The Right Model, the Right Context: Building Better Legal AI

The newest AI model isn’t always the best choice for legal work. At Draftwise, we look beyond what a model can do on its own and instead ask: How much should it have to figure out in the first place?

A more powerful model might take longer to find the right provision, return more information than a lawyer needs, or produce a technically correct redline that changes far too much.

Better reasoning doesn’t automatically lead to better legal work. Selecting the right model means testing for what actually matters: useful output, delivered quickly and efficiently.

Model selection is only part of the equation. Through Ontonomi, our legal ontology platform, we organize agreements, provisions, and negotiated positions into connected knowledge that AI can use directly. That can spare a model the work of reconstructing context from documents every time.

Choosing the right model matters. Giving it a better starting point matters just as much.

That principle shapes how we evaluate new models, coordinate them across legal tasks, and build the knowledge foundation they work from.

1. Newer doesn't always mean better

AI models don't improve uniformly across every task. Each new release reflects trade-offs between reasoning capabilities, speed, and cost.

For example, a recent legal AI benchmark comparison found that a newer model was nearly four times faster and cost about half as much to run as its predecessor, but performed slightly worse on the benchmark's strict pass-or-fail test.

That's an important distinction. A faster, more affordable model might be the better choice for certain workflows, even if it scores lower on a particular evaluation.

We've seen similar trade-offs in our own evaluations.

When testing a move from GPT-4.1 to newer GPT-5-series models for contract retrieval, we observed noticeably longer response times and a greater tendency to return more information than necessary.

In technical terms, the newer models favored recall over precision: finding everything that might be relevant, rather than focusing tightly on what actually matters.

For lawyers, that distinction is important. A comprehensive answer isn't necessarily a useful one if it requires additional time to separate relevant findings from unnecessary information.

And benchmarks don't always capture what lawyers care about most.

A model might pass a legal drafting benchmark but still produce awkward language, introduce unnecessary edits, or deviate from a firm's preferred drafting style.

Many benchmarks also test models at their highest reasoning settings, which can come with substantial time and token costs.

The real measure of performance is whether the output is accurate, usable, and delivered efficiently.

2. The system matters as much as the model

Legal AI workflows involve different types of work, each requiring different capabilities.

Finding a specific clause isn't the same as analyzing a complex contractual relationship. And neither requires exactly the same capabilities as producing a carefully targeted redline.

That's why using the most powerful model for every task isn't necessarily the smartest approach.

Model orchestration allows AI systems to coordinate different models and tools according to the requirements of each step.

A faster, less expensive model might be sufficient for a straightforward retrieval task, while more sophisticated reasoning may be justified for complex legal analysis.

The goal is to find the right balance between accuracy, speed, and cost.

But there's another consideration: models don't operate in isolation.

How a model interacts with retrieval systems, prompts, and surrounding architecture can significantly affect its performance. Upgrading to a newer model may require changes to the system itself before its capabilities translate into better results.

In other words, better models don't automatically create better AI systems.

3. Better context can be more valuable than more reasoning

Here's where things get particularly interesting.

Newer AI models increasingly rely on iterative reasoning: examining information, checking their conclusions, and correcting potential mistakes.

That can improve accuracy on complex tasks. But it also consumes time and tokens.

What if some of that work could be avoided altogether?

Consider a lawyer comparing negotiated positions across two versions of a contract.

A conventional document-based AI workflow may need to read both agreements, locate the relevant provisions, interpret each position, and determine what changed. If important context is missing or buried in lengthy documents, additional reasoning may be needed to resolve uncertainty.

Legal ontology offers a different approach.

An ontology structures legal knowledge into connected entities and relationships: clients, matters, agreements, provisions, negotiated positions, and decisions.

Rather than repeatedly asking a model to reconstruct those relationships from documents, an AI agent can access information that has already been identified, organized, and connected.

For our contract comparison example, that means the agent can work directly with structured negotiated positions and their relevant context.

The result is a simpler task: compare the positions and explain the differences.

This doesn't eliminate the need for reasoning or verification. But it can reduce the amount of information the model needs to process and the work required to reach a reliable conclusion.

That has implications for both accuracy and token efficiency.

Sometimes the biggest improvement in AI performance comes from reducing how much the model needs to figure out.

Building legal AI that improves as models evolve

At Draftwise, we've approached legal AI with this principle from the beginning.

Our earliest precedent-driven drafting capabilities used structured representations of clauses and definitions to help lawyers find and apply relevant firm knowledge. That foundation has evolved into Ontonomi, our legal ontology platform.

At the same time, we continuously evaluate models and architectures against real legal tasks, balancing output quality, latency, and cost. As the technology changes, we're willing to rethink our approach based on empirical results.

Because the best model today may not be the best model tomorrow.

What should endure is the firm's institutional intelligence: its knowledge of clients, matters, agreements, and negotiated positions, organized in a way that enables different AI models to access it reliably.

Foundation models will continue to become faster, smarter, and more efficient. But the firms that benefit most won't necessarily be those using the newest model at every opportunity.

They'll be the ones with the right AI architecture and structured knowledge to make those models effective.