Home / Articles / Legal on demand for startups: agent teams that do the reading, lawyers who sign, and a meter instead of a retainer, with three vaults that show the founder's desk, the investor's desk and the deal room

Legal on demand for startups: agent teams that do the reading, lawyers who sign, and a meter instead of a retainer, with three vaults that show the founder's desk, the investor's desk and the deal room

By Dinis Cruz · 2026-10-27 · article v1.0.0 · newsroom v0.13.109 · legalstartupsinvestmentagentspkivaultsbusiness-modelsarticle

Peer review Not reviewed yet. Review it

Abstract: Legal advice is in every part of a startup, from shares and votes to IP and people, and most founders have little experience of it and little access to good lawyers. So they sign deals that are worse than they need to be, often with no malice on anyone's side. Legal services never scaled because the work took time, much of it repetitive, and good people did not want to do it. GenAI changes that: models are very good at reading every detail and finding inconsistencies, and an agent team can do the reading, the checking and the drafting so that a lawyer's time goes on judgement and signature. That makes a service that scales: on the founder's side, on the investor's side, and in a deal room where both teams meet, with every understanding, clause and version signed and verifiable. Paid per task or by a monthly plan, not a yearly retainer. I think someone should build it, and most of what it needs exists today in open source. To show it, three made-up vaults: Tallowmere Counsel's founder desk, investor desk and deal room, with real signatures your browser checks.

In short

• Legal is in everything a startup does, and most founders have neither the experience nor the lawyers.
• Bad deals are often not malicious. If you hold the power, why wouldn't you ask for more?
• Legal never scaled. The time, the onboarding, the repetitive reading.
• Models read every detail and find inconsistencies. That is most of the repetitive work.
• Agent teams drive the lawyers' efficiency. Not a human in the loop: humans inside a workflow.
• Charge per task or per month, never a yearly retainer.
• Investors need it too, and will lose deals to investors who have it.
• Sign everything: principles, terms, clauses, versions, understandings.
• Publish your terms, and let investors debate them with their own teams.
• Three vaults show it working, with made-up companies and real signatures.

The problem: legal is everywhere, and founders are new to it

This follows the jobs that never scaled, applied to the law. Having someone who understands the legal side in the conversation is critical in almost every part of a startup: shares and the cap table, voting rights, intellectual property, contracts with customers, human resources, data protection, and above all the investment. And most founders have little experience of any of it. They don't know which questions to ask, what a cap table will look like after two rounds, or what a participating preference does to them in a sale. They often don't have good lawyers, or any, or not ones who know the particular thing that matters.

So they get worse deals than they should. Sometimes it isn't even malicious: if you hold the power, why wouldn't you ask for more? But the question I keep asking, as in the controllable founder, the exit is when you sell control and sell, don't issue, is why an investor wouldn't align its terms with what the founder is happy with, rather than with what it can get. Part of the answer is that the founder can't argue well enough. That part can be fixed.

Why legal never scaled, and why it can now

Legal services never scaled because they were expensive, and they were expensive because of the time: onboarding, reading, checking, re-reading the next draft, a lot of repetitive work that the best people did not want to do, and that clients paid senior rates for anyway.

Models change exactly that part. I have used ChatGPT and Claude for real legal work, quite successfully, and here is what people miss: even when they hallucinate, they are very good at reading every detail and picking up inconsistencies. Three things I have done that worked:

• Mirror terms. If a contract asks me for a particular level of indemnity, or a particular term, ask for the same on the other side. Terms and conditions became much fairer that way.
• "Here is my understanding." When I am not fully confident, because I am not a lawyer, I write to the other side's lawyers: here is my understanding of what this clause does, do you agree? If they say it says X and the contract says twice X, that is now their problem, not mine, and if the text changes halfway through, it is misleading. Those things matter in a relationship, and especially at the final due diligence.
• Infographics of terms. Turning terms and conditions into a picture shows what they actually ask.

And when the terms are right, the models agree with each other. Ask ChatGPT and Claude separately, and when they both say "this now covers your bases", it usually does. They are also very good at finding blind spots. None of this replaces a lawyer. It replaces the part of a lawyer's time that should never have cost what it did.

