You Don’t Have to Build AI From Scratch. Most Organisations Shouldn’t.

How foundation models and pretrained AI are changing what’s possible — and the questions you need to ask before using them.

When most people imagine building an AI system, they imagine something like this: a team of data scientists assembles a massive dataset, designs a neural network architecture, trains it from scratch on thousands of GPU-hours, and after months of iteration produces a custom model tailored to the organisation’s specific needs.

Sometimes that’s the right approach. Most of the time, it isn’t. And organisations that default to building from scratch — out of a belief that custom always means better, or out of unfamiliarity with the alternative — are spending ten to twenty times more than they need to, and taking six to twelve months longer than necessary.

The alternative is to build on the work of others. To start from a model that has already learned to understand language, or to recognise images, or to generate code — and to adapt that capability to your specific domain, at a fraction of the cost and time. This is the world of pretrained models and foundation models. And understanding it is one of the most practically valuable things an AI programme manager can learn.

What pretrained models are and why they exist

A pretrained model is a machine learning model that has already been trained on a large, general dataset, and whose learned representations can be applied to new tasks with significantly less additional training. The concept emerged from a fundamental observation in deep learning: the patterns a model learns when trained on a large general dataset — patterns of language, visual structure, audio characteristics — are remarkably transferable to specific domains.

Google’s BERT, released in 2018, was trained on the entire English Wikipedia and a corpus of books — approximately 3.3 billion words. That training produced a model that deeply understood the structure and semantics of English text. A pharmaceutical company that wants to build a system for extracting information from clinical trial papers doesn’t need to start from scratch. They start from BERT, expose it to a much smaller corpus of clinical literature, and fine-tune it for their specific task. Weeks of effort instead of months. Thousands of examples instead of millions.

Foundation models are to AI what the internet was to software development — a shared infrastructure layer that dramatically reduces the cost and time of building on top of it. The organisations that understand this are moving faster than those that don’t.

Foundation models — the new infrastructure layer

Foundation models are a category of pretrained model that takes this concept to a different scale. Models like GPT-4, Gemini, Claude, and Llama have been trained on vast portions of the public internet, producing models with broad, general capabilities across language, reasoning, and code. These are not narrow specialist systems — they are general-purpose language intelligences that can be applied to an enormous range of tasks through prompting, fine-tuning, or retrieval-augmented generation.

The practical implication for organisations is significant. Tasks that previously required custom model development — summarising documents, answering questions about internal policies, generating first drafts of structured content, classifying customer feedback — can now be accomplished by connecting a foundation model to your data, through an API, with days of integration work rather than months of model development.

GitHub Copilot is built on foundation model technology. Instead of Microsoft training a custom code generation model from scratch for every programming language and framework, Copilot uses a foundation model trained on vast quantities of public code, fine-tuned on specific coding tasks. The result is a system that helps developers across dozens of languages and frameworks — a scope that would have been unachievable with traditional custom model development.

The build vs. buy vs. adapt decision

Understanding pretrained and foundation models reframes the fundamental AI build decision from binary (build or buy) to a three-way choice: build from scratch, adapt an existing model, or deploy a model as-is through an API.

Build from scratch is appropriate when you have a highly domain-specific problem for which no general pretrained capability exists, you have the data volume and quality to train a reliable model from scratch, and you have specific performance, security, or compliance requirements that preclude using external model providers. This is the right choice for genuinely novel capabilities — but it’s the right choice for a much smaller proportion of AI initiatives than most organisations assume.

Adapting a pretrained model — fine-tuning it on domain-specific data — is appropriate for most enterprise AI applications. You get the benefit of the general capability the foundation model brings, while adapting its behaviour to your specific domain, tone, and requirements. A legal services firm that fine-tunes a foundation model on thousands of contracts and legal judgments gets a system that understands legal language with a depth no general model will match, at a fraction of the cost of building from scratch.

Deploying through an API is appropriate for tasks where the general capability of the foundation model is sufficient, where the data involved is not sensitive enough to create compliance risks with external processing, and where the cost structure of API-based access makes sense relative to usage volume. Many content generation, summarisation, and classification tasks fall into this category.

The questions you must ask before using a foundation model

Foundation models are powerful and accessible — which makes them easy to deploy without proper governance. These questions need to be answered before any foundation model is integrated into an organisational system.

• What data will be sent to the model? If any of it is sensitive, confidential, or personally identifiable, external model providers may create data privacy risks that conflict with GDPR, DPDP Act, or sector-specific regulations.

• What are the terms of service for the model provider? Does the provider use inputs to further train the model? Does sending proprietary information to the provider create IP risks?

• How will you manage hallucination in production? For every use case, you need an explicit human review policy that defines when AI outputs can be acted on directly and when they require checking.

• What is your vendor dependency risk? If your deployment relies on a single external model provider, what happens when they change their pricing, their API, or their model behaviour?

• How will you evaluate the model’s performance on your specific task? General benchmarks don’t tell you how the model performs on your data. You need a domain-specific evaluation dataset and a defined performance threshold.

Retrieval-Augmented Generation — the enterprise pattern

The most important practical development in enterprise foundation model deployment is Retrieval-Augmented Generation, or RAG. RAG addresses one of the core limitations of foundation models: they have a knowledge cutoff date and don’t know about your organisation’s specific information.

RAG connects a foundation model to a retrieval system that can access your documents, databases, and knowledge bases at query time. When a user asks a question, the system retrieves the most relevant information from your knowledge base and provides it to the model as context. The model then generates a response grounded in your actual organisational knowledge rather than its general training.

The result: a system that can answer questions about your specific policies, products, contracts, and procedures — accurately, with citations, without the hallucination risk of relying on the model’s general knowledge. This is the architecture behind the internal knowledge assistant tools being deployed across consulting firms, financial services companies, and technology organisations globally.

You don’t have to build AI from scratch. Most organisations shouldn’t. Understanding what’s already been built — and how to adapt it intelligently — is the skill that separates the AI initiatives that accelerate in 2025 from the ones that spend the year reinventing wheels.