Twenty-One Minutes

Cooper’s floor had entered the quiet season. The summer associates attended their last social happy hour event a week ago and had turned in their badges as they headed back to school for their 2L or 3L year. There would be a few weeks before the next batch of fall associates arrived. It was also the period where more emails were answered by out-of-office auto-responses than actual replies. This is the little stretch of a few weeks where he considered this to be the transition to the firm’s true new year, when the office empties out enough for a person to really have time to think for themselves.

Cooper planned to use this time to knock out what he had assumed to be the easy part of his new job. His name was on the card for the new evaluation harness, with quarterly review requirement that he had no intention of missing, and the first deliverable was the trial set. Forty closed matters were assigned. Billed, paid, and no additional litigation troubles. The rubrics would be built from these forty matters, gathered from the attorneys who worked on them, and then frozen in the same way they did the surveys to the firm. That was the design that Cooper, Maya, Leo, and Jesse set up in the windowless conference room. Now all he had to do was actually select the forty matters.

He sent a request to the accounting department for an export of closed matters going back three years. It came back the same afternoon. A spreadsheet with eleven thousand rows. Within two hours of scrolling through the spreadsheet, Cooper understood that picking which forty matters was not going to be the easy part at all.

The initial problem started in the practice area column. The firm had an official list of practice groups. Cooper had actually helped update it two years ago. What the intake system collected, however, was a text field, and text fields are the place that official lists go to die. The same corporate work showed up as M&A, as Mergers & Acquisitions, as Corporate - M&A, and even one weirdly labeled as m+a misc. Those were four descriptions of the same kind of matter, written by four very busy legal assistants, on very busy days. The system just accepted every one of them without any pushback.

The industry column was worse. Many of the cells were just empty. Of the first sixty matters that Cooper flagged as candidates, nineteen of them had nothing in the field at all. A rubric that ask whether the work product served the client’s industry can’t do much with a blank.

Then there was the client he found filed under three differently spelled names. Same client. Three versions of the name, two of them abbreviated differently, all three carrying separate matter histories. The only reason Cooper caught that one was that he knew this client and wondered why the matter count was so low.

The one that made him put down his coffee and shake his head was a pair of matters that turned out to be one matter. It was opened twice, eight days apart, by two different assistants in different cities. Two different client matter numbers. Both went through conflicts. Both were staffed. Somebody eventually noticed, because the billing discrepancies. The second matter was closed with a note saying see other file. The note did not actually give the matter number.

After working on the list for a few hours, Cooper flipped open his notebook and started a list. He titled it: Ways we disagree with ourselves. By the time he wrote out what he already knew, he had eleven entries. He stared at it for a few minutes and then wrote right below it. The evaluation harness needs forty client matters we agree on. Before that, we need to agree on what an actual matter means.

At that point, he didn’t realize that a Berkeley Computer Science professor was about to tell him the same thing.

Cooper had first seen reference to the video on LinkedIn. Then he saw it in a newsletter he subscribed to under the title of Ontologies are making a comeback. That didn’t catch his attention because quite honestly, he wouldn’t have been able to define ontologies if you held a gun to his head. He did remember that he’d covered the topic in library school, but that was a lifetime ago, and probably a required class he really didn’t want to take in the first place. Then the video was forwarded by two people within a few minutes of each other. One of them was a client he occasionally talked tech with. When the firm’s clients start sending you tech videos to watch, you tend to watch them, even if you don’t want to.

The video was twenty-one minutes long and Cooper watched it at 1.25x speed. It was from an AI conference that Cooper had briefly considered going to earlier that year, even though it wasn’t a legal industry conference. The professor opened up with a sort of hippy-artist’s creed, borrowed from a nun of all people who taught art in the sixties, and made famous from an avant-garde composer a few years later. “Nothing is a mistake. There’s only make.” As Cooper let that bounce in his head, it took on a Yoda-like feel to the way it sounded.

The professor pulled on this idea to the five thousand engineers sitting in the room. Get your hands moving and learn by making things. He asked them to stop typing out their notes and get back to using a pen, because typing makes your brain think about the keyboard, and when you write it out by hand, your whole intelligence shows up and participates.

Cooper looked over at his notebook and pen lying there next to his lunch. He felt sufficiently validated, and he moved the speed of the video back to 1.0x speed.

Cooper reconstructed the Berkeley professor’s case after watching the video, and rewatching parts of it a second time.

The thinking behind AI agents comes from the decades old dream that computers can play out situations, make decisions, and act out the scenarios. Large Language Models and Generative AI tools are sold at trade shows that sell this dream. Cooper watched it first-hand in Las Vegas when the same claims investigation ran twice.

Ontologies go back millenniums. The professor traced it back to at least Aristotle, who wanted to formalize a method of sorting what existed into categories. The modern version of ontologies got a working definition back in the nineties. Cooper wrote it down: a formal, shared way of writing down how a domain fits together. Stripped down to the basics, ontology is simply a dictionary with rules attached to it. Here are the things that exist in our world. Here are the types of relationships they are allowed to have. Here’s their properties which they carry, and the values that each of those properties can take on. These are written down in a way that a machine can check its work against it.

The data exists as a graph. The things, and the lines between those things.

The next part in the talk made Cooper set down his sandwich. The model by itself is a phenomenal talker, locked in a room. It can reason, it can propose ideas, it can advise you on what to do next. It cannot actually do anything. Cooper had heard this type of philosophy before from Jesse, furnished as an office location with a genius inside. It was somewhat comforting to hear the professor describe the same tenant.

