Generative AI in the Enterprise: Buy, Build or Platform?

Observations

Enterprises of all sizes in all economies are implementing GenAI in their employee workflows, customer service interfaces and developer tools to improve productivity and create competitive advantage simply by shipping products quicker.

Buying a GenAI solution off-the-shelf or building in-house is not an easy decision because it has long-term organisational implications. Getting the decision wrong carries significant business and investment risks. When correctly integrated, GenAI can open new revenue streams and deliver real business value by transforming existing products and services.

The choice between buy or build ultimately depends on your risk appetite and time horizon. If GenAI is going to be a key differentiating factor for your business, you may sacrifice short term competitive advantage to build an in-house GenAI solution to develop your unique advantage long term. However, given the growing trend towards deeply integrated AI platforms (such as Google Vertex AI and Microsoft Azure AI), building a custom AI application with platform vendors, and their partners, provides a middle ground in terms of agility and risk, while still creating an intellectual property advantage.

IBRS noted that implementing a custom in-house developed AI project from scratch is not recommended for most enterprises. Software-as-a-Service (SaaS) vendors¹ regularly invest significantly in existing functionality that may meet your needs at less risk. Thanks to GenAI models offered by hyperscale vendors and stable open-source models on commercial Cloud platforms, most enterprises do not need to pre-train their foundation GenAI model. Moreover, most organisations lack the skills to undertake such foundational AI activities.

Even so, there may be edge cases where an enterprise needs to quickly build something completely in-house and play the long-term game.

The following decision framework can assist organisations choose between buy versus build for GenAI solutions.

Define Business Needs

Is AI necessary to satisfy a commodity capability, or can it be a business differentiator? A commodity capability such as a generic text summarisation or content repurpose tool for marketing teams may not be business critical and can be easily obtained from a vendor. However, for a media company, the requirement for similar tools could justify a greater investment that focuses on differentiations that deliver competitive advantage through novel business models.

Cross-Functional Evaluation and Requirements Gathering

Irrespective of whether you choose a ready-made solution or go with a vendor that allows customisation, a cross-functional evaluation team must represent all stakeholders to ensure the solution architecture captures clear requirements to address conflict. Stakeholders should include:

1. Cross-Functional Domain Experts: ensure collaborators discuss a variety of AI capabilities to address real or emerging business challenges and opportunities, and not just technology exploration.
2. Product Managers: document detailed inputs on the specific requirements.
3. End Users: engage with early prototype testing and give critical feedback.
4. Data Engineers: ensure integrity, reliability, and availability of relevant datasets.
5. Software Engineers: explain the integration of the solution into existing or new infrastructure.
6. Security and Compliance: understand and minimise the organisation’s exposure to various risks.

Technology Evaluation

In practice, non-functional requirements are best established with domain experts before determining the functional requirements. Implementation decisions for your AI solution can then be aligned to include non-functional and functional requirements. IBRS has observed that many leading platforms now contain GenAI solutions, so any decision to build in-house AI capability would need to consider potential duplication of functionality across AI projects.

1. Availability: how to determine the appropriate level of downtime the solution can afford?
2. Performance: how does the solution scale with an increase in demand? What is the quality of output required? How can we assess and benchmark performance?
3. User Experience (UX): the type of UX solution may depend on several factors including whether it will be for customer-facing or for internal consumption.
4. Security Compliance: what are the data security requirements from a technical perspective?
5. Support: examine solution complexity to minimise the support model. A multi-product or multivendor solution may increase support complexity. This may require a service integration and management (SIAM) team to be established to manage vendor interactions.
6. Monitoring: how will the AI solutions be monitored and assessed to ensure trust in terms of performance, reliability, and security considerations?

Proof of Concept and Value

When evaluating any AI solution, a proof of concept (POC) is the best way to evaluate functional and non-functional requirements.

During the POC, pay attention to the following:

1. Vendor’s Philosophy: examine the flexibility to comply with your requirements. Do they want the end solution to look a specific way?
2. Functional Assessment: Eexamine features and functions of the AI tool sets using real business use cases.
3. Total Cost of Ownership (TCO): at this stage, you can estimate how much the solution will cost after you have factored in the real cost of software development.

AI Platforms

AI platforms from vendors like Google (Vertex AI) and Microsoft (Azure AI) are cohesive and comprehensive in providing databases, tools, pipelines, and orchestration workflows for AI workloads of all scales and complexity.

Their developer and partner ecosystems contain sourced talent and integration partners that provide access to the latest state-of-the-art GenAI models. Foundation model vendors include OpenAI, Anthropic, and Stability AI. Leading open-source models include Meta LLama, Mistral and their in-house models. Evaluation tools, frameworks and datasets enable rapid customisation and iteration on these existing AI platforms.

Case Studies

Several large enterprises at an early stage in their assessment of an AI maturity model² have managed to undertake innovative projects and deliver impactful business results.

The CIO of Bennett & Coleman, Asia’s largest newspaper, created new revenue streams by digitising old content using machine learning (ML) and GenAI³. The project took nearly 2 years to deliver and provide content to previously untapped markets. Prior domain knowledge and understanding the challenges of working with legacy unstructured data helped the team choose Google for their AI and Cloud data platform. The product uses its proprietary data to retain all IP ownership.

Anand and Anand, a 100-year-old legal firm deployed GenAI to predict the outcome of their current cases, thus saving thousands of man-hours for their legal experts⁴. Building on Microsoft Copilot, a finely tuned model provided relevant text extraction and summary for a junior lawyer to board a new case. The project was done in-house and all the data is protected at both tenant and environment levels.

Next Steps

1. Assess your AI need by utilising the following IBRS models – IBRS Data⁵ and AI⁶ maturity .
2. Examining the coupling/decoupling of AI to your existing systems should modernisation of your data platform be some way off.
3. Examine legacy platform capability to integrate with your chosen AI platform or in-house developed solution.

Footnotes

1. ‘DIY or Ready-Made? Choose Your AI Adoption Path Carefully’, IBRS, 2018.
2. ‘AI Maturity Model’, IBRS, 2023.
3. ‘How a Media Company Used Its Old Assets to Open a New Revenue Stream’, IBRS 2024.
4. ‘A 100-Year-Old Legal Firm Uses AI to Predict Outcomes of Today’s Legal Cases’, IBRS 2024.
5. ‘Data Management Maturity Model’, IBRS, 2021.
6. ‘Analytics Artificial Intelligence Maturity Model’, IBRS, 2021.