The service

So here is the service. A legal desk for startup funding, where agentic workflows drive the efficiency of the humans inside them. This is not a human in the loop, approving what a machine did. It is agents doing the reading, the extraction, the mirroring, the cap table models, the plain-language explanations and the checking, so that each lawyer's hour goes on judgement, negotiation and signature.

The founder does as much as they can themselves, with the platform: depending on their skills and their time, more or less. The platform tunes its prompts and workflows, not its models; fine-tuning a model is the last thing you should do. And the founder can always go back to their own Claude or ChatGPT and ask "what do you think of this?", which a good service should welcome.

The commercial model follows from the economics. With agents doing the repetitive work, and lawyers working virtually with low overheads, the service can charge a fair rate for an hour, half a day or a day of a lawyer's time, plus a price per task. A monthly plan you can cancel any month is fine. A yearly subscription makes no sense for this: it is a SaaS model that suits the vendor's recurring revenue, not the client. Better still, pay per use, which I would argue is also better for the platform.

Three markets: the founder, the investor, and both

The founder's side. Counsel paid by the founder, on the founder's team, who understands the intricacies, and who doesn't cost a ridiculous amount of money. Investors benefit too: by the time they arrive, the data room is in order and the due diligence has largely been done.

The investor's side. I would argue investors need this even more. The difference is between someone working in a spreadsheet and email, and someone with an agentic workflow that can read a whole data room, collect everything, raise the questions and draft the memo in hours. Investment is a competitive sport: as a founder, I want an investor who works at this speed, and an investor who works at analogue speed will lose deals to one who doesn't.

Both, in one room. Each side goes to its own team, agents plus humans, and says: this is what I want, these are my terms, this is my situation. The two teams hash it out. And who pays? I would want the investor to pay for its own diligence: why should a founder pay to answer questions for an investor who may not make an offer? Whoever pays the bill will have the service, and even the models, a little more aligned with them. That is fine, as long as each side has its own.

Sign everything, and keep the intent

The part that makes this more than faster drafting is the signing. With a public key infrastructure, every party has a key pair, and everything that matters is signed: the founder's principles, the term sheet, each clause, each version, each "here is our understanding" and its confirmation. A signed clause cannot change without the signature failing, and every change between versions is a diff with a name on it.

I also think contracts should become a graph, what I call a fractal document: the intent at the top, the principles under it, then the terms, then the clauses, then the plain-language versions in each language. You sign off the intent and the principles; the legal clauses are the way to enforce them. Across languages and cultures, the plain principles travel better than the legal text. And you could argue that a clause that contradicts the signed principles should not be valid: that is a proposal, and the wording still has to be right, but it is the direction.

We did an early version of this in a vault for risk acceptance, before vaults had append lanes. What can be built now is a different order of thing.

Publish your terms

The last part is where I am going with this myself. I am going to raise, on the terms I want, and because everything I do is open, I am going to publish them: the term sheet, what I want, what I am interested in, and the materials. Then I can tell investors: hire this kind of service, or use your own, and debate it with mine until you reach a conclusion. Ask questions, find things, argue. But here are the terms, and this is my line.

That scales for me, because it is a page on my site. I have done versions of this before, for The Cyber Boardroom and for MyFeeds, but I never had a product with real demand. Now I do, which is what makes the conversation with investors interesting.

See it working: three vaults

To make this concrete, I asked the agents to build it. Not a description, a working demonstration with made-up companies: Tallowmere Counsel, the service; Brindlewood Security, a startup that makes an open-source secrets scanner, raising £1.2 million; and Kestrel Row Ventures, the investor. Every person, number and clause is invented, the legal wording is illustrative and not legal advice, and no model is called: the runs are scripted. But the signatures are real: each party has an ECDSA key pair, and every page checks every signature with your browser's own cryptography.

The founder desk is Brindlewood's side: the founder's published intent and eight principles, signed, with a QR code to share them; a readiness checklist and a data room of 18 documents, each hashed; a cap table modeller showing the founder keeping control; the mirror check of the investor's first term sheet, every obligation asked of the other side