ChatGPT vs. a legal AI platform: why a custom GPT isn't enough

You open ChatGPT, ask a question about tenancy law, and get a reasonable answer. You set up a custom GPT with some instructions about your firm, and it starts to feel like you've got a junior on hand. But the moment you try to use it properly for client work, something falls apart.

The session knows nothing about your matter. You're starting from scratch every time. The document you discussed yesterday? Gone. The firm guidelines you entered? Lost the moment you open a new conversation. And exporting in your house style? That's on you entirely.

None of this is a criticism of ChatGPT or Claude. They're powerful AI models. We use them ourselves, several of them in fact, because different tasks call for different strengths. But a model isn't the same as a platform. And in legal practice, that distinction matters rather a lot.

What a solicitor actually needs day to day

Let's be honest: most solicitors using AI today are doing so through ChatGPT or something similar. That's understandable. It's accessible, it's quick, and it gives you the sense that you're ahead of the curve. But the reality of running a legal practice doesn't fit neatly into a single chat window.

You don't work on isolated questions. You work on matters. Each matter has its own file, its own correspondence, its own court documents. If you're going to use AI properly, it needs to understand which matter you're working on, what's been discussed before, which documents are in play, and how your firm operates. Not as a nice-to-have, but because it's the only way to avoid errors and genuinely save time.

At Blokzijl, this was the very first request: "Stop making me explain the case from the beginning every time." At LawBeam in London, it was much the same. Solicitors don't want to become prompt engineers. They want a tool that understands their practice.

The reality of "just using ChatGPT"

Bar associations are now testing how lawyers should work with AI. And the honest summary isn't particularly encouraging: you need to write an extensive prompt with all relevant context, you need to verify every piece of output for errors and hallucinations, and each follow-up question potentially adds more noise to the conversation. After three or four iterations, you're spending more time managing the AI than doing the legal work itself.

To be clear: with Andri ai, the solicitor always remains the one with final responsibility. We're not lawyers ourselves, and we wouldn't pretend to be. Every output needs to be reviewed by a qualified professional before it goes out. The difference is that a platform makes that review easier: source references with every answer, a transparent reasoning process you can follow step by step, and matter context that ensures the output is relevant to your file. Reviewing is quite different from having to rebuild everything from scratch. In an independent test by the Advocatenblad, the journal of the Netherlands Bar Association, a panel of twelve lawyers deliberately tried to make the platform hallucinate. Andri ai didn't fall for it. That doesn't mean it can never happen. But it does show that the threshold is considerably higher than with a standalone model.

The question you have to ask yourself at that point is: are you a solicitor, or are you an IT consultant? Because what these guidelines essentially describe is manually doing everything that a proper platform should handle automatically: providing context, verifying output against sources, maintaining consistency across sessions. That's not a sustainable way of working. That's a workaround for the limitations of these models. And yes, everyone says it's all getting better. AGI, deep research, ever smarter models. But look at the reality: for every step forward, there are just as many voices pointing out the fundamental limitations. The question isn't whether AI will one day be perfect. The question is: do you take on a tool now that simply gets on with it and manages all those limitations for you, or do you keep waiting?

And it's worth asking yourself: are you really going to handle a complex matter on a £200-a-month subscription? A matter where you need to work through hundreds of emails, analyse audio recordings for exactly what was said, and carry out digital forensic analysis (checking the EXIF data of photographs to determine which camera was used, where it was taken and when)? Or process iXBRL filings and HMRC tax returns that your client's accountant has sent over? These aren't hypothetical scenarios. This is what solicitors need in practice. And it simply isn't realistic to expect a consumer product, however powerful the model behind it, to deliver that kind of depth.

And then there's the point nobody particularly likes to say out loud: all your client data sits with a provider you have no data processing agreement with. With ChatGPT, including the £200-a-month Pro subscription, the use of your data for model training is switched on by default. You have to dig into a submenu to turn it off yourself. A data processing agreement? That only exists for Enterprise customers, not for you on your Pro plan. Deleted conversations? They're kept for at least 30 days, and sometimes longer due to ongoing litigation. Don't take our word for it. Read their terms of use yourself.

