Building a Legal-AI Tool Solicitors Will Actually Trust

The hard part of legal AI is not the intelligence. It is earning the trust of a profession that is right to be sceptical. These are the design decisions that put trust first, and why each one is also the commercially smart choice.

When I started building Probatur, I assumed the difficulty would be technical. It was not. Plenty of systems can search case law or summarise a judgment competently enough. The real problem is that the people I want to use the tool have every reason to distrust a machine that offers them legal material, and the ones who have trusted the wrong machine have paid for it in open court. Once you accept that, the product changes shape. Trust stops being a compliance box at the end of the build and becomes the first constraint that everything else has to satisfy. What follows is an account of the decisions that fall out of taking that seriously.

The scepticism is earned, and the data proves it

It is tempting to read professional caution about AI as technophobia. The data says something more precise, and more useful to anyone building in this space.

Lawyers are adopting AI quickly. By one widely cited measure, seventy-nine per cent now use it in some capacity, up from nineteen per cent only two years earlier. But the confidence underneath that adoption is thin, and the worries are specific. In one professional survey, three-quarters of lawyers said concerns about AI making things up were the reason they had held back from using it. In another, among those who felt AI had no place in their daily work, accuracy and reliability were the primary concern for forty per cent, almost double any other objection. A recent trust index put accuracy as the top worry for fifty-six per cent of legal professionals, with data security close behind at fifty-three, driven by an obvious fear about what happens to confidential client information fed into a general model.

None of this is irrational. In a profession where a single fabricated citation can produce a wasted-costs order or a negligence claim, "mostly right" is a failing grade. And the failures are not hypothetical. US courts recorded hundreds of instances of AI errors in filings during 2025, more than ten times the previous year's total, with qualified lawyers responsible for well over a third of them. For anyone building here, that scepticism is not an obstacle. It is the market telling you, in detail, exactly what to build for.

Decision one: verifiability over cleverness

The most important decision in Probatur is also the least glamorous. Every output is grounded in the primary source, and that source is always one click away. The tool does not hand you a confident answer. It hands you organised evidence that points straight back to the judgment it came from.

This is a direct answer to the problem the data describes. Independent testing has found that even the leading legal AI tools return fabricated or poorly grounded answers somewhere between roughly one in six and one in three of the time. A tool that gives a busy solicitor a plausible-looking answer they cannot instantly verify is not saving them work. It is adding a checking burden and hiding it, which is worse than doing nothing. Grounding the output in a source the user can see is not a nice extra. It is the whole point, and the cleverness of the underlying model is close to worthless without it.

Decision two: disclaimers as product, not paperwork

Most tools bury the "not legal advice" line in a footer and treat it as something the lawyers made them write. I treat it as part of the interface. Every output states plainly what it is, which is preparation for a professional's own judgment, and what it is not, which is advice, a prediction, or a decision.

Part of this is regulatory. The SRA and the Bar bodies are consistent that the professional remains accountable for how any tool is used, and a product that blurs that line is doing its users a disservice. But the deeper reason is that a good disclaimer is information, not indemnity. Telling someone exactly how far to rely on an output is one of the most useful things you can tell them. A caveat that changes how the tool is used is not a hedge against liability. It is a feature that makes the tool safer to lean on, and I would rather ship that than a bolder claim that reads better in a pitch.

Decision three: name the blind spots

A tool that wants to be trusted has to be honest about what it cannot see. The published record Probatur reads is genuinely powerful, and it is also partial. Most disputes settle without a reasoned judgment, most judgments are never formally reported, and coverage of the lower courts is thin. Rather than paper over that, the product surfaces its coverage and declines to imply it has seen everything.

This runs against the instinct of a young company, which is to project omniscience. But for this audience the instinct is wrong. Litigators test evidence for a living, and a source that names its own limits reads as more credible than one that claims none. Admitting what the tool does not know is, counter-intuitively, one of the strongest trust signals it can send.

Decision four: a verified profession only

Access to Probatur is gated behind a real SRA, BSB or CILEx number, verified before anyone gets in. That choice does two jobs at once.

It keeps the tool in the hands of people trained to exercise the professional judgment it is built to support, which is exactly where the responsibility for any output should sit. And it tells every user something about the company they are dealing with: this is made for professionals and held to professional standards, not a consumer novelty dressed up in legal language. The gate slows sign-up, and I have kept it anyway, because it is as much a statement of who the tool is for as it is a control on who gets in.

Decision five: fictional demos, real integrity

A smaller decision, which matters more than its size suggests. An early version of the demo showed a real judge's name attached to an illustrative statistic. On reflection that was a trust error in miniature. It paired a real, named individual with a figure the tool had not earned the right to assert, and in doing so it modelled precisely the overclaiming the whole product exists to avoid.

So the demo now uses a fictional composite. It shows honestly what the tool does, without making a claim about any real person that it cannot stand behind. If you are asking a cautious profession to trust your restraint, the demo is the very first place they will look for evidence of it, and it had better be there.

Why trust-first is also the winning strategy

It would be easy to read all of this as an ethics tax, the slow and careful road taken at the expense of the fast one. The evidence points the other way.

The barrier to adoption in legal AI is not capability. It is confidence. The tools struggling in this market are not the least intelligent ones; they are the least trusted ones. Survey after survey finds that lawyers extend the most trust to tools grounded in verifiable sources and built around their real work, and the least to general-purpose systems operating as black boxes. Trust-first is not the timid strategy in a regulated, reputation-sensitive profession. It is the only one that scales. And it happens to be the story that a serious acquirer, running its own diligence years from now, will most want to find: a product that understood its market's deepest reservation from the first line of code and built for it deliberately.

I did not set out to build the cleverest legal AI. I set out to build the kind a careful solicitor would put their name behind, because in this profession that is the only kind worth building. Every decision above costs something up front: a caveat where a bolder line would sell harder, a gate that slows growth, a demo that refuses to overreach. Each of them is the reason the tool has any chance of being trusted at all. In legal AI, trust is not a constraint on the product. It is the product.