Make or buy? The key question when integrating AI tools into an organisation

Introduction

After an experimental phase, many companies are now ready to integrate generative AI (GenAI) tools into their production processes to capitalize on the growth and profit benefits reported by pioneers and analysts.

This presents a significant opportunity to address the skills shortage by generating more business with the same workforce rather than scaling up staff numbers.

However, the risk lies in getting lost in an AI "zoo" of unmet expectations, unforeseen risks, and redundant investments.

From a corporate development and innovation management perspective, I, too, plan to translate our GenAI experiences from pilot projects into significant process improvements.

A key question I've recently resolved for myself is whether to make or buy: Is it wiser to wait for commercial AI products or develop our own AI tools?

Thesis

In the past years, many Organisations have shied away from developing their own IT solutions due to a shortage of skilled workers and challenges related to time, quality, and cost in software development projects (except companies whose core business revolves around software-based processes, like Zalando).

I'm here to tell you that it's worthwhile to shift our mindset and relearn how to develop AI tools in project mode, either internally or through outsourcing. Here's why:

  1. Off-the-shelf AI products address standard issues, but businesses need to tailor processes to their specific needs.
  2. Creating custom AI tools might be easier than you think.
  3. Licenses for general AI tools become costlier as user numbers increase, whereas developing in-house usually involves costs based on usage.
  4. Centralizing scaling through one platform improves interface management, governance, and compliance.
  5. Developing AI internally offers valuable insights for new products and business models.

Reasoning

Off-the-shelf AI products solve standard issues, but companies need to optimize bespoke processes.

Like any other, AI products target standard issues to appeal to a broad customer base. However, companies should aim for comprehensive, end-to-end optimization, focusing on specific processes.

Typically, no single IT application or AI bot solves all problems, given the variety of providers within the processes, such as Microsoft (Copilot), SAP (Joules), or Atlassian (Intelligence).

There are two options:

  1. License an AI from a provider and adapt it to other environments,
  2. Develop a custom AI application.

For instance, Microsoft Copilot can be enhanced by integrating additional systems and workflows by using APIs from other systems ("Plugin") or by expanding the Microsoft Graph with more data, documents, and information.

However, this requires all involved employees to have a Copilot license, and each expansion of the Microsoft Graph incurs additional costs.

Often, it's more cost-effective to develop the AI tool in-house.

In this picture you can see a cost distribution key when AI tools are produced in-house.

Creating custom AI tools may be easier than initially thought.

Numerous complex and challenging AI applications exist, such as enhancing customer support with AI-driven telephony.

However, most current AI applications in production within companies are relatively straightforward, often resembling a "Chat with your Document" model. This involves integrating internal documents using RAG architectures with large language models.

Examples include:

  • Chatting with contracts
  • Chatting with intranet content
  • Chatting with knowledge from Confluence
  • Chatting with source codes or documentation
  • Chatting with policy and corporate strategy documents.

These types of AI applications are relatively easy to implement, as 80% of the solutions are typically the same, with only the remaining 20% needing specific new development (e.g. for internal interfaces).

GenAI tool licenses cost more with additional users, while self-production is charged based on usage.

Imagine a company with 1000 employees who want to deploy an AI chatbot to handle internal contracts.

Costs for GenAI tool licenses (buy) usually increase with user numbers, while in-house production costs are calculated based on manufacturing and AI usage (make).

Buy Option:

  • Annual cost for 1000 Copilot licenses: €337,200 (recurring) [1]
  • Total cost over 5 years: €1,686,000
  • Requirement: All contracts must be stored on SharePoint or OneDrive to avoid extra migration or interface integration costs.
  • Benefit:
    • The license can be used for additional use cases.

Make Option:

  • Initial costs: €100,000 in the first year, including 100 person-days for software/cloud engineers and learning to use LLM APIs, then €20,000 annually for operation and maintenance.
  • Total cost over five years: €180,000.
  • Requirement: Access to software/cloud engineers.
  • Benefit:
    • Flexible document storage; interface costs included in the initial €100,000.
    • Reusable code, making subsequent similar AI use cases significantly cheaper.

Over a 5-year period, in-house production is 89.3% cheaper than purchasing the solution.

The in-house production of AI use cases has reuse and management advantages.

Uniform scaling across a platform enables reuse, along with centralized interface, governance, and compliance management.

In the above Contract AI scenario, I estimate that €100,000 will be used to develop the initial use case. This figure is on the higher side since it's meant as the first use case of this kind.

Many processes require specific document interaction tools. For example, if you have 40 processes, you might need to consider 40 such AI tools. Although these tools are about 80% similar, they differ in special features, system prompts, and integration with existing systems.

Using a SaaS-Builder platform allows you to leverage the reuse of functions and components across apps, reducing marginal costs with each additional AI use case. While the first use case costs €100,000, the 40th might only need one production day to assemble from existing components.

Such a platform also enhances governance and risk management. Unlike externally sourced AI tools, which increase in complexity and require individual management, this platform offers centralized monitoring of all AI use cases. This allows you to track requests to the cloud-based language model, check system integrations, evaluate user feedback on AI response quality, and ensure AI compliance rules are met.

Those who build AI tools themselves should rely on a platform approach.

Developing our own AI provides valuable insights for new products and business models.

Companies developing their own AI tools gain valuable experience to create innovative AI-based products and business models.

Predictions that software would dominate the world were overly optimistic in timing, yet the trend remains clear: Traditional businesses and mid-sized companies are increasingly becoming software-driven. This shift means services are now often delivered through purely software-controlled processes rather than a mix of human labour and IT tools.

Artificial intelligence accelerates this trend. Companies active in software must integrate AI into their processes to intelligently handle unexpected situations and meet user interface expectations.

The capabilities companies acquire today by developing their own AI tools—either internally or through proven IT outsourcing partners—enable them to tap into new markets and revenue streams with innovative, software- and AI-driven products.

My Recommendation

The market and technology are rapidly changing. Thus, there's no clear right or wrong in building or buying AI tools. My belief that in-house development should increase may soon be outdated since competition among AI tool providers will likely drive down prices, and as AI platforms advance, the range of application scenarios will also expand.

But at the very least, I recommend testing creating AI tools yourself to gain valuable insights, which can simplify future decisions about using AI in your processes. This doesn't necessarily mean hiring software developers and AI experts right away. Trying out make-projects with students who tweak open-source architectures or a co-innovation project with your IT provider might suffice for starters.

Invitation to Discuss

What do you think?

Is the IT industry on the brink of a custom software renaissance, or do IT and department heads still prefer productized solutions to avoid the stress of development projects and maintenance, even if they're more expensive?

Feel free to share your thoughts in the comments!

Best regards,
– Martin (martin@deliberate-diligence.com)