proprietary from competitive AI is essential even while the rules are still being defined. 
  • Ethics: Defining a clear policy with clear guidelines for the ethical use of AI is a challenge that everyone is grappling with, including governments. Though it’s hard, creating a plan that details what is and is not permitted from an ethics standpoint will be an essential safeguard to keep AI use ethical.
  • Customer Policies & Requirements: If you are a B2B business working with privileged access or customer information, you need to ensure that your policies are compatible with theirs and that you are compliant with both your internal and customer requirements. If the customer doesn’t have a policy, it would be a good idea to recommend they create one.
  • Update Intervals: AI is expected to evolve at breakneck speeds, so the policy you create today could quickly become outdated. A well-designed plan should include regular intervals to reevaluate your policy to ensure it is still practical and up-to-date. 

Given the broad scope and potential for impact, drafting a policy can be daunting. However, policies don’t need to be perfect and should not be set in stone. Pulling together company leadership and creating a conservative yet clear initial draft can serve as the foundation for regular updates as the market evolves and AI use within your business matures. In this case, something is better than nothing, and perfect is the enemy of good.

3. Learn & Experiment

AI is an evolving and complex field with ever-changing limitations and capabilities. To identify the suitable AI applications that could benefit your business, you will need a deep understanding of what applications are valid use cases based on current limitations and what is just a “parlor trick” or demo of a potential but currently unproven future that will require significant time, investment, and innovation to pull off.

A healthy balance of optimism and skepticism is vital as you move forward in AI, and the best way to avoid confirmation bias is to look for content and experiments that will both confirm and deny your existing assumptions. Finding blogs like Google’s “Debunking five generative AI misconceptions” can help to delineate marketing from reality.

However, the best way to separate marketing and FUD from fact is to try it yourself and get real hands-on experience.

Experimentation doesn’t need to be a large or expensive undertaking. Starting small using the freely available AI products, tools, and libraries and slowly going deeper can provide the right balance of knowledge and value relative to the time invested. There are likely many curious and driven folks in your organization doing this already, and encouraging and supporting them is a great way to slowly build an understanding of how AI can work for your business. Providing a shared space like a Slack channel or weekly meeting to share what’s working and what is not can further amplify the value of these experiments and encourage others to participate.

Eventually, experimentation should evolve into developing proof of concept (POC) projects. Designing a small-scale project or POC, including success criteria and testing the feasibility and effectiveness of your chosen AI solutions. Our blogs on AI-generated marketing emails and talking to your data with AI are great examples of POC projects you can do yourself.

With a collection of POCs under your belt, you will be well equipped to choose the right AI technology, identify known risks and challenges, and better understand the required budget and resources when moving on to developing an AI strategy and choosing production use cases. You will also have a clearer picture of the team and the skills required to be successful in AI.

4. Develop an AI Strategy

An AI strategy will be different from business to business. For some, the primary goal may be to power up users across the business internally and drive efficiency and productivity using AI products and tools. For others, it may be to enhance an existing product using AI technology or to build an AI-centric product or platform. What is most important is to have a clear strategy, regardless of what the strategy is.

good strategy will align with business, company, and technical realities guiding tradeoffs and explaining the rationale behind that guidance, evolving as more context and patterns are discovered. Defining a strategy acts as a forcing function that requires you to think critically about what is essential and how to prioritize it. It helps you to face challenges and constraints head-on and to come up with realistic plans and solutions.

Once defined, it serves as a reminder and guide of principles and priorities to drive alignment across the company, ultimately leading to more effective work and a higher likelihood of success. It’s a compass you can use to check that incremental decisions will ultimately lead to your desired outcomes or that it is time to change routes.

complete strategy will consist of at least these three parts:

  • Diagnosis: What’s the current status?

  • Vision: What is your desired end state? Where are you trying to go? What does it look like to be ‘good’?

  • Plan: What are the highest impact steps to reach your desired end state? What is the use case or project you need to take on first to achieve your vision?

