The impact of AI on law firms
Two years of building software for lawyers, and most of what I assumed at the start turned out to be wrong.
The first time I demoed an early build of Mandato to a lawyer, he let me talk for about four minutes and then asked whether the thing could tell him which clients he had not replied to that week.
That was the meeting. I had brought document analysis, clause extraction, a summarisation feature I was rather pleased with. He wanted a list of people he owed an answer to.
I have thought about that conversation more than any other in the last two years.
The demo problem
Legal AI demos are built to impress people who do not practise law. Feed in a two-hundred-page share purchase agreement, watch a memo appear. It is genuinely impressive. It is also aimed at the part of the job lawyers are already good at, spent seven years learning, and mostly enjoy.
The bottleneck in a small firm is never the analysis. It is everything wrapped around it. An enquiry that arrived on Friday and got buried under a completion. A nota simple nobody remembered to order. A buyer in Oslo waiting four days for a yes or no that would take ninety seconds to give, if anyone had a reliable picture of what was outstanding.
So the most valuable feature we have shipped is embarrassingly dull. It reads incoming mail and works out which matter it belongs to. That is the whole thing. No reasoning fireworks, nothing you would put on a slide. It saves each lawyer perhaps twenty minutes a day, and it is what they mention when they renew.
A single property purchase at Frank & Partners might involve a Norwegian buyer, a Spanish seller, two agents, a gestoría, a notary and a bank, half of them writing in different languages and none of them using a subject line. The correspondence fragments across four inboxes within a fortnight. That is the actual problem. It has been the actual problem since long before anyone had a model to point at it.
Spain is not a smaller version of the US
Almost all of the money and most of the product thinking in legal AI is built for a firm that has associates, a knowledge management function, and somebody whose job title contains the word innovation. That firm exists. It is not the Spanish market.
The overwhelming majority of firms here are a handful of people, often one or two, frequently a partner who also does the invoicing and answers the phone. There is no procurement cycle, no pilot committee, no six-month evaluation. This cuts both ways, and both ways are sharper than I expected.
On the good side, adoption is direct. Nobody has to be convinced by a business case. A partner tries it on a Tuesday and by Thursday two people in the same colegio have heard about it, which no sales team I could afford would achieve.
On the hard side, there is nobody to configure anything. If a firm has to map its matter types before the product does something useful, the product will not get used. We learned this the slow way. The first version asked for a setup that took a couple of hours, which I thought was reasonable, and which turned out to be roughly two hours longer than anyone had.
Then there is the relationship, which is the part I think foreign observers get most wrong. In a large firm the client's relationship is with the institution. Here it is with a specific person, usually by mobile, often over years, sometimes across a family. That trust is the firm's only real asset, and it is not the sort of thing you route through a chatbot to see what happens.
Which brings up secrecy. Secreto profesional is not a compliance box for a Spanish lawyer; it is closer to a professional identity. Ask one to send client documents to a model provider in another jurisdiction and you get a very particular kind of silence. We answer with data residency, retention limits and a clear account of what leaves the tenancy, and not everyone believes us, and I do not think they are wrong to keep asking. The AI Act is arriving in stages on top of that, and nobody I know can tell you with confidence what it will eventually require of a five-person practice in Málaga.
The cheap model and the expensive one
An architectural decision I did not anticipate having to make, and which turned out to shape the product more than any feature.
We started by sending everything to the best model available, because quality mattered and we were small enough not to care. The unit economics were absurd. A firm paying a few hundred euros a month was generating thousands of calls, and the great majority of those calls were doing work that did not need a frontier model at all — deciding whether an email was about a live matter or a newsletter, pulling a reference number out of a PDF, tagging a document type.
So we tiered it. Small, fast, cheap models for classification, routing and extraction. The expensive one for drafting, for cross-referencing an agreement against what a client actually instructed three months earlier, for anything where being wrong has a cost somebody can feel.
The tiering is the easy half. The hard half is deciding which tier a task belongs to, because that decision is itself a judgement, and being wrong in the cheap direction is far more expensive than being wrong in the other. We ended up with a rule that sounds obvious written down and took months to hold ourselves to:
If a mistake could reach a client without a lawyer having seen it, it goes to the good model. If a lawyer sees it first, the cheap one is fine. Cost is not the axis. Blast radius is.
There was a side effect nobody warned me about. Cheap models are fast. The routing we built to control spend also made the product feel immediate — and in the end, feeling immediate did more for adoption than any accuracy figure we could have put in a deck.
What clients are actually asking for
No client of a law firm has ever wanted their lawyer to use AI. Not one, not once, in any conversation I have had. What they want is a reply.
A buyer writes to ask whether a discrepancy between the registry description and the physical property will hold up completion. It is a real question. It is an hour of work for someone who knows the answer. What that client remembers, a year later, is not the elegance of the answer. It is that it arrived the same afternoon instead of the following Tuesday.
This is why I am sceptical of the AI-lawyer positioning, and why we took nearly all of the AI language off the client-facing side of Mandato. Internally it is everywhere. Externally the client should see a firm that answers quickly and remembers what they said last time. If they ever notice the software, something has gone wrong.
What has surprised me most is how little of this is really about the technology. The models are good enough, and have been for a while. Whether a firm gets better because of them depends on things that have nothing to do with us: whether the partners agree on what a good week looks like, whether anyone is willing to change a habit they have had since 2011.
And I am not yet sure which way it goes. There is a version of this where lawyers get back the part of the job they went into law for, and a version where the same person is simply expected to carry forty matters instead of twenty-five, faster, with the boring parts removed and the pressure left in. I have watched both start to happen in the same office. The tooling does not decide it. The firm does.
The lawyer from that first demo renewed, by the way. I asked him recently what he uses most. Same answer.