You asked for comments and anything else high level. Here it is, plus the two tabs showing what I thought before the ride and how much of it the ride destroyed. I score myself in public so you know what my calls are actually worth.
Your ten in full, each one with my marks underneath it. Six I would not change. Two I want to push on. Two need something added.
An AI company for the legal industry. Software, services, or both. The form follows what the market tells us.
Nothing to add. Form follows market is right, with one caveat that lands in point 4.
The blocker in legal AI was never the tech. It's confidentiality, ethics rules, and fear. Every vendor pitching law firms speaks one language. We speak both: you build with domain experts; I live inside the rules they're afraid of. That's the moat no matter which product we pick.
You wrote this better than I did. I had it as a regulatory cost argument. You have it as a language argument, which is the real one.
The moat has an expiration date. Every year more lawyers become AI-literate and more vendors hire compliance people. This advantage is widest right now and narrows from here. It is an argument for moving in months, not quarters.
Big firms spend millions on AI already. Solo lawyers won't pay. The middle, firms of roughly 5 to 50 lawyers, knows it's getting left behind and doesn't know who to trust. Thousands of them, real budgets, no credible option.
Five to fifty is a much better cut than my "mid-market." Solo won't pay is the part most people get wrong.
Within five to fifty, which practice areas? Construction litigation is a real segment, it is yours, and firms in it are exactly this size. Starting where you already have standing beats starting where the market is theoretically largest.
All open, all to be explored:
My first pass at this said "do not rank them, sequence them." That was too clean and I was wrong about it. Two of your six are sellable next month, so sequencing alone still leaves you choosing between them, which is a ranking problem.
The real structure is rank inside a tier, sequence across tiers, where the tiers are set by clock. Ranking a three-week engagement against a six-month build is the category error. Ranking two three-week engagements against each other is the actual decision.
All six worked through with marks, plus the criteria and one reversal of something I told you in the car. Its own tab.
The market decides, not us. Conversations with mid-market decision-makers, one question underneath all of them: what would you actually pay for this? Rank the board by what we hear, then bet.
Agree the market decides. Disagree with the abstraction.
"What would you actually pay for" asked cold gets polite nonsense. Buyers describe what they think they should want. They react honestly to something concrete. Bring one specific thing to every conversation and watch what they do with it, rather than asking them to design it.
You already know one thing they would pay for, because you raised it yourself. See the additions tab.
Your sparring model: both prep solo, exchange, iterate hard, be brutal, repeat. Point a direction, narrow it, run the experiment, learn, pivot. Every cycle sharpens the offer even when it kills an idea.
Nothing to add. You already ran the first cycle faster than I expected.
You: build, operate, run the client-facing motion, own delivery. Me: legal domain depth, the compliance architecture that makes firms trust us, network into the buyer side, strategy. Structured so nothing bottlenecks on either of us.
You gave me the client-facing motion and delivery. In the car you told me your prior company was effectively a sales company and that a paying customer in thirty days was not a question.
So which is it. Is "network into the buyer side" you making warm introductions and me closing, or is that a bandwidth call, or is it a status call you have already thought through and not said out loud?
Any of those answers is fine. I just do not want to build a plan around a role split that quietly exists for a reason neither of us has named. If it is the third one, that is more evidence the layer underneath this cannot stay deferred.
Time and intellect first. No one quits anything, no big capital moves. Cheap experiments, honest kill criteria, repeat until something hits.
Kill criteria get written before the experiment, not after. Honest kill criteria decided at the end are just rationalization with better vocabulary. Write down what result would make us stop, then run it.
And put a date on cycle one. Thirty days from an actual day, not thirty days from whenever we get around to it.
Formalize when exploration says there's something real. When we do: you in operational control, equity reflecting contribution over time, with a rebalancing mechanism from day one so early logistics never harden into the permanent split.
The rebalancing mechanism is the right call and rarer than you might think. Most partnerships fix the split during the month when contribution is least representative of what either person will actually do, then spend years resenting it.
Dynamic equity also happens to be friendly to the layer you set aside. A stake that starts small and grows on contribution is easier to structure around a work authorization constraint than a fixed split negotiated up front.
