Before you begin, check out our “Four Steps to Create an AI Readiness Assessment for Your Business”
In the rapidly evolving landscape around AI, it's easy to feel lost amidst the clamor to compete. Promised productivity boosts are tantalizing (some projecting as much as 7%, or $7 trillion uplift to global GDP), yet the roadmap to harness them remains elusive, especially for businesses. For leaders, aligning AI endeavors with business objectives, foreseeing risks, and charting a clear path forward is paramount. Now, more than ever, it's critical to plan your path and develop a roadmap. Central to this roadmap is the pivotal decision: To build or to buy?
In the previous article in this series, we walked through how to create and run an AI readiness assessment. Now that you have identified your goal, gained a better understanding of your technical readiness and cultural attitudes, and selected a business use case, you can begin setting out a roadmap for AI adoption within your company.
What is an AI roadmap?
An AI roadmap is more than just a plan; it's the compass guiding an organization's AI journey. It charts out the steps, timelines, and milestones to seamlessly integrate AI solutions. With nuances akin to other roadmaps—be it product, tech, or business—it offers clarity on AI objectives and the sequence to achieve them.
How do I know whether to build or to buy your AI solution?
The "build vs. buy" debate has been a longstanding one, and it centers around whether an organization should create custom IT solutions in-house (build) or purchase off-the-shelf products (buy). Both approaches have their advantages and disadvantages, and the decision often hinges on the unique needs, resources, and strategic goals of an organization. Here are the primary arguments for both sides:
We’ll outline the Buy vs. Build considerations as you consider factors in the roadmap related to budget, time, complexity, resources and strategic importance to your core business.
How to create your company’s AI roadmap
Remember, an AI roadmap is not set in stone. It's a fluid document, open to revisions as tech landscapes evolve. It's essential to co-create this roadmap, roping in diverse stakeholders to ensure holistic insights. While milestones and timelines lend structure, flexibility ensures adaptability.
To create an AI roadmap, you should include the following:
Clearly outline goals for what you aim to achieve with AI given your selected initial use case. This could be enhancing a particular business process or achieving specific performance metrics. Establishing your desired outcomes (like faster ticket closures or less inbound requests) and the key performance indicators will give you a tangible measure of success.
Identify the data sources required to build out the AI-assisted workflow for your use case for AI implementation. This could include applications such as your knowledge management portal, customer service chatbots, product recommendation systems, or ticketing system. Depending on the data involved, you may need to develop additional plans for data collection, cleaning, preprocessing, and augmentation.
Now that you understand the objective for your use case, the workflow you intend to enhance with AI, and the data sources needed to do this, it’s time to document the anticipated system features, constraints, and behaviors that will meet the needs and expectations of your stakeholders. This is also the point to evaluate the trade-offs between building or buying an AI solution for automating your workflow. Your completed AI readiness assessment is key to detail the resources and limitations of your business currently.
Skills and Training
At this stage, you will need to allocate team resources to the project and determine the skills required for the AI program implementation based on your requirements and whether there's a need to upskill existing teams or hire new talent.
Building: This may involve hiring dedicated personnel to manage your LLMs and AI implementation. Engineers in this domain average $160K annual base salary according to Indeed.
Buying: If buying AI-as-a-service, your solution provider may provide training and guides for your teams to help reduce the need for upskilling. A tool like Knode, with a conversational chatbot built right into Slack, makes the AI tool immediately accessible to users.
Model Development and Validation
There are many steps involved in designing, training, validating, and fine-tuning AI models. This might also include considerations for transfer learning (i.e. context), the use of pre-trained models, etc.
Building: Building an LLM is expensive and cost prohibitive for all but the largest companies.. According to a technical overview of OpenAI's GPT-3 language model, each training run required at least $5M USD worth of GPUs. A 30B parameter model (like CPT-3) can cost approximately $450K USD while smaller 7B models can be trained for around $30K USD. If you choose to build on top of an existing model, there are resources that can help equip your team to choose, train and validate AI models but these still introduce costs for your organization. For example, evaluation tools and employee time just to evaluate the initial model you build upon.
Buying: Since model development and administration is not an issue when buying, the focus with AIaaS tools is in validating the response and performance of the tool. With many AIaaS tools, much of this work is included. For example, with Knode we perform custom training and validation on an ongoing basis as customers use the tool - but it is important for customers to perform their own validation as well as providing ongoing feedback to improve the system. With a solution like this, you get the best of LLMs at a recurring monthly cost per user - giving you the billing reliability to manage your IT budget.
Technical Deployment Strategy
Outline how the AI model will be deployed, whether it's in a cloud environment, on edge devices, or integrated into existing software infrastructure. This is where multiple stakeholders will be needed as you should actively consult your IT team and CIO to make sure your deployment includes best practices in security.
