The First Thing She Asked For Was the One Thing I Wouldn't Build

A Tuesday signing, a Thursday phone call

My wife runs a family law practice. She signed a new divorce client on a Tuesday. Intake was still in progress on Thursday morning - staff collecting documents, filling in the gaps - when the client called to say she believed her husband was about to empty their accounts. She had heard it through a mutual friend.

Everything the client could reach started arriving that afternoon. Statements, letters, screenshots of account pages, forwarded email threads. More than four hundred documents, in no particular order, some duplicated, plenty of it irrelevant.

The job was to turn that pile into one list: every account, every institution, complete enough that nothing was missed. It had to be done by Friday.

Between the attorney and a paralegal, it took eight to twelve hours. My wife worked late into the evening, drafting the motion and the letters to each institution.

They made Friday.

What that work actually was

There is a fair question buried in that story. Why was the attorney reading bank statements at all? That is what a paralegal is for.

Half of it, she should not have been. Pulling an account number off a statement does not take a law degree. Neither does noticing that March is missing, or that this is the fourth copy of the same page, or that a screenshot and a PDF are the same account. Doing that four hundred times, against a deadline, without missing one, is not a legal problem. It is a throughput problem, and it was being solved by the two people in the firm who could least afford the hours.

The other half could not be handed to anyone. Which accounts were community property and which were his before the marriage. Whether a forwarded screenshot was evidence of anything or just a client being helpful. Whether the pile showed money that was actually about to move, or only a rumour from a mutual friend. Those are the calls she trained for and the ones the client was paying for, and there is no way to make them without looking at the documents. The list also had her name on it at the end. She was the one certifying it was complete, and you cannot certify a list you assembled by proxy - not when the account you missed is the one that gets emptied on Monday.

So both halves sat in the same pile, and by hand there is no way to pull them apart. To reach the handful of decisions that needed a lawyer, she had to go through four hundred documents that did not.

The drafting waited on the same thing. The letters could not go out until the list of institutions was final, and the list was not final until the reading was done. So the motion and the letters - the part of that week that unambiguously required an attorney - started in the late evening, after the part that did not had finished. The clerical work was not just expensive. It was in front.

The other way it goes wrong

A while later, an assistant missed an email from opposing counsel. Inside it was a date: a deliverable due on a specific day. Nobody calendared it. It surfaced when opposing counsel wrote on the day itself, asking where the documents were.

My wife got an extra day and delivered. It cost her a phone call she should not have had to make. It was not the first time something like that had happened.

Both failures are the same failure wearing different clothes. In one, the information was in the pile and there was no time to find it. In the other, it arrived in an inbox and nobody saw it. Neither is a failure of legal judgment.

Five hundred emails, four hundred documents

A typical family matter carries something like five hundred emails and four hundred documents by the time it closes.

Nearly every one of those emails produces something. A task. A date. A fact that changed - and in family law the facts move constantly, the financial ones most of all. Each of those has to land somewhere a human will see it again: a to-do list, a note in the practice management system, an entry on a calendar.

That transfer is manual. It happens hundreds of times per matter, and it only has to fail once to matter.

Why nothing on the market fit

Practice management systems are good at what they were built for, which is storing things and billing for them. They are a filing cabinet with an invoice attached. Getting real value out of one means feeding it constantly, and the feeding is the work.

The AI tools were split two ways. Most were built for firms big enough to have someone whose job is adopting them. The rest did one slice well - draft this document, review this contract - and handed everything else back.

Nothing took a matter end to end. Nothing was built for a firm where the attorney is the throughput.

The first version

I have spent my career building software platforms, most recently as chief product officer at a company whose system runs thousands of field service jobs a day across a twenty-thousand-person workforce.

Looking at that pile of four hundred documents, I could see two days of reading collapsing into an afternoon. It was an ingestion problem with a ruleset on top.

So I built it. A small thing, for one job.