You can't ring them. Your only point of contact is a chatbot that keeps you going round in circles. And if they're breached, your clients' files are exposed, with nothing you can do about it. That's not a theoretical risk. That's the reality of using a consumer product for confidential professional data.

And this isn't just an OpenAI problem. Anthropic, the company behind Claude, treats its consumer subscriptions (Pro, Max, even Team) the same way, despite the professional-sounding names. Training on your data is switched on by default, and turning it off requires a two-step opt-out that most users never discover. A data processing agreement is only available under their Commercial Terms: Enterprise or API. Which is exactly the route Andri ai takes.

That this isn't a theoretical risk was made painfully clear by a federal court ruling in New York in February 2026. In United States v. Heppner, Judge Jed S. Rakoff held that documents generated using a consumer AI tool are not protected by attorney-client privilege. The reasoning: the AI is not a lawyer, there is no reasonable expectation of privacy when using a third-party platform, and the predominant purpose was not obtaining legal advice. In plain terms: if you use a consumer chatbot for client work, your output is potentially discoverable by the opposing party.

We're not solicitors, but this doesn't seem right to us. A data processing agreement with these providers only becomes available once you're spending half a million a month. That's not an exaggeration; that's their Enterprise programme. For a sole practitioner or a small firm, that's simply not an option. The spirit of data protection law is clear: the entire processing chain should be responsible for safeguarding personal data. But in practice, that's not what we see. Bar associations and regulators ought to set clear rules here, so that solicitors aren't left to work out for themselves whether the terms and conditions of an American technology company are compatible with their professional duties of confidentiality. But until that happens, every firm has to make that choice on its own. And the question becomes: do you use a tool that solves that problem for you, or do you carry that risk yourself?

And even if you switch to a tool like Anthropic's Co-work or a similar desktop assistant, you'll run into the same walls. These tools don't work through your matter each time. They don't know your file, they don't remember your earlier questions, and they don't understand how your firm operates. It's the same chat interface in a different wrapper. There's a reason building a proper legal AI platform isn't straightforward: the talent required is scarce and being actively recruited to America with salaries of two to four times what the Prime Minister earns. That translates to annual packages of well over half a million pounds. You don't fund that kind of engineering with a consumer subscription. And if you're thinking "we'll build it in-house": with whom, exactly? The engineers who can do this work in San Francisco, not in your IT department.

The difference between a skill and a platform

A ChatGPT skill or custom GPT is, at its core, a set of instructions layered on top of a general-purpose model: a prompt running in a loop. That's all it is. And we've seen how fragile that foundation can be. A new model drops, stock markets tumble, and suddenly AI is meant to do everything. But the reality is that every model, including Claude, which we use ourselves, has fundamental limitations you need to engineer around. That's not "vibe coding," as everyone seems to call it these days. And there's no such thing as a "vibe lawyer" either: typing a quick prompt and hoping the output holds up. When vibe-coded software fails, an application crashes. When you give legal advice based on an unverified AI output, you lose your privilege, as we described above. This is years of experience understanding what these models can and cannot do, and building a reliable product around those constraints.

Matter context. In Andri ai, you're always working within a matter. Everything you discuss, upload and produce is tied to that matter. Your AI assistant knows the file, the questions you've asked before, the documents you've shared. You don't start over each time. The context is simply there.

Firm knowledge. On top of the matter layer sits a firm layer. Your firm guidelines, your preferences, your house style: the platform understands how your firm works. Not because you explain it every session, but because it's built into the structure.

Production workflows by practice area. A tenancy dispute requires different documents than an employment matter or a judicial review. In Andri ai, you ask for the production you need, and the platform knows what's involved. You move from your research into the right drafting flow. Not a one-size-fits-all template, but a workflow that fits the practice area.

From draft to finished product. You start with a research question, move into a draft, ask questions about your document, make your final adjustments, and export in your own firm style as a Word document. That's a complete chain, not a chat window where you're copying and pasting everything together yourself.