My own firm is coming. A live proving ground and reference customer. Everything we build gets battle-tested on real matters before we sell it to anyone. Almost no legal-tech company gets that. Not the plan; the tailwind.
You buried the biggest thing in the document at number ten and called it a tailwind. It is more than that, and it collides with the layer you set aside.
A lawyer starting their own firm is self-employment. That is precisely the shape the 51% advice exists to prevent. So point 10 and "structure and work authorization are their own layer beneath this" are not two separate topics. Point 10 is gated by that layer.
And it changes my read on the ABS. I told you it was the wrong first move. That was correct when the assumption was that we build a services company alongside your existing job. If a firm of yours is coming, the ABS stops being an exotic option and becomes the structure that lets me co-own the thing you are already going to build. Arizona is the only state where that is legal.
Which makes the question I asked in the last doc suddenly load-bearing rather than academic: are you admitted in Arizona, or practicing on the California license? If the firm is going to be an Arizona firm, that is the first domino.
Five criteria. The one that matters most early is not revenue.
For the first tier, rank on evidence yield, not revenue. Tier one's job is not to be the business. It is to buy access and information cheaply while someone else pays. Ranking it on revenue picks the wrong one.
Zero build. These are where cycle one lives.
The most underrated thing on your board, and I think you filed it too low.
The unlock is CLE. Arizona requires 15 hours a year with 3 in professional responsibility, on an annual cycle, and at least 10 of the 15 have to be live or interactive. Arizona also does not pre-approve providers, attorneys self-evaluate and report by affidavit. So the accreditation barrier is close to zero, the ethics category is the hardest one for lawyers to fill, and AI ethics sits directly in it.
That changes what you are selling. Not training, which is discretionary, but a requirement they already have to satisfy, in the category they least want to spend it on. You are also not cold-calling anymore. You are standing in a room full of exactly the buyers we want, as the credible person, before anyone has been pitched anything.
Recurring by construction, since the requirement resets every year. Compounds, because the material is built once. And it needs you specifically, because a non-lawyer teaching lawyers their own ethics rules does not survive the room.
Higher ticket, and the natural close from the room the training fills. Firms are half-buying this from people who cannot deliver it, which is your own phrasing and it is correct.
The deliverable is concrete: an AI governance policy, approved and prohibited tools, a confidentiality analysis, a supervision protocol, a client disclosure position, and a written record. ABA Op. 512 directs managerial lawyers to establish clear policies, which most mid-market firms have not done and know they have not done.
Ranked second only because it does not compound on its own. Nobody buys a governance policy twice. It is a door, and a good one, but the door has to lead somewhere.
Highest evidence yield on the entire board. You are inside the actual workflow watching what people really do, instead of hearing a description of it on a call.
And the most dangerous one. This is the classic path to becoming an agency: every dollar tied to hours, two people, nothing compounds, and you wake up in a year with revenue and no asset.
The rule that makes it safe: before any per-firm build, name the component we expect to reuse. If we cannot name it, we are taking the engagement for cash and evidence only, and we say that to each other out loud rather than pretending it is product work.
You listed "one practice area done end to end deep" without naming one. Name it construction, because it is yours and the fit is unusually good.
Change orders, RFIs, submittals, daily reports, schedules. Delay and disruption claims are essentially a document-timeline reconstruction problem across enormous volumes, which is close to the best case that exists for this technology. The market is narrow enough that nobody large is building it and wide enough to be a company.
Highest defensibility on the board, and it needs everything above it to have happened first.
I am walking part of this back. In the car I called it the wedge. On more thought it is a strong product and a bad first move, and the reason is data rather than AI.
Prior testimony and deposition transcripts are not sitting in one accessible place. Federal filings are reachable, depositions largely are not. The moat here is data access, and there are funded incumbents who already spent years assembling it.
What keeps it on the board is the buyer. This gets expensed to a client matter rather than fought for out of a firm's software budget, which is a materially easier sale than anything else here. Still the right long bet, just not the opening one.
Real problem, and mid-market firms genuinely cannot adopt agentic tooling because they have no engineers. Also the thing I would personally most enjoy building, which is exactly why I am ranking it last.
It is a platform bet, and platforms need scale two people do not have. It is also the most exposed to being commoditized by the model providers themselves, who are moving toward this on their own timeline and are not going to stop.
