Three doors into AI.

Most boards are being asked to "do something with AI" without a clear picture of what the options actually are. The decision is usually framed as a choice of vendor or a choice of project. Underneath, it is really a choice between three quite different kinds of work, with very different cost, risk, and pay-off profiles.

The three doors

Building, owning, or applying. They are not the same job and they do not cost the same money.

Underneath every "AI strategy" is one of these three choices. Most organisations need only one of them. Some need two. Almost none need all three. Naming them honestly is the first step to choosing well.

1. Building a model

Training a neural network on raw data, from scratch or near-scratch. This is research and engineering work, with research costs and research timelines. The output is a model that did not exist before.

Who this is for: a very small set of organisations with a real research thesis, the data scale to support it, and the budget to fund a multi-year programme.

2. Owning a model on your data

Taking a base model that already exists and adapting it to your proprietary data: fine-tuning, retrieval, evaluation harnesses, deployment. The base model is borrowed; the adaptation is yours.

Who this is for: organisations whose data is a real moat - regulated, specialist, or generated through operations that competitors cannot easily replicate.

3. Applying industry LLMs

Using industry-standard models (the big-name LLMs from the big-name labs) to automate or augment work that already happens in your business. With or without retrieval. With or without agents.

Who this is for: almost everyone else. This is where most of the business value of AI lives in 2026, and where most of our client work sits.

The honest test is: which door do your problem and your data actually point at?

People rarely arrive at the right door by working through a feature comparison. They arrive at the right door by being honest about three things: what they actually need to do, what data they have, and what they can afford to maintain over the next five years.

Start from the problem, not the technology."We need AI" is not a problem statement. "Our underwriters spend three days a week reading documents that the current system already has structured data for" is a problem statement. It points at door three.
Look honestly at your data. Door two only pays back when the data is genuinely a moat. If your data is the same data your competitors have, fine-tuning on it does not build advantage; it just rebuilds the public model with extra steps.
Cost the maintenance, not just the build. Door one and door two both require a team that can keep a custom model alive as the world changes. Door three lets you ride the upgrade curve of the major labs.
Be willing to walk away. Sometimes the right answer is "not yet" or "not AI". A good advisor tells you that before you commit.

Where Probity sits

Whichever door you walk through, the decisions still need governing.

Rubicon Probity is not any of the three doors. It is the layer that sits above them. Whether you build a model, own a model, or apply someone else's model, your business still has to decide what to do with the output. Probity captures those decisions, surfaces the biases at play, and keeps the audit trail that regulators and boards increasingly require.

Read about Rubicon Probity See where AI fits your business

One conversation, no commitment. We listen first, then suggest a path.

Ready to see where you stand?

Two minutes. No email required. Find out where AI fits your business - and where the gaps are.

Get a clear picture

Or skip ahead and start a conversation