Build vs. buy: when should we use AI?

By Matillion General Counsel & Summize Power Counsel Chair Kimberly Trull

There’s a question legal teams are increasingly likely to hear from the wider business: “Why do we need a specialist legal tool? Can’t we just use…?” – and insert their chosen general-purpose AI model, whether it’s Claude, ChatGPT, Gemini, Perplexity or one of the many others on the market.

It’s a reasonable question. General-purpose AI can do a lot, and legal teams should absolutely be looking at where it can take work off their plates.

Take NDAs. If AI can review a high volume of straightforward agreements, identify the ones that need attention and surface the issues that matter, lawyers can spend far less time working through routine contracts. That’s a genuinely useful application for an AI tool like Claude or ChatGPT.

But there’s a difference between using one of these models and using them as the system that runs an entire legal workflow.

That distinction is becoming increasingly important as organizations look for ways to consolidate technology costs. When someone asks why Legal needs a specialist tool when the business already has access to a general-purpose AI assistant, the answer shouldn’t simply be that the AI assistant can’t do it. The better question is: what would it actually take to make it work?

That’s where the build-versus-buy question comes in.

What does AI actually need to work well in Legal?

There’s a natural assumption that if an AI model can do something, it should be relatively straightforward to build a workflow around it, but capability is only one part of the equation.

Legal work often depends on context that isn’t contained in the contract you attach or prompt you write. You know that a particular clause is unusual. The team might have a documented fallback position. The organization might have a commercial reason for accepting a particular level of risk. Often, that’s knowledge that’s built up over years.

Historical decisions, negotiated positions, playbooks, templates, approval processes and previous contracts all contribute to understanding what a particular legal team considers acceptable. For AI to apply that knowledge consistently, it needs access to the right context and a way to use it reliably. If the goal is to reduce repetitive work, that’s especially important.

How do you push back when someone says “just use the model we already have”?

The answer to that question shouldn’t be an instinctive “no”. General-purpose AI models are powerful and flexible, and there are plenty of legal tasks where it is the right tool. But it isn’t always the right system.

I’ve experienced this first-hand when being asked: “Why can’t we build our CLM inside our Gen-AI tool??”

On paper, it’s easy to make that case. Gen-AI is capable of handling a huge range of language-based tasks. If the question is simply whether it can help with contracts, the answer is yes. But that’s not the question.

The question is what your legal team needs the system to do, what needs to be built around the model and who’s responsible for maintaining it. You need to consider how it’ll access your organizational knowledge, how it will apply that knowledge consistently, how it connects to the rest of the contracting process and what happens when something changes.

And of course, someone has to own all of that.

So if you’re asked why Legal needs a specialist tool when the organization already has a general-purpose AI assistant, don’t just argue about what the model can or can’t do – ask what it would take to make it work properly.

The cost of build-versus-buy

“Build versus buy” can sound like a purely financial decision, but it isn’t. Cost matters, but that calculation has to include everything involved in making an AI workflow useful and sustainable. And of course, AI workflows consume tokens. A single prompt and response might look inexpensive, but that calculation changes when you’re applying a workflow across hundreds of thousands of contracts and providing context alongside them.

Token limits and usage costs become a practical consideration when you’re being asked to consolidate work into a general-purpose AI tools. You can quickly use up available capacity when you’re repeatedly passing in contracts, organizational context and instructions and generating outputs at scale.

That doesn’t mean token costs automatically make a general-purpose AI tool the wrong choice – it just means they need to be part of the calculation.

What else has to be built?

The other consideration is that the model itself is only one part of the system. If you build a workflow around a general-purpose AI model, someone still needs to:

• Connect it to the right information
• Keep that information up to date
• Build and maintain the workflow
• Test and monitor its outputs
• Manage permissions and access
• Make sure the workflow continues to work as the business changes
• Decide what happens when the AI output needs human intervention

None of that disappears because the underlying model already exists, and when consistency is the goal, all of it matters.

If you’re reviewing hundreds of similar agreements, the value isn’t just that AI can read them – it’s that it can apply the same agreed approach repeatedly, while surfacing the agreements or issues that need a lawyer’s attention.  

That’s where specialist legal technology can play a role. Understanding contracts is one thing, but building and maintaining everything around the model in order to build a reliable process is another.

Ultimately, the build-versus-buy decision needs to come down to how Legal gets the outcome it actually needs. Sometimes that might be a general-purpose assistant, sometimes it’ll be a specialist tool, and sometimes it’ll even be a combination of both.

What should you ask before deciding to build?

Before deciding to build something, here are a few straightforward questions I would ask:

• What are we actually trying to achieve?
• How much organizational context do these tasks require?
• Do we need the same approach applied consistently across a high volume of work?
• Who will build and maintain the workflow?
• What will the ongoing usage and infrastructure costs be?
• How much human review will still be required?
• Do we need structured data, permissions, auditability or integrations?
• Is there already a specialist tool designed around this workflow?
• Would building it ourselves genuinely be a better use of our resources?

The alternative to buying specialist technology could mean a significant amount of internal time spent building, maintaining and checking something that was never designed specifically for the legal workflow in the first place.  

AI can remove work from Legal. But if it’s being used to build an entire new system, it can also create new work around it. Understanding that difference makes it much easier to have a constructive conversation when someone asks “why can’t we just use the AI assistant we already have?”

Kimberly Trull is the chair of Summize’s customer advisory board Power Counsel and General Counsel of Matillion. Explore more of her insights on legal tech, AI and the future of AI in legal, follow her on LinkedIn.