We Use the Same AI Models as Claude. That's Not the Point.

tobias.warken@boutiqai.com (Tobias Warken) — Wed, 10 Jun 2026 16:06:48 GMT

Here's something most legal tech vendors won't put in writing: nearly all of us run on the same handful of AI models. Claude from Anthropic. GPT from OpenAI. Gemini from Google. When a vendor claims to have built their own "legal LLM," they usually mean they've fine-tuned or carefully prompted one of these three.

That includes us. Treasure Contracts runs on Claude, GPT and Gemini, picked per task.

Why admit this in a blog post? Because it leads to the question we hear in almost every demo: "If you use the same models, why shouldn't our legal team just use ChatGPT?"

Fair question. It deserves a straight answer instead of a scare story. So here it is: for some tasks, ChatGPT or Claude on their own will serve you well, and we'll tell you which ones. For contract work at company scale, the chat window itself becomes the problem. And in Germany, there's a criminal statute worth reading before anyone pastes a contract into a consumer chatbot.

Key takeaways

General-purpose AI (ChatGPT, Claude, Gemini) is genuinely good at first-pass contract review, summaries and drafting. We use the same models.

The weaknesses are structural, not technical: chats lose context across iterations, knowledge stays trapped in individual conversations, collaboration turns into email ping-pong, and nothing is auditable.

In Germany, § 203 StGB (professional secrecy) and the GDPR set hard limits on feeding confidential contract data into consumer AI tools.

A contract platform uses the same models but wraps them in matters, playbooks, audit trails and integrations. The model is the engine. It was never the whole car.

What ChatGPT and Claude genuinely do well for legal work

Quite a lot, honestly. The models can read a 40-page agreement and summarize it in plain language, flag unusual clauses, suggest redline language, and explain the difference between an indemnity and a limitation of liability to a sales colleague at 11pm. For a first pass over a low-risk NDA, they're fast and surprisingly thorough.

Legal teams have noticed. A survey of 160+ in-house lawyers found that 58% already use AI for contract review tasks — mostly first-pass review, summaries and risk spotting. Another survey shows that 89% of legal professionals now use AI tools, and an ACC/Everlaw survey found that generative AI use in legal departments more than doubled within a year.

The model makers are leaning in, too. In May 2026, Anthropic published claude-for-legal, an open-source suite of twelve practice-area plugins for Claude — contract reviewers, DSAR responders, diligence agents. It's well built, it's free, and every output is explicitly framed as a draft for attorney review.

We're not going to pretend any of this is bad. It's the same technology we build on. The honest comparison starts one level up.

Where the chat window becomes the bottleneck

The failures we see in legal teams that run their contract work through ChatGPT or Claude have little to do with model quality. They're structural. Six patterns come up again and again:

1. The work stays in the chat. The review happens, the output is good — and then what? Someone copies it into Word, emails it around, and the analysis is gone. The reasoning, the accepted and rejected positions, the context: none of it lands anywhere your team can find it next quarter.

2. Chats sprawl. Forty conversations in, you're scrolling a sidebar wondering whether the indemnity discussion was in the chat called "NDA question" or the one called "quick check pls". Sounds trivial. With three lawyers and a few hundred contracts a year, it's hours of searching — and the searching usually fails.

3. Context drifts across iterations. This is the one that bites hardest. In a long chat, you ask for changes to clause 12 and the model quietly rewords clause 4 — the one everyone signed off on twenty messages earlier. Nobody re-reads the full document every round, so the regression ships. We've watched this happen in real negotiations (and in ai software development).

4. Collaboration means email ping-pong. A chat belongs to one person. The moment a second lawyer, a sales manager or the counterparty gets involved, you're back to NDA_v7_final_FINAL2.docx and an inbox thread with nine participants. The AI made the drafting faster and the coordination slower.

5. No repeatability, no playbook. Ask the same question twice and you may get two different answers. You can paste your playbook into the prompt — and the model will guess at how to apply it, while your colleague next door forgets to paste it at all. A standard isn't a standard if it depends on who's prompting.

6. No audit trail. Who reviewed this contract, against which standard, and who approved the deviation? When an auditor or regulator asks, "it's somewhere in someone's chat history" is not an answer. We went through ISO 27001 certification ourselves this year; this question is no longer hypothetical for anyone.

None of these are bugs OpenAI or Anthropic will fix, because none of them are model problems. A chat is simply the wrong container for a process that involves multiple people, repeating standards and accountability.

Can German legal teams even use consumer AI for contracts?

Now the part that's specific to our home market — and the part most English-language commentary skips entirely.

(The usual caveat: this is a blog post, not legal advice. Talk to your own counsel about your setup.)

Germany protects professional secrets with criminal law. Under § 203 StGB, lawyers — including in-house counsel admitted as Syndikusrechtsanwälte — face criminal liability for disclosing protected secrets. The prevailing view in German legal commentary is strict: the mere possibility that an external provider can access client-related data already counts as disclosure. The German Lawyer Association's (Anwaltsverein) Anwaltsblatt sees no fundamental barrier to using generative AI, but flags cloud-based tools as the critical risk area for confidentiality under § 43a BRAO and § 203 StGB.