It may feel like AI is moving so fast you don’t have time to develop a detailed strategy, but slowing down now will help you to be much faster and more effective as you move forward. You don’t need to pause anything while developing the strategy.

Developing the strategy in parallel with experimentation and POC development will help to iteratively understand your organization’s goals, challenges, and desired outcomes from AI implementation and establish a comprehensive AI strategy that aligns with your business objectives.

Appointing a single owner or team to drive the definition and maintenance of your AI strategy across the company now will lead to long-term success as your business embarks on its AI journey.

5. Create a Skilled Team

Depending on your investment in AI, your team might be a handful of generalists, including those that helped with the foundational steps above, or a large team of profoundly skilled specialists and domain experts. Hiring and training can be difficult, so now is the time to start building or hiring a team with the required diverse skill sets. Some of the roles on a typical AI product team might include the following:

  • AI Research Scientists: These individuals create new AI algorithms and models. They often have a deep theoretical understanding of machine learning and stay current with the latest research. They are usually responsible for creating novel approaches to AI problems.
  • Data Scientist: Data Scientists are responsible for analyzing and interpreting complex datasets. They use their mathematics, statistics, and programming skills to find patterns and trends, which can be used to train and tune AI models. They also communicate the results of these models to business stakeholders.
  • Machine Learning Engineer: Machine Learning Engineers are responsible for taking the theoretical models built by AI Research Scientists and implementing them in a way that they can be used at scale. This often involves software engineering, data engineering, and system design skills.
  • Prompt Engineer: Prompt Engineers are responsible for crafting context and prompts as input to the AI models in a way that maximizes the quality of the AI’s output​ and achieves the desired results reliably. Additionally, prompt engineers may be tasked with fine-tuning and testing AI models to ensure the AI performs within spec and does not “hallucinate” too much or provide biased responses. 
  • Data Engineer: Data Engineers are responsible for the data infrastructure. They design and build systems that allow data to be collected, stored, and processed so machine learning models can use it effectively. This often involves skills in databases, distributed systems, and ETL (Extract, Transform, Load) processes.
  • AI Operations Specialist: AI Operations Specialists (AIOps/MLOPs) ensure that AI systems run smoothly in production. They monitor system performance, troubleshoot issues, and coordinate with other teams to implement fixes. They also ensure that the systems are scalable, reliable, and secure.
  • Domain Experts: These individuals are subject matter experts who provide critical insights that enhance the performance of an AI system. They help answer questions like the most valuable insights, whether the data collected about the domain is trusted, and whether the derived insights make sense.

The specific composition of a team will vary from organization to organization. Additionally, more typical software project roles like product manager, project manager, business analyst, UI/UX engineer, software engineer, and test engineer may be required. 

These roles can overlap, and each function does not have to be fulfilled by a separate person. For example, in a smaller company, a single individual might take on the roles of a data scientist, machine learning engineer, and data engineer. In a larger organization, these roles might be more specialized.

Of course, if hiring doesn’t make sense or you would like to augment your team, phData has seasoned experts available to jump-start your AI initiatives.

phData's History of Applied AI

Over the last decade, phData has helped many companies leverage AI and machine learning to give our enterprise customers an edge over their competitors and harness the power of AI in sustainable, maintainable, and operational data applications critical to their daily business function and success.

You can learn more about some of these projects in these detailed real-world case studies:

In Closing

It is clear that AI is being used to power up businesses now more than ever, and companies that don’t use AI risk being outcompeted by those who do. There are clear steps you can take to move your enterprise in the right direction.

First and foremost, building a solid data foundation on a modern data platform is paramount.

At phData, we have years of industry experience and have delivered numerous successful AI projects. Our engineers deliver end-to-end production-ready data applications, analytics, and AI services on the modern data platform daily.

If you’re looking for additional help along your enterprise AI journey, we invite you to attend one of our free AI roadmap workshops with an expert. In a one-on-one setting, we will help you take the next step in positioning your business to succeed with your enterprise AI initiatives.

Data Coach is our premium analytics training program with one-on-one coaching from renowned experts.

Accelerate and automate your data projects with the phData Toolkit