The role of context engineering in building trusted, self-service legal AI

When building AI tools for legal and compliance work, accuracy isn’t just a technical challenge. It’s a product challenge.

Josef CEO Tom Dreyfus explores how even the most powerful model in the world won’t help your user if it misunderstands the question or draws on the wrong source.

This, as Tom explains, is where context engineering comes in.

Context engineering is a foundational part of how we build Josef Q. It is especially important because our tools aren’t just for lawyers and other subject matter experts. They’re designed to be used directly by business users in self-service settings.

So, what is context engineering?

Context engineering is the deliberate design of dynamic systems to deliver the right information and inputs to an AI model. Its purpose is to ensure the model can generate accurate, relevant, and trustworthy responses.

In Josef Q, context engineering has two equally critical components:

1. The knowledge base: the body of legal and compliance material the tool draws from
2. The user input: the question or task the user brings to the tool in real time

By shaping both sides carefully, we create the conditions for the tool to perform well. This matters most in high-stakes environments like legal and compliance work, where mistakes can be very costly.

Engineering the knowledge base

Every Josef Q tool is powered by curated source material such as policies, regulations, contracts, standards, and playbooks. But we don’t just ingest these documents and hope for the best.

Using our document pre-processing engine, we transform them into a structured, searchable system that highlights what matters most, so that tools on Josef Q can locate the right passage, interpret it correctly, and generate a correct answer in seconds. This can include things like jurisdictional scope, approval requirements, version history, and relationships between clauses or definitions.

The key is precision and control. Josef Q operates in a closed domain system, answering questions based on the actual documents teams work with, as they are written, reviewed, and governed.

Engineering the user input

Just as important is what the user provides.

Josef Q tools are built for real-world business users, like project managers, engineers, HR professionals, and operations leads. Often, these users don’t know exactly what context the tool needs to generate the best answer. And unlike legal experts, they may not be equipped to verify whether an answer is correct.

Rather than relying on the user asking perfect questions, Josef Q guides users in real time, suggesting clearer phrasing, prompting for relevant details, and adapting to the user’s role, location, and task.

An example

In this Josef Q tool, the user starts with a broad question: “Which security standards do I need to follow?”

To help narrow things down, Josef Q engages a context agent to assess the question and suggest a more apt version: “As a structural engineer working on a big project in Denmark, what security standards do I need to follow?”

This rephrasing adds three key pieces of context:

• The user’s role (structural engineer)
• The type of work (a major project)
• The location (Denmark)

With this added detail, Josef Q returns a response that is not only accurate, but actionable. It identifies the relevant standards, highlights the need for encryption and compliance checks, and links directly to next steps such as reviewing the policy or applying for an exemption.

This is context engineering in action: vague questions are transformed in order to deliver precise, trustworthy answers.

Why does this all matter?

Legal and compliance work demands rigour, clarity, and confidence. Josef Q accounts for this, but it also factors in scale and speed too, enabling non-specialist users to get answers without needing to escalate every question.

How do you ensure tools still get it right, even when the user might not know all the relevant details to include?

Context engineering! By carefully designing both the knowledge Josef Q draws from and the way users interact with it, we’re able to create tools that are not only smart, but also safe, consistent, and useful.

That’s how we build legal AI people actually trust. Not just with better models, but with better context.