Why is legaltech so disappointing?

Document assembly has been around for a good 15 years — they thought it was “Lawyer-killing disruptive technology” in 2006, and, well, the cockroaches — we cockroaches — are still here, lazengem, and document assembly technology still doesn’t work very well. And nor, for all the promise, do many of the other heralded applications in the vanguard of the reg-tech revolution. The things that were supposed to revolutionise legal practice, putting junior lawyers out of work — the chatbots; the natural language parsing; the data-extraction — still seem to be eluding us.

JC pontificates about technology An occasional series. →Go uptown For your monitor, sir. Technology Tell me more Sign up for our newsletter — or just get in touch: for ½ a weekly 🍺 you get to consult JC. Ask about it here.

“Unlike a covered call, which is about promising to sell what you actually own, a naked call is about promising to sell what you don’t actually own.

Like wearing a nice sweatshirt, learning the lingo, and hanging out at a hackerspace with a code editor open, looking the part, but only scrambling to learn a new skill if somebody actually hints they might want to hire you if their funding comes through in a few months.

That’s selling a naked call option. Faking it till you make it. Ironically, it calls for careful dressing up.”

—Venkatesh Rao, The Premium Mediocre Life of Maya Millennial (2017)

Why is reg tech so disappointing?

Why?

If advanced technology is magic, then “magic” is in the eye of, and measured from the perspective of, the beholder. When the beholder in question inhabits the legal or compliance department the technology doesn’t have to be awfully advanced to seem magical. Especially in a proof of concept. Your salesguy airily drops “blockchain”, “chatbot”, “natural language processing”, “algorithm” and “AI” into his patter and he will sail through.

And so he does.

Music as the exemplar

If you want to see real AI and really powerful algorithms at work have a look at modern music production software.

IK Multimedia's Amplitube and Tonex emulates and models any guitar amplifier (including your own!), down to the cabinet, speaker, microphone, placement (off/on axis!) and hall dynamics.

Apple's Drummer can play along even with my dyspraxic time-keeping, inserting fills and anticipating bars, turnarounds, choruses and middle-eights.

AI can (and does) exactly replicate Tony Visconti's “Bowie Histrionics” microphone set up at Hansa Studios in Berlin.

Izotope automatically mixes and masters — this is truly a dark art that no ordinary mortal can do — compressing, limiting, dynamically equalising — it even replicates the production values your favourite Trevor Horn reference track.

In short, everything you need to recreate Abbey Road studios: its acoustics, amplifiers, mics, pianos, synths, analogue keyboards, autotune, the BBC orchestra, compression, limiting, mixing and mastering — everything can be done inside a laptop, thanks to the power of AI and algorithm, using software that costs less than a couple of grand.

The better question: why is AI elsewhere so much worse than that what currently exists in music production.

We wonder if what is going on here is a basic misunderstanding of sensible and silly applications of AI.

AI assisting artists to make art — fulfilling limited, directed tasks, as do the music applications listed above — is liberating and expansionary. It lets human creators get on with the interesting parts of the creative process, while outsourcing the drudgerous and technical ones to a machine that won't complain or get them wrong, and will basically follow human instructions. The artist remains in executive control of the enterprise.

But AI replacing artists to make art is, after half an hour, rubbish, unless you like a horse with three legs, one of which comes out of its head. AI legal tech is, out of the gate, rubbish, and basically good for nothing.

No-one wants music that can write itself. Writing music is fun. Artists don't create good music or paint good pictures in order to generate commoditisable art: that is the goal of music publishing: that is a second order derivative of art — commercial activity that depends on art. Artists write and paint because they have something to say. It is a very human impulse to project. Creating is what enlightened human beings were designed for. Asking algorithms to create music misses the basic point of music, which is human: a cultural sharing and a non-verbal conveyance of mood and sentiment between people. It is to skip the creative process altogether and go straight to publishing. The thing that the commercial activity at first depended on is missing.

If you want random, machine-generated rhythmic noise you don't need AI: you just need a sewing machine.

The other amazing thing about music tech is the absence of rent extraction. The software is designed to be run by the user without help, intermediation or a “service”; the expectation is no software-as-a-service because the software is so good you don’t need any service.

What reg tech ought to do

• Allows infinite flexibility: In the olden days you needed a typist with some carbon paper: there was a real cost to manipulating words. You were trained to be economical. Modern information technology allows us to freely manipulate, desiccate, desecrate, defibrillate and duplicate data. A good enough algorithm can, in theory, handle any kind of syntactical complexity, costlessly ingesting and processing the densest textual construction. With a simple cut-and-paste we can replicate, vary and augment at will. But this generates what we call the “Yngwie Malmsteen paradox”: Just because guitar technology means you can play 64th note flattened mixolydian arpeggios at 200 bpm doesn’t mean you should.
  

