Work
Work
Three write-ups of the kind of brief we take. They are here so you can see the grain of the work before we talk.
These are illustrative composites built from the kinds of requests we take. No client is named and no figure here should be read as a measured result.
Write-up 01
A small legal practice and an archive of agreements.
The practice had years of executed contracts in shared folders, named by habit rather than by a scheme a new associate could trust. When a partner needed a clause they remembered signing, the search was a morning of opening PDFs. The knowledge was in the files. It was not in a place a person could query without already knowing the filename.
What got in the way was not a shortage of software. The firm already had a document store. The store could find a title. It could not answer “which of our supplier agreements lets us terminate on thirty days, and on which page.” Associates were reconstructing that answer from memory and from the last person who had looked.
We built an assistant over the firm’s own agreements. Each reply pointed to the PDF and the page. Access followed the existing folder permissions, so a trainee who could not open a matter could not ask about it either. A log of queries gave the partners a view of what people were actually hunting for, which later fed the evaluation set.
Checking meant a hundred or so questions taken from real matters, with answers the partners agreed. We re-ran the set whenever a new batch of files was added. The assistant was allowed to say it did not know. A confident wrong clause is worse than an empty result that sends someone back to the folder.
What stayed with the practice: the assistant, the citation path, the log, the question set, and a short guide for the person who adds new executed files. Orienting figures we use when talking about this shape of work: a search that used to take the better part of forty minutes now starts with a shortlist in a few minutes, still read by a lawyer before anyone relies on it.
What the tool does not do
- It does not draft new agreements or rewrite a clause.
- It does not give a legal opinion. A named lawyer still signs the advice.
- It does not read files outside the folders it was given.
- It does not replace the document store. It sits on top of it.
Write-up 02
A logistics dispatch desk and a noisy inbox.
The desk received booking changes, gate queries and “can you take this as well” notes as ordinary mail. Each message was a small puzzle: which booking, which field, which sheet row. The staff knew the puzzles. The volume meant the morning was typing rather than deciding. Nothing in the mail was exotic. It was simply unsorted.
The obstacle was the gap between free text and the table the warehouse already trusted. Forwarding everything to a chatbot would not have helped, because the chatbot would not have written the row, and would not have known which messages were the odd ones that still needed a person. The desk did not want a new system of record. They wanted the existing sheet filled with fewer keystrokes.
We built a pipeline that read the incoming mail, guessed the type, pulled the fields the sheet already had columns for, and wrote a draft row. Messages that did not match a type, or that conflicted with a known booking, went to a short queue. The queue was visible on the same screen the desk already used. Nobody had to open a second tool to find the leftovers.
We checked the path on a few hundred historic messages, with the desk marking which draft rows were fit to keep. The types were allowed to be few. A pipeline that claims to understand every polite ramble will hide the cases you most need a person to see. After that, new mail ran through the same path, and the evaluation set was a sample of each type plus a bundle of the leftovers.
What stayed on the desk: the path, the queue, the map of types, and a way to add a type without calling us for every new phrase a customer invents. Orienting figure: in a week of this shape, about seven messages in ten could be filed without a person touching every field. The other three were the job.
What the tool does not do
- It does not send mail back to a customer on its own.
- It does not change a booking in the warehouse system without a person confirming the row.
- It does not read attachments it has not been shown how to parse.
- It does not decide which customer is more important.
Write-up 03
Clinic administration and letters that still need a person.
An administrative team in a clinic prepared discharge summaries and letters from notes that already lived in the record. The clinical decisions were not theirs to make. The typing was. A letter that should have taken twenty minutes was taking the better part of an hour because the notes were long, the template was strict, and every letter still needed a clinician’s eye before it left the building.
What slowed the team was the fear of missing a line that a later reader would treat as fact. They copied carefully. They still worried. A tool that wrote the letter and sent it would have been the wrong tool. The brief, once we sat with them, was narrower: assemble a draft from the notes, show the source of each block, and refuse to send anything until a named person had ticked the page.
We built a drafting path that pulled from the record the team already used, filled the template, and highlighted the spans it had used. The send button stayed with the clinician. The log recorded who had ticked, and which version of the notes had been in force. Access was limited to the administrative role and the clinicians who already opened those records.
Evaluation was a set of historic letters, with the team marking where a draft would have been unsafe to tick. We cared less about fluency than about omissions and about invented details that were not in the notes. The harness failed a draft that added a medication the record did not contain, even if the sentence was well written.
What stayed with the clinic: the drafting path, the mandatory tick, the log, and the question set. The letters still left the building under a person’s name. That was the requirement, and it is the requirement we will not bargain away on work of this kind.
What the tool does not do
- It does not diagnose, prescribe, or choose a treatment.
- It does not send a letter without a named person’s confirmation.
- It does not invent a fact that is missing from the notes. It leaves a gap.
- It does not replace the clinical record. It reads from it.
Pattern
What work of this kind tends to share.
The useful material is already there. It is in folders, inboxes and records that a team has lived with for years. The job is to make a Tuesday faster without pretending the archive is a blank page that a model should fill.
A person remains in the chain wherever the output changes something in the world: a clause relied on, a booking altered, a letter leaving a clinic. The tool prepares. The person still signs.
The evaluation set is not a flourish at the end. It is how we agree, in writing, what “good” means, and how we notice when a later change has undone it. Without that set we will not take the build.
The tool is allowed to be small. A search that cites a page, a pipeline that fills a sheet, a draft that cannot send itself: these are complete products. They are also the ones a team of five to sixty can keep after we have gone back to Bussorah Street.