Sources that are always there. In the background, the platform is connected to case law, judgments and legal literature. Not as a bolt-on, but as an integral part of every interaction. It can also pull information directly from company registries and land registries. Without you having to switch between systems.

When things get large

Not everything is a quick question. Sometimes you've got a substantial production: dozens of documents that need processing, an extensive analysis that takes time. In ChatGPT, you're bound by the limits of a single conversation.

In Andri ai, you set a task running in the background. You carry on working on other matters whilst the production runs. When it's finished, you get the result: complete, in context, with source references.

Client communication

Something we see constantly in practice: solicitors spend a remarkable amount of time answering routine questions from clients. "How's my case progressing?" "What does this document mean?" "What happens next?"

With the matter email feature in Andri ai, a client can ask a question and receive an answer grounded in the knowledge from their matter file, under the solicitor's supervision. It doesn't just save time. It means clients are informed more quickly and consistently.

Data alone isn't enough

Some legal tech tools position themselves as unique because they've "got all the case law in their platform." But a database of jurisprudence with a search function isn't a working platform. It's a starting point: perhaps 1 to 2 per cent of what a solicitor needs to run a matter. Search itself is getting remarkably good, remarkably quickly. The added value of "we've got all the data" shrinks by the day. The real work begins after you've found the right case law: how do you apply it to this particular matter, with these facts, for this client? That's where the difference lies, and you don't solve that with a database.

Multiple models, one platform

Solicitors think it matters which AI model you use. "I'm already on GPT-5" or "Claude Sonnet 4.6 is better for legal work." The reality is that the model is just one component in the chain: perhaps 40 per cent of the solution.

We use multiple AI models ourselves, because different tasks genuinely require different capabilities. Every step is optimised for what produces the best result. It happens automatically. As a solicitor, you won't notice any of it, except that the output is better.

When a better model becomes available, our clients are running it within 24 hours. Not blindly: every new model is first tested against an evaluation set we've built together with practising lawyers. Only once it passes that legal quality bar does it go live. No announcement, no upgrade button. It's simply there.

Every one of those models runs within the same secure environment that's been penetration tested by NCC Group (Fox-IT).

Audio and voice

You can upload audio recordings too: think witness statements, client interviews or dictations. You get a transcript straight away that feeds into your matter. Or you speak to Andri ai directly to formulate your question, rather than typing it out. Not everyone thinks most clearly behind a keyboard.

What solicitors say about it

We could say a great deal about what Andri ai does. But it's the lawyers who use it every day who describe the difference best.

Personal injury lawyer Maarten de Klerk called Andri ai "a Rolls-Royce — fully equipped, and impeccably built" in the Advocatenblad, the journal of the Netherlands Bar Association. He now uses the platform across his entire caseload.

Lawyer Laurens Nooijen simulated the work on a complex dossier: 400 hours of work reduced to 4. What would normally take weeks of review and analysis was completed in a single afternoon.

And Roshi Sharma of LawBEAM in London, who began his career at Slaughter and May and now serves over a hundred clients with a team of six: "Near-zero hallucinations. Full transparency in its reasoning. It shows you how it thinks, step-by-step, with sources. Critical when your advice has to be bulletproof." On a recent dispute, Andri ai cut document review by 90%.

What it comes down to

ChatGPT and Claude are powerful tools. But they're built as general-purpose chatbots, not as legal working platforms. The difference isn't the model. The difference is everything around it.

A skill solves one step. A platform solves the entire chain: from matter context and firm knowledge to production workflows, source verification, client communication and export in your own house style. With the security a law firm deserves.

We've nothing against ChatGPT or Claude. They're genuinely strong models. But we choose our model partners based on the agreements we can make about data processing, not just performance. With some providers, you simply cannot make the arrangements necessary when you're handling confidential legal files. That's a deliberate choice, not a technical limitation.

What we build on top of those models enables solicitors to put AI to work properly: without becoming a prompt engineer, without re-entering your case knowledge every session, and without compromising the confidentiality your clients expect from you.

Curious to see what it looks like in practice? Give it a try or get in touch. We'd be happy to walk you through it.