• Disintermediate. Like, really disintermediate.: the heat signature of the information revolution is its capacity to disintermediate. Suddenly, any random could publish anything to anyone, free of charge. Teenagers in London could engage manufacturers in Pakistan to produce custom cricket merchandise. Fat middle aged lawyers can partially fulfil teenage dreams to be record rock music and publish it to the world. But, inside the great steampunk Bolshevik machine that is a modern financial services firm, the organisational psychology militates against it. The great orthodoxy will insist on total top-down control in the form a bureaucratic chain of command: procurement, internal IT, management, a process literally intended to remove the optionality, flexibility and improvisational utility that disintermediation promises: whatever value the concept had will be bloated, deprecated, rigidised and commoditised to the point where using it is a chore. This is not how it was meant to be. Few successful innovations in the history of the world have made things more tedious for workers than they already were.
• Software as a service: Software developers — especially bad ones — have no greater interest in disintermediating themselves from their product than do their Marxist paymasters. For re-intermediation — I beg your pardon: software as a service — is how they take their cut. They are rentiers. This would be more defensible were the reg tech products they flog unique, imaginative or even any good, but they tend to be generic and underwhelming. The real special sauce comes from the user painstakingly training the software to do its bidding. A clever vendor will harness that training, repackage it and sell it to other clients as the “service” the software is providing. This is an excellent chiz, by the way: charging one client to train your stupid software so you can flog it to another client, who will train it a bit more, so you can flog it to another cl ... Well, you get the point.
• Doesn’t provide user flexibility: policy will see to that. The product will calcify, it is too hard, requiring too many approvals and too many business cases to develop.
• Doesn’t provide out of the box usable content: to be usable the will require lawyers, and there are generally precious few of those, and they generally are refuseniks and low-cost-location rent-a-seat types who can follow instructions but aren't any good at writing them.

What none of this does is put useful tools in the hands of the user.

• Don't be a rentier: How do I make money off something which is basically a simple idea that doesn’t require a lot of maintenance? The whole point of this tech is it is meant to be labour saving, right? I can’t do it per unit - the whole point is to eliminate the cost of having meatware do manual, repetitive tasks, and — once you have set it up — there is no actual cost to having a machine do it. So trying to act like a rentier is (a) a dick move and (b) is going to get you killed, because your big idea isn’t that flash, and someone will do it, and undercut you. See Roger Martin’s the The Design of Business: Why Design is the Next Competitive Advantage
• Remember the meatware: If you convert your user experience from “answering nuanced legal questions” into “completing a mandatory questionnaire”, you have lost. Document assembly applications: I’m talking to you. You are trying to make humans behave like machines. That is stupid. Humans aren’t good at emulating machines. Humans are better than machines precisely because they aren’t machine-like. If you have reduced your process to a rules-based questionnaire, you don’t need humans at all. Get a machine to do it - hook it up to the trading system directly.

What reg tech should do

The aim of reg tech should be to work with lawyers and to respect this divide between things machines are good at (accurately, cheaply and quickly following orders) and things the meatware is good at (interpretation; judgment; lateral thinking; dealing with conundrums; figuring out what to do when the instructions run out), and to divide labour accordingly:

• Reduce risk: reviewing a contract more systematically than a lawyer by reference to pre-described policies and therefore more reliably picking up things that a human, however skilled, might not.
• Reduce waste: Make the process more efficient by dealing cheaply with formal issues so that the expensive unit (the in-house lawyer) is only engaged in the stuff really requiring skilled legal attention. Thus, the “touch time” is shorter, and the cost of negotiating the contract lower. Since the machine handles all the basic stuff it allows the lawyer to concentrate her judgment on the difficult issues without getting bogged down on the tedious stuff that is always the same.
• Empower the user: AI should function as a para-legal: It should respond to the lawyer’s simple instructions: “do this/don’t do that/care about this/don’t care about that” and the application should go away and do it. This makes the lawyer more valuable and incentivises her to use the AI: the really dull stuff will just get done automatically, and the next layer — points that are important to spot, but trivial to articulate — the lawyer can quickly spot these (the valuable part) and hand them off to the machine for articulation (the trivial part).

What reg tech often does do

In this user’s experience reg tech tools tend to invert these priorities:

• Risk reduction: Unless your natural language processing is pretty special, machines cannot pick up all syntactical, legal nuances. They should be a first-line triage, not a last line: in other words “whatever else you pick up, consider the following”. If the lawyer assumes the machine has read everything other than what it highlights (“I’ve read the whole document and I have found this...” she puts her faith in the interpretative judgment of the machine. Good luck. Risk reduction thus won’t significantly reduce lawyer time (if it does, consider whether you are really risk reducing). It is an overlay to lawyer time: the lawyer will need to review the whole contract those things beyond the scope of the approved algorithm.
• Waste reduction: Well-targeted AI can take care of simple drafting work on the simple points it does pick up. This would require some skill in natural language parsing but should not be beyond the wit of decent AI. The application would do the “eighty” while the lawyer does the “twenty”. Typically, reg tech applications get this back to front, purporting to do the big picture stuff (there’s an indemnity!) and leaving the lawyer to sweat the details in terms of drafting. It would be better for the lawyer to read the draft cold and go “fix the indemnity, change the term to 2 years, restrict disclosure to persons with a need to know, remove the requirement to notify for regulatory disclosures and take out the non-solicitation clause”.
• Empower: Getting this backward will disempower the lawyer: This is the machine purporting to spot the big issues and assigning the lawyer to do the manual process. This reduces the lawyer to being a sort of grammarian.

See also

• Reg tech