Building: Depending on the use case, you should consider your infrastructure (scaling, latency, hardware), deployment strategy (cloud vs on-premises, edge deployment if needed), versioning, monitoring, redundancy and more.
Buying: Speak with your AIaaS provider about your options. Solutions like Knode will give you the choice of shared cloud, dedicated cloud, or on-premises options. Many of the other considerations are part of the included service.
Maintenance and Iteration
Outline your plan for maintaining and updating the model over time. What happens as your organization grows and new data is introduced? AI models may degrade over time or as data changes, so it's essential to have a strategy for continuous learning and adaptation.
Building: Given the speed of change in AI, you should expect building and maintaining your own solution to be a resource-intensive exercise. If AI engineering expertise is truly a core competency required for your business to be successful, this is the right choice.
Buying: The benefit of using AI-as-a-service in this instance is that the provider will maintain and iterate on your behalf. Many providers, including Knode, often give you the option to use your preferred LLM or will recommend one based on your use case.
Budget and Resource Allocation
Estimating the costs associated with implementing the AI solutions and how resources will be allocated over the project's duration.
Building: Unlike buying, building your own solution may introduce more variable costs based on usage and your deployment. Don’t forget to factor in talent and technical costs into the equation as well as adding potential padding to your budget to account for any additional build timelines. For example, aside from computing fees, the annual salary for a software engineer is between $120-250k USD at Anthropic and $240-300K with a total compensation of $925K at OpenAI.
Buying: Check out the billing and pricing for your solution provider to estimate your cost. Be sure to note any discounts for longer term contracts as well as usage limits that may push you in or out of a specific tier. Consider your primary use case now and your long term vision as you estimate costs. Many providers like Knode also have enterprise solutions that can be customized and billed based on your needs. For example, for organizations above 100 users, Knode offers a flexible user rate and custom feature packages with annual and quarterly billing. Other solutions like Microsoft’s AI feature add-on is a flat $30 per month that can be included in your Microsoft 365 billing.
Create a clear schedule indicating when each step or milestone will be completed. This will help to ensure that the organization is on track to achieve its goals.
Building: Your roadmap should include a timeline for developing, implementing, and adopting the AI applications. Don’t forget to build in a window for initial feedback and tweaks as part of your implementation phase.
Buying: Typically, buying a solution will be less time to value for an organization and will focus on successful onboarding and adoption. Regardless, you should inquire with potential vendors about estimated onboarding time and include this in your roadmap.
Identify potential risks or challenges that might arise during the implementation and strategies to mitigate them.
Building: You will need to address potential issues like data privacy, permissions, and regulatory considerations in addition to understanding and managing biases, as they can skew model outcomes .
Buying: With vendors, you will want to follow any existing vendor requirements around SOC II certifications, GDPR compliance and more. In addition, you may also want to understand the stage of the company and their insurance coverage to make sure the vendor is well managed.
No matter whether you are building or buying AI solutions, addressing potential ethical issues, biases in data, transparency, and explainability of AI models, and their impact on stakeholders is crucial for the long term success of the program. While biases in data can be a systematic issue, it is important to consider this when embarking on your AI journey. Adding accountability based on the source and the query will help monitor where this bias can be corrected. AI productivity tools like Knode can help with this. Knode follows a “source-first” principle that will detail out which sources are used to generate a response so users can follow the paper trail as needed.
Now that you know all the factors involved in your AI Roadmap, use our AI Roadmap Template to get started implementing your company’s AI strategy.
Overall, AI roadmaps are an important tool for organizations that are serious about using AI to achieve their business goals but - remember - AI is not a silver bullet. It is important to have a realistic roadmap for its adoption. This should be a guiding document to ensure that all teams are aligned, resources are correctly allocated, and potential challenges are addressed proactively. Whether you're building or buying, organizations must evaluate their specific circumstances, weighing the pros and cons of each approach, to make the most informed decision and build their competitive advantage.
Every business is unique. If you'd like to discuss your journey to AI transformation, schedule time with our team of experts.
Frequently Asked Questions (FAQs)
What is an AI roadmap and why is it important for my business?
An AI roadmap is a strategic plan that outlines the steps, timelines, and milestones your organization needs to follow to successfully integrate AI solutions into your operations. It's crucial because it serves as a guide, helping ensure that your AI initiatives are aligned with your business objectives, and it provides a clear path for implementation, helping to navigate the complexities of AI adoption.
How do I decide whether to build or buy an AI solution for my company?
What are the key components of an effective AI roadmap?
Can the AI roadmap change over time, and how should my business handle it?
What are some common challenges businesses face when implementing their AI roadmap, and how can they be addressed?