She used it, and liked it. Then she asked what else it could do, which is the question that turns a script into a company. The honest answer was: most of it, but not quickly. It meant connecting the mail, the calendar, the documents and the systems the firm already ran. It meant encoding rules that change by county. It meant handling privileged data properly rather than eventually. And it meant all of that being usable by people who had no interest in any of it.

That is a platform, not a weekend.

Why family law first

Family law has the most moving parts I have seen up close. Multiple parties in constant contact, facts that change month to month, procedure that varies by county, and filings built from numbers somebody has to find in a document first.

The platform is not limited to it. Other practice areas need their own guardrails and their own rules, and those are coming. But a system that can hold a contested family matter together is not going to struggle with a simpler one.

I am not a lawyer

I am not, and I do not pretend otherwise. What I have is a decade of proximity to the operational half of a practice, and a career spent on automation, scale, security and data privacy.

Two practising family law attorneys advise the product. One of them is my wife. The other is a colleague of hers working in the same areas. They tell me what the work actually is, so I do not have to guess.

The first thing she asked for

Here is the part I keep coming back to.

Once the system could read a matter, what my advisors asked for was straightforward and, in their terms, obviously right. The number is already in the record. Stop making me type it again. Put it in the summary, fill it into the form, move on.

I said no, and it took me a while to explain why, because the disagreement was not really about the feature.

To software, a fact is a value with a confidence score attached. The extraction either worked or it did not, and if the model is sure enough, the value is good. That is the natural way to build it, and every instinct I have from building platforms points that way.

That is not what a fact is in a practice. There, a fact is a claim with a history: where it came from, who said it, and what it would be worth if somebody challenged it. A support figure in a certified order and the same figure a client mentioned on a call are the same number and completely different objects. No confidence score can tell those apart, because the difference was never about how sure anyone is. It is about what the document is.

So my model had one kind of fact where the practice had several. Build it that way and it works for months, and then one afternoon it puts a phone call into a filing.

What they were describing was the outcome - the record should not have to be re-entered by hand. They were right about that, and it is what the product does. The part they had no reason to specify was the shape of the thing underneath, and the obvious shape was the wrong one.

So every fact carries where it came from and how well the record supports it, ranked from a certified document down to a phone call. Nothing becomes trusted because the machine was confident about it. An attorney reviews it, or it stays a candidate.

That is what the review queue is. It is slower than what was asked for. It is also the only version I would put my wife’s signature behind.

What other attorneys have changed

For the last few months we have been putting the product in front of family law attorneys beyond the two who advise us. The pattern in what came back has moved the roadmap several times, and each time it has been the team that turned the correction into something shipped.

The first was about scope. We had assumed a firm would want one system. Nearly every attorney told us the opposite: they have a practice management system, they have paid for it, their staff know it, and they are not migrating for anybody. What they want is the part that system has never done - the reading, the watching, the deadline that computes itself. So we stopped trying to be the system of record and rebuilt around working next to one.

The second was about counties. We knew procedure varied. We did not appreciate how much of the variance is unwritten - what a particular court will accept, what gets bounced at the window, how long a clerk really takes. Attorneys who practise across county lines were blunt about it: a rules engine that only knows the codified version will be wrong in a specific and embarrassing way. The team moved the jurisdiction rules out of the code and into configuration a person can correct without waiting for a release.

The third was about who inside a firm sees what. Our model was per-firm - connect the mailbox, everyone works the matters. Several attorneys stopped us on it. A firm has matters where one particular person must not be in the file, and that is not a preference to be configured later, it is an obligation. Per-matter access, confidential matters and ethical walls came out of those conversations, and reworking the access model around them cost the team the better part of a release. None of it had been on the plan.

The fourth we hear every time, in some form: do not send anything on my behalf that I have not read. Not because the drafting is bad. Because the signature is theirs. Every outbound action in the product is approval-gated for that reason, and no amount of accuracy would change it.

None of this came out of a roadmap exercise. It came from attorneys with no stake in being polite about it, and from a team willing to throw away finished work when they turned out to be right.