Consumer AI tools sit on the wrong side of that line. A few facts worth knowing:

OpenAI's own documentation confirms that consumer ChatGPT conversations are used for model training by default; business and enterprise tiers are excluded by default.
In ongoing US litigation, a court ordered OpenAI to preserve consumer chat logs — including conversations users believed they had deleted. Data you paste into a consumer chatbot has a longer life than the chat sidebar suggests.
A widely reported case from Samsung showed engineers pasting confidential source code into ChatGPT on three separate occasions — at a company with world-class security. The same shadow-AI pattern exists in legal teams today; it just involves contracts instead of code.
A GDPR data processing agreement under Art. 28 does not solve the criminal-law problem. § 203 Abs. 4 StGB requires a separate confidentiality commitment from the service provider. Two different legal regimes, two different agreements — consumer plans offer neither.

Enterprise agreements with EU hosting close part of this gap, and we'd never claim otherwise — we rely on such agreements ourselves. But they don't fix the chat problems above, and they don't stop employees from quietly using their private accounts because the official tool is missing or clunky. The most effective answer to shadow AI isn't a ban. It's a sanctioned tool that's genuinely better.

What changes when the AI works inside a system instead of a chat

This is where we stop being neutral, openly.

Treasure Contracts puts the same models inside a structure built for contract work. Three things make the difference:

Workspaces and matters instead of conversations. Every contract lives in a matter that holds the documents, the parties, the negotiation history, the applicable playbook and the related knowledge. When the AI reviews a draft, it works with all of that — not with whatever one person remembered to paste into a prompt. Ask about the indemnity position three months later and the answer is in the matter, not in chat #37.

An ontology that connects three kinds of knowledge. Contract review isn't only about legal rules. It's about your rules: which entity signs, which liability cap procurement accepted last time, which clauses your industry regulator cares about. Our ontology links company knowledge, contract knowledge and legal knowledge, and feeds all three into every review and every draft. That's the part you cannot paste into a chat window, because it lives in hundreds of documents and a dozen heads.

Auditing without anyone doing the auditing. Every check, every change, every approval is logged as a side effect of doing the work. When the ISO auditor or a regulator asks who approved the deviation from the standard liability clause, the answer takes one click. Nobody reconstructs anything from email threads.

The best legal AI is the one your colleagues never see

Here's the part that surprises people in demos: most users of Treasure Contracts never open Treasure Contracts.

Sales requests and tracks NDAs from the CRM. Procurement checks supplier terms from the purchasing tool. The legal logic — playbooks, fallback positions, approval rules — runs underneath, where legal defined it once. Legal keeps control without becoming the bottleneck, and the business stops emailing contracts around because the workflow lives where they already work.

ChatGPT can't do this, and won't: a general-purpose chat is by definition a destination you go to. Embedded legal logic is the opposite — it comes to the work.

One process instead of fifty chats

A contract isn't a document, it's a process: template → draft → review → approval → signature → archive → retrieval. Treasure Contracts covers that full path — templating, AI review against your playbook, approval flows, signing, and a searchable archive where "find every contract with an uncapped liability" is a query, not a project.

A chat covers exactly one step of that path, for one person, once.

FAQ

Can ChatGPT review contracts? Yes, and for a first pass on low-risk documents it does a decent job — summaries, risk flags, clause explanations. What it can't do: enforce your playbook consistently, keep context stable across long negotiations, support multi-person workflows, or produce an audit trail. And on consumer plans, confidential contract data shouldn't go in at all.

Is it legal to use ChatGPT for contracts in Germany? It depends on the plan and the data. Feeding client-identifying or secret-protected information into consumer versions exposes professional secrecy holders to risk under § 203 StGB, with GDPR issues on top. Enterprise agreements with EU processing and a § 203-compliant confidentiality commitment change the picture considerably. Get advice on your specific setup.

What's the difference between ChatGPT and a CLM platform like Treasure Contracts? The engine is similar — we use Claude, GPT and Gemini too. The difference is everything around it: matters that hold context permanently, an ontology that knows your company, playbook enforcement, approval and signature flows, integrations into CRM and procurement tools, and a complete audit trail.

Does Treasure Contracts use ChatGPT? We use the models behind it. GPT runs alongside Claude and Gemini in our platform, selected per task, under enterprise agreements that exclude training on customer data and keep processing in the EU.

Where this leaves you

If your legal work is occasional and low-stakes — the odd NDA summary, a clause explained — use Claude or ChatGPT, set up the enterprise tier properly, and you'll be fine. The claude-for-legal plugins are a good starting point and cost nothing.

If contracts are a process in your company — volume, multiple departments, standards you actually want enforced, auditors who ask questions — then the chat window is where your efficiency goes to die quietly.

We'd be glad to show you the difference on one of your own contracts. Book a demo — bring your messiest NDA.

Tobias Warken is CEO and co-founder of Boutiq AI, the company behind Treasure Contracts. Before founding Boutiq AI, he spent five years at Freshfields, one of the largest law firms in the world, leading 20+ legal tech and automation projects. This article reflects practical experience, not legal advice.

Hidden Gems Blog