That comfort was broken slightly when the professor continued on past the idea of the office location. The walls guide the genius on what it can touch. They do not define what the words mean. And this ambiguity is what causes agents to break in production. Not dramatically, but quietly, plausibly, in the gaps between what prompting instructions say and what a definition would enforce.

Cooper leaned in on the screen as the professor gave three examples of the kind of errors that a well-meaning prompt cannot reliably stop.

• A second refund issued on the same order.

• A payment sent to the support desk instead of the customer.

• An order status that read “probably shipped.”

Every one of those, he said, dies instantly when you have a one-line rule in a formal dictionary. A refund can only happen once per order, so the second refund is a contradiction, killed before it executes. A customer is a different entity than a support rep, so money can only be routed to one type of entity and not the other. When a status can only be one of three values so that things like probably shipped is immediately rejected because it’s not one of those three on the list.

Cooper moved his uneaten lunch to the side, pulled his notebook closer and turned back to the page he wrote: Ways we disagree with ourselves.

At the fourteen-minute mark, Cooper paused the video again and began writing in the margins.

The second refund on the same order. He began writing next to it: our matter opened twice. Two different offices, just eight days apart, they’d been issued different matter numbers and had both gone through the conflicts process. Nothing in the firm’s intake system knew to identify that a matter opening is the kind of thing that only happens once. No rule anywhere that could flag a second matter opening and recognize it as a contradiction in process.

And the payment that was routed to the wrong party. That one had an even colder translation, but Cooper wrote it down anyway: a document crossing to the wrong side of a wall. The walls were design to hold because the systems were set up to enforce who could reach what. But the professor’s point was a level deeper than just access. What saves you in this instance is that it knows the two groups are defined as not to overlap at all. That way membership in both is not a permission issue, it becomes an impossibility. And the machine is simply not allowed to do it.

Then came the probably shipped issue. Cooper looked it over and then began listing all of the “probablys” the firm had. Statuses that live in spreadsheets and in answers you get in the hallway, rather than in any system with a list of actual allowed values. Probably conflict checked. Probably filed with the courts. Probably the final version of the contract. And the one he knew Maya would not like, probably privileged. Right now, somewhere in the firm, a review status has the effective status of probably privileged, and there wasn’t a machine anywhere that would refuse it, because none of the systems are set up to say privilege review only has three outcomes and the word probably is not in any of those three.

He hit the spacebar to restart the video and one minute later pressed the spacebar again to pause it. This was the part that reminded him of Leo’s third folder, where a supply agreement was labeled as an NDA filename, and the harness rejected it. The harness saw that the structure didn’t match. Cooper placed in the margin: Ask Jesse about this. When the harness rejected the third folder, what was it checking against? It must have known what an NDA looks like. How did it learn that? Where does that knowledge live?

The last part of the video focused on history, and Cooper found this as interesting as the rest.

The professor discussed living through the eighties version of what were called expert systems. Where the plan was to write down every rule and then let the rules do the thinking. Entire companies were established on this idea. Lots of money was thrown at it. Even the professor’s own son studied Japanese in college because they were leading in this area of rules-based machines. But the systems became overwhelmed and couldn’t scale to the level needed to keep up with more and more rules. The whole field went into an AI-winter that took decades to thaw.

Cooper noticed that the professor chose the wording of his conclusion very carefully. Rules fail. Models drift. It takes a pairing to make it useful: use the model for the reasoning, and a written dictionary is kept outside of that model, to check the work. The dictionary keeps the model honest. The model makes the dictionary worth maintaining, because for the first time the dictionary has a tireless reader.

The video closed much like it opened, instructing everyone to go make something. Cooper looked at the final timestamp of twenty-one minutes and seventeen seconds. Then he refreshed the page and watched the entire presentation again. He had his pen ready to add to the notes. This was the first time in a long time that he remembered doing this.

The payoff of rewatching it was capturing a sentence that had been eluding him all afternoon. The firm had been building rules for a year. The binder he had is full of these rules, and the boxes Jesse drew on his whiteboard had turned the best of those rules into code. What the firm lacked was agreeing, in writing, in a single place, on what its own words really mean. The system was set up on guesses. The spreadsheet with eleven thousand rows was just an exported version of those guesses.

Cooper stayed at his desk while the floor went from quiet to empty. He opened up his email and sent Jesse a quick four sentence note. He’d found a video worthy of Jesse’s time. He had a list of eleven ways that the firm’s own data disagreed with itself, and he attached it. He wanted an answer to Leo’s third folder result, and thought he might know the answer. Lastly, he calendared time on Jesse’s schedule for first thing in the morning.

The response came back before Cooper had returned from the breakroom. Jesse had watched the video last week. He finished off the email with bring the list, this should be fun.

Cooper pulled open the notebook one more time before he left for the night. He flipped through to the page after the printed whiteboard photo, and wrote in the final lines for the day.

The professor spent twenty-one minutes listing the errors that instructions alone simply can’t stop. Every one of those were recognizable in the work we do in this building.

Then directly under it:

A rule a machine can enforce begins as a definition someone writes down. We have binders full of rules. Instead of finding our definitions today, I found a client spelled five ways.

LINKS:
Why Agentic Systems Need Ontologies - Frank Coyle, UC Berkeley