For the past year, the technology hype cycle belongs to “agentic” AI. Is it different than the AI chatbots that dominated the previous few years? It depends on how you look at it. The agentic pitch is that users can outsource more of mental drudgery to the AI because it can figure out how to reach your goal without having to be prompted every step of the way.
But the way it does this is, to oversimplify a bit, by deploying the same old AI and running it through a cascading stream of prompts composed by the AI itself. Which moves a lot of work off the user’s plate. It also introduces more points of potential failure.
A year ago, I expected that “agentic” would fail as a legal tech selling point because lawyers would never accept a marketing pitch based on the premise that they should forfeit more decision making. Vendors using the agentic lingo could just as easily describe their products as “vetted workflow automations” or something equally reassuring for lawyers. The stuff being called “agentic” in the legal context always came with enough caveats and guardrails to distance the products from what the general tech community branded as agentic. Why embrace the Silicon Valley term designed for the express purpose of impressing incels who want a robot to autonomously handle their Hinge accounts?
However, I’d underestimated the noxious spread of publicity talking points in the wild. The term has embedded itself in the discourse and we all just have to live with it.
For one thing, most lawyers don’t want to be prompt engineers. Legal tech has always been about trusting experienced providers to build products that get a lawyer closer to their goals without having to look under the hood. Lawyers don’t generally like hearing that they’re not special snowflakes imparting bespoke wisdom with every touch of the document, but the tech overlords have apparently hammered “agentic” enough that even the lawyers are willing to accept it.
And the Valley needs everyone to embrace agents. Tokens keep getting cheaper at the precise moment that AI companies need to massively grow revenue to avoid collapsing under the weight of profligate spending. So the only hope for the industry’s biggest players is a comically enormous increase in token consumption. Agents are ready, willing, and able to get the job done because they will churn tokens like a nervous third-year after checking their billing target in November.
A chat query makes one call, while an agent plans, calls a tool, reads the result, calls another tool, drafts, checks itself, revises, and loops. Gartner puts agentic workflows at five to 30 times the tokens of a single query. On the one hand, if a professional is putting their faith in an agent, it’s nice to know it checks its work obsessively. On the other hand, this has all the hallmarks of investors cheerleading gross inefficiency in an effort to claw back the money they’ve burned chasing revenue that never arrived.
Legal AI installs more protections than just running a free bot and taking your chances, but it doesn’t alleviate lawyers of the responsibility to check their work. The problem with all the “human-in-the-loop” pledges is that it matters where in the loop the human finds themselves. And the more automated the workflow, the more likely the human ends up on the dumb end of the loop.
Work product that arrives looking finished blunts the review. No one wants to admit that, but it’s true. The human brain just approaches clean copy with a lighter hand than a messy work-in-progress. When AI took the document on a speedrun through steps that used to take days, it’s hard to mentally gear up to check every step along the way. Vendors brag about providing an audit trail, but how discerning is a lawyer going to be sifting through all the back and forth along a 12-step reasoning log at midnight? When they used to intervene on each stage with 20-minute office drive-bys during a four-day project, it made sense.
Catching mistakes early matters, because errors don’t average out. At 95 percent reliability per step, each step has an exceptionally good chance of working without a hitch. But screwing up step two just feeds that result into step three. By the end of a 10-step workflow, you’ve played Telephone with your professional output. Researchers have found that per-step accuracy really does degrade as models condition on their own prior mistakes. That paper doesn’t write off AI by any means — the authors argue that scaling can overcome these limitations — but that’s minimal comfort to someone using the technology today.
Most of the anti-agentic backlash points to security. People fret that SkyNet is about to become self-aware, but instead of a nuclear strike, it’s going to email your vacation photos to your boss. But security strikes me as a solvable problem. For instance, a working paper posted to Zenodo this past week laid out a plan for agents without resorting to categorical bans. Eric Swidey’s “Bounded Execution” architecture gives the agent no credentials, routes every consequential action through an independent control point, verifies its factual assertions against authoritative records… and treats this as a design requirement. Which is all to say that whatever technical problems exist, someone will solve with a technical solution.
Human judgment is another story. That’s the tool that can’t be sped up. Unless you trust that Elon Musk embedding chips in monkey brains will end in the technological singularity and not, say, a bunch of otherwise normal people who suddenly have strong feelings about white replacement theory.
This is the dromological displacement argument — the notion that accelerating a process doesn’t merely compress it but fundamentally changes it. The way the human brain thinks about a problem while that document bounces around the team org chart for five days is different than how it thinks about “just check and clean up this AI draft.” This might be over romanticizing the work of an attorney, but there are epiphanies that arrive in the shower on Thursday because the matter was still open Wednesday. Compress the drafting process to 11 minutes and — even after the human-in-the-loop edit — you don’t actually get the same brief but faster. It’s a different animal.
The apocryphal story is that we’ve been saddled with the QWERTY keyboard layout because its inventor separated common letter pairs like “th” and “he” to cure the constant jamming of early typewriters that couldn’t handle the speed of typists triggering these neighboring hammers in rapid succession. People developed a more inefficient interface to prevent a technological failure. This probably wasn’t the reason for the bizarre keyboard we all know and love… a 2011 Kyoto University paper argues that the layout evolved to serve telegraph operators transcribing Morse rather than to slow anybody down.
Here I am comparing AI to typewriters like a hack. But stick with me because unlike with that other guy, my story carries a lesson.
AI continues to improve and its capacity to take a legal task from concept to completion gets better all the time. But the human brain isn’t advancing as quickly, and that’s the stumbling block. If “human-in-the-loop” means anything — and especially if it means “let’s not commit malpractice” — then the process needs to slow down to let the humans keep up.
Lawyers need to keep using AI because it’s a disservice to clients to tie a hand behind your back. But remain cognizant that giving a 30-minute edit to an AI brief is not the same as when four lawyers read it a combined 30 times over the course of a week. There’s an anti-AI trope that “writing is thinking,” which misses the mark. Allowing AI to write a draft is not a wholesale abdication of thinking any more than trusting a first-year to take a stab at the first draft was. But “writing” includes editing and editing is thinking. Human lawyers shouldn’t shortchange that process.
And that’s the real challenge of “agentic” AI. The quality these days will be at least good enough. But making sure it’s excellent for the client requires lawyers to fiercely defend the slow process of editing and mid-stream correction. Unfortunately, the economics of law will push back on this. Clients will wonder why it’s taking so long, firms will push lawyers to get the draft out the door and start billing on something else, and management will complain about hiring another set of eyes after spending so much on AI… but the human workflow needs to be adapted and protected in this process.
Or we’ll all get working monkey brain chips and then it’s fine. Whichever comes first.
Joe Patrice is a senior editor at Above the Law and co-host of Thinking Like A Lawyer. Feel free to email any tips, questions, or comments. Follow him on Twitter or Bluesky if you’re interested in law, politics, and a healthy dose of college sports news.