For most of the past two years, choosing an AI tool meant choosing a model. You picked one provider, learned its habits and worked around its weaknesses. That is no longer how the better legal platforms are built. The larger tools now sit on top of several models at once and move work between them depending on the task. Harvey, one of the better known legal platforms, draws on models from OpenAI, Anthropic, Google and Mistral rather than tying itself to any single one. The change is quiet, because the name on the product does not shift, and it matters for how you buy and how you supervise.
One platform, several engines
A multi-model design lets a platform send a drafting task to one model, a research query to another and a summary to a third, choosing on cost, speed or accuracy. The general providers have moved the same way from their own side. Anthropic released a legal version of Claude this year and OpenAI has set out a legal offering of its own, each presenting a broad model with a legal mode rather than a system built only for law. The result is that the tool on a fee earner's screen is rarely a single model. It works closer to a switchboard that decides, matter by matter, which engine does the work, and the routing can change as new versions arrive.
The question is no longer which model
This changes what your due diligence should ask. Comparing two models on a published benchmark tells you little when the product in front of you calls on four and alters the mix over time. The useful questions are about the switchboard. Which models can this platform reach, who decides which one runs a given task, and can your firm see or limit that choice on sensitive matters. A firm that wants a particular model kept away from privileged work should be able to ask for that and receive a plain answer. We set out the wider method in choosing a legal AI tool that fits your firm.
Where each model sends your data
The confidentiality question grows with the number of models. Every model a platform routes to is a separate place your client's information travels, each with its own retention terms and its own stance on training. A tool that uses four models carries four sets of terms to read, not one. Ask where each model runs, whether any provider keeps prompts after the session ends and whether client data trains any of them. A promise of zero retention and no training on your data belongs in writing, and it should hold for every model in the chain rather than the headline one. Our note on client confidentiality when using AI tools covers the wider duty.
Cost and the freedom to switch
There is an upside to this design for the buyer. A platform that can route between providers is less exposed when one model rises in price or slips in quality, and a firm that keeps its data and templates portable stays less bound to any single supplier. Ask how your work and prompts would move if you left, and whether your material remains yours. Much of the value in a multi-model tool sits in the choice it protects, so hold that choice in the contract rather than trusting it to stay open on goodwill.
The regulator's position holds through all of it. The SRA expects appropriate human review and effective supervision of AI output whatever produced it, and the named solicitor answers for the work that leaves the firm. A multi-model platform can serve a practice well, giving the stronger model for each task without locking you to one provider. The discipline is to treat the platform as several suppliers rather than one, to get its data terms in writing across every model it reaches, and to keep a person accountable for the result.
The data protection side of any platform is measured against the ICO's guidance on AI and data protection.
If a platform pitch is on your desk and the model names mean little, we translate them into risks and prices before you commit: get in touch.