Adapters Are Where Business IP Starts to Live

For years, business AI knowledge has been trapped in weak places.

It sits in prompts, chat histories, undocumented corrections, personal workflows, and tacit judgement inside senior people’s heads. That is useful while the person is present and the conversation is open. It is fragile when the business needs repeatability.

Adapters point to a better pattern.

An adapter is a small trainable layer that changes how a base model behaves without retraining the whole model. In practical terms, it gives a business a way to teach a model a specific domain, task, tone, judgement pattern, or operating constraint without owning or rebuilding the entire foundation model.

That is commercially important because it moves value from temporary prompting into persistent capability.

From prompt skill to operating asset

A prompt is an instruction. An adapter is closer to learned behaviour.

The difference matters. If a specialist has to keep rewriting prompts to get the right output, the knowledge remains dependent on that specialist. If the same judgement can be expressed through training examples, specifications, tests, and adapters, the knowledge becomes easier to reuse.

This is where business intellectual property starts to become operational AI infrastructure.

The asset is not just the adapter itself. It is the full learning system around it:

• the specification that defines the task
• the examples that teach the behaviour
• the evaluation set that tests performance
• the governance rules that define boundaries
• the deployment code that controls when it runs
• the audit trail that proves what happened

Together, these assets become a business-owned expression of expertise.

Why adapters suit smaller models

Smaller models often become valuable when they are specialised. They may not be the best general reasoner, but they can be trained to perform a narrower task reliably.

Adapter-led tuning supports that specialisation. It can be lighter, cheaper, and more modular than full fine-tuning. Different adapters can represent different domains, clients, workflows, or control regimes.

That modularity matters for agentic systems. A business may not need one giant model trying to understand everything. It may need a set of smaller, specialised capabilities, each trained for a clear role and orchestrated by deterministic code.

The strategic point

The future of business AI will not only be about who has access to the strongest model. Most serious organisations will have access to strong models.

The advantage will sit in the taught layer: the specifications, examples, adapters, tests, and operating model that encode how a specific business thinks and works.

That is where the IP starts to live.

Further context

This post supports Jonny Bowker’s work on upskilling AI and the wider spec-driven AI methodology.