No tier three build starts until a tier one or two engagement has paid for it and produced the specific evidence that picks it.
That single rule does most of the work your point 5 was reaching for. The market still decides, but it decides by what firms pay us for and what we see inside them, rather than by what they tell us on a discovery call. It also means we are never sixty days in with nothing shipped, because tier one ships in weeks.
Four things at the same altitude as your ten.
It is not in your ten. It was the sharpest thing you said in the car.
Every vendor calling these firms is selling hour savings into an hourly business. They are pitching the buyer on destroying his own revenue and cannot understand why the deals stall. Nobody in that queue is solving the compensation problem, because solving it requires knowing how firms actually make money.
Under ABA Op. 512 billed time has to reflect actual time spent, so a firm cannot quietly keep billing the old hours either. The way through is capacity, throughput, work they currently turn away, or moving specific matter types to flat fee where compression becomes margin.
That is a consulting engagement a firm would buy tomorrow, it requires nothing built, and it is the one thing on the board only you can credibly sell.
The duty to independently check AI output does not transfer to a vendor and does not go away. Uncritical reliance is described in ABA Op. 512 as almost certainly malpractice.
So whatever we build has to emit a trail a partner can check in minutes. If checking takes as long as doing, we have sold nothing. This is not a disclaimer at the bottom of a page, it is a design constraint from the first line of code, and it is a real reason to build rather than resell someone else's tool.
Expanding point 4. Consulting and training are revenue in weeks and the best research money can buy, because you are inside the firm watching the real workflow instead of hearing a description of it.
The product bets should be chosen from that, not from discovery calls. It also means the thirty-day bet is winnable, which matters more than which product we eventually build.
Two things are closing. Lawyers are getting AI-literate, which shrinks the fear that is currently our moat. And the vendors pitching these firms weekly will eventually hire compliance people and start speaking both languages.
Neither closes this year. Both close. It is the argument for shipping something imperfect into a real firm soon rather than exploring six directions well.
Built cold, from a voice memo and public rules, before I knew anything about you. Eight calls.
A company that sells to lawyers, or an Arizona ABS that employs them. Everything else is a variation.
The ABS needs court licensing, an Arizona-admitted compliance lawyer, a governance plan, and drags in California fee-sharing restrictions.
Prior testimony, exclusions, Daubert history. Narrow enough to ship, valuable enough to price.
The duty to check does not transfer, so verification speed is the feature.
Ownership and control are separable. Minority stake, limited voting rights, you keep your job.
A California card does not fill an Arizona seat, and admission on motion runs on reciprocity California does not offer.
In-house counsel authority reaches the employer only.
A domain expert with a demanding job is enthusiastic in the room and gone in a month. Find out who signs and whether there is real room for this.
You had not read anything before getting in the car and opened by saying the two avenues were the exact same ones you had come to.
Stronger than predicted. I argued for the services path on regulatory cost. You argued for it on market size, which is the better reason. Note that P2 is now partly reversed by your point 10, see the marked-up tab.
You laid out how strictly the field is governed and how afraid lawyers are of breaking a rule, then said unprompted that the dissonance itself is the product. We landed on it in the same minute.
You confirmed expert witness research and immediately moved to the real question, how wide versus how specific. That question is now the disagreement in point 4 of your list.
Working only from a voice memo, I had you as a domain expert with a demanding job, which meant the risk I was solving for was attention. Would this survive your actual bandwidth, and could you get anything bought. So I went in planning to test for commitment.
You volunteered it before I asked, and the premise was wrong anyway. You sold a home automation company that white-labeled across cameras, locks and thermostats before Ring and Nest owned the category, and got out when the giants moved.
The correction: you are not a domain expert I sell through. You are the commercial half of this.
Which is why your point 7 caught me. You handed me the client-facing motion and delivery and kept strategy and network. That is the opposite of what the car suggested, and it is the question I flagged on your list.
Correct given the assumption I was working from, which was that we build alongside your existing job. Your point 10 breaks that assumption, and the conclusion goes with it.
Law firm versus construction company, and Arizona admission. I rated these the top blockers, we talked for thirty minutes, and neither surfaced. Not a failure of the night, but the Arizona admission question went from academic to first domino the moment you wrote point 10.