The Idea in 60 Seconds

  1. Generative AI projects fail 70%-85% of the time, making them riskier than standard IT projects.
  2. Off-the-shelf AI tools are emerging, with 80% of software vendors embedding AI by 2026.
  3. That said, Horizontal tools (e.g., ChatGPT, Microsoft 365 Copilot) and vertical solutions (e.g., Harvey, DeepScribe) already deliver value.
  4. Build internally only for unique, high-cost problems or proprietary data advantages, with ROI under a year.
  5. Focus on AI readiness: governance, ethical safeguards, data preparation, and staff training.
  6. Avoid building now; instead, prepare to adopt proven tools as the market matures.

Why You Shouldn’t Build Your Generative AI Solution

Businesses are investing heavily in AI. Even in early 2024, 87% of companies were piloting or deploying Generative AI in some form

However, most generative AI projects fail. Estimates on the exact proportion vary but consistently, failure rates of between 70%-85% pop up. That is to say, Generative AI projects fail twice as much as standard IT projects.

And, on the horizon, is a material change in how Generative AI will be deployed. Gartner say that, by 2026, 80% of software vendors will have embedded AI in their products. That’s a material change to which I will return in a later article.

So, Why Build A Generative AI Tool Internally At All?

The answer more often than not is that you shouldn’t.

The Same Thing Happened With E-commerce

Not so long ago, every company wanted an e-commerce platform. So they hired developers, wrote custom code, and built fragile, expensive systems. I worked at Vodafone and, even as late as 2010, we spent $millions designing and building bespoke online stores.

A few years later, a plethora of off the shelf products, such as Shopify, Magento, and BigCommerce arrived. They were faster, cheaper, and better maintained. Importantly, they were modular so connections between the components had been considered. The cost of integrating these ready-made solutions, proved far less than bespoke development. (Of course, ongoing licensing and support costs do add up but overall, my view is that splitting the development costs over a larger number of users generally keeps the overall cost lower.)

The same pattern is unfolding with GenAI:

  • Phase 1: Everyone builds their own
  • Phase 2: The free market produces off the shelf products for common use cases.
  • Phase 3: Most realise the real value lies in how you use the tools, not how you invent them

Trying to build GenAI solutions internally today is, in many cases, the equivalent of hard coding a shopping cart 20 years ago.

Where GenAI Is Already Working: Buy, Don’t Build

That said, there are already horizontal and vertical GenAI products delivering materials value.

Horizontal GenAI (General Knowledge Work)

These are tools that supercharge existing workflows—writing, summarising, analysing:

  • ChatGPT – or secure, internal versions of it which don’t leak your data.
  • Microsoft 365 Copilot – Drafts documents, creates summaries, analyses trends
  • Notion AI – Assists with writing, brainstorming, and note-taking
  • Google Workspace Gemini – Embedded GenAI across Gmail, Docs, Sheets

No custom model is needed for these and they have unarguably huge productivity benefits when users are provided adequate training.

Vertical GenAI (Domain-Specific Products)

For high-value, specialised workflows, purpose-built SaaS is already ahead:

  • Harvey – Legal research and drafting for law firms
  • DeepScribe – Medical documentation and transcription for clinicians
  • Fin by Intercom – GenAI support agent for customer service
  • Jurisage – Case law summarisation and legal memo automation

Each solves a real operational problem. Each is already in production. Each is safer, faster, and more maintainable than anything people could build from scratch. There will be an increasing number of these tools available.

When You Should Build GenAI Internally

Building still makes sense but only under very specific conditions.

Build if:

  • You’re solving a real, high-cost bottleneck for your vertical

 The problem is painful, measurable, and not solvable by existing tools.

  • There’s no viable off-the-shelf solution – and you can’t see one coming soon.

This applies in domains like law enforcement (where I work), defence, or niche industrial settings, where highly sensitive data, unique regulatory requirements, or extremely specialised operational workflows are not covered by general solutions.

  • Your proprietary data creates defensible advantage

E.g., domain-specific RAG (Retrieval Augmented Generation) search over internal policies or confidential data sets. The competitive advantage here comes from how you fine-tune or deploy existing models with your unique, proprietary data and processes.

  • You’re building internal capability, not permanent infrastructure

Prototyping as training—for engineers, risk teams, policy authors.

But even then, builds should offer demonstrable, fast, clear ROI of less than a year.

Otherwise, you’re spending time and money on what are essentially vanity projects that will be obsolete (replaced by off the shelf versions) by the time they stabilise, especially given the rapid pace of GenAI development.

So Where Should You Invest Now?

Even if you’re not building GenAI, there’s no excuse for inaction. Smart organisations are preparing for what’s coming by investing in readiness:

 Capability Layer

  • AI governance – Decision rights, risk tolerances, red-teaming protocols
  • Assurance tooling – Model evaluation checklists, approval workflows, escalation paths
  • Data readiness – Tagging, access controls, versioning, sensitivity mapping
  • Ethical guardrails – Bias mitigation, hallucination risk profiling, explainability documentation
  • Change management – Staff education, prompt literacy, use case intake forms

None of this effort is wasted. These are the systems you’ll need to scale fast when the GenAI market settles.

Final Thought: Wait, But Prepare

If you’re not saving time, cutting costs, or solving a real problem within a very short payback period, don’t build a GenAI tool. In the meantime, there’s a lot of value in preparing your organisation to adopt at speed—once the terrain is stable.