Why you need an AI factory: A CIO’s guide to generative AI

1. Don’t be tentative in adopting GenAI — start fast and at scale

There’s a great way to achieve significant ROI quickly from generative AI. Take advantage of how a single, repeatable “pattern” of generative AI training and deployment can apply across your value chain. With conventional AI, you may have needed a different, custom-built AI model for every use case. With GenAI, a single pattern of deployment may be able to help produce or troubleshoot software code in many functions or lines of business.  

Since generative AI models come “pre-trained,” requiring only adaptation and customization, you can often conduct a proof of concept quickly or even go straight to a pilot. Naturally, you won’t roll out GenAI everywhere all at once. You’ll start with a few use cases and learn from mistakes and successes. But with the right preparation, you can move far more quickly than with conventional AI. A 90-day “sprint” is often enough to get initial use cases up and running and ready to scale.

To create a roadmap for generative AI deployment, bring together technology and business leads. As you identify use cases, assess their value and look for patterns that can be replicated to deliver value elsewhere. 

Identify data, technology and skills gaps and then determine the cost and timeframe for closing these gaps. Common gaps may include: 

• Insufficient, poorly organized and hard-to-access data sets.
  
• A need for more cloud-based computing power.
  
• Cloud engineering capabilities and new APIs.
  
• Lack of both specialized data science skills and broad-based knowledge of responsible AI practices. 
  

Your roadmap should also include trust-by-design — a plan to embed responsible AI from day one into the technology and processes needed to validate results, monitor ROI and manage and mitigate risks now and in the long term.

2. Select your model — and prepare your data

Several specialized companies, including the major cloud platform providers AWS, Google and Microsoft, as well as OpenAI, offer generative AI “foundation models” — deep learning algorithms pre-trained on massive amounts of public data. Once selected, you’ll then want to work with your cloud provider to establish a private version of the model within your firewall.

You can, for example, have your most experienced specialists embed your own data and IP to be used by a GenAI model. Then everyone in your organization will have access to their expertise — as if GenAI were providing them with personalized guidance from a mentor. 

As you upgrade internal technology and cloud capabilities for GenAI, pay special attention to data. Generative AI can make sense of unstructured data, but you may still have to cleanse and organize at least some of it — mitigating, where appropriate, the risks of bias in data sets. You’ll likely also need to create data pipelines, complete with new APIs, to provide generative AI with continual access to the most up-to-date data available. To safeguard your data and intellectual property, you may also need to upgrade both data governance and cybersecurity for GenAI’s potential risks.

3. Stand up a generative AI factory for fast, repeatable results

To make deployment and execution fast and repeatable, build an AI factory — an operating model built on pods to identify and assess use cases in a selected domain or line of business — and then adapt the foundation model to deliver value. Each pod contains six roles.

• Pod leader: Determines and monitors overall business objectives and value of each potential use case.
  
• Business analyst: Provides detailed breakdown of use case objectives and success criteria. Gathers and confirms requirements. Sets and monitors weekly deliverables. Should have experience in the business process that GenAI will augment or automate.
  
• Prompt engineer: Designs prompts to customize the model’s outputs for greater accuracy and relevance. Works with business analyst and model mechanic to validate outputs, then adjusts prompts as needed.
  
• Model mechanic: With help from prompt engineer and business analyst, customizes the AI model to further enhance results. Will “own” the technical side of the solution after deployment begins. 
  
• Data engineer: Accesses, prepares and organizes data and embeddings to meet the AI model’s requirements.
• Data scientist: Informs work of the prompt engineer and model mechanic to maximize accuracy and performance. Divides the code used into components to make it easier to deploy in other solutions or future updates.

But GenAI’s remarkable scalability can help ease the demands on your workforce too. Technology specialists (such as model mechanics and data scientists) can usually work many pods at once. As we do at PwC, internally and with clients, you can also use pre-built toolkits with ready-to-use software code and prompts. These toolkits can help speed up the deployment and reduce the costs of GenAI for things like personalized customer experiences, content creation, research, agile software delivery, support services, report generation, deep retrieval, smart summaries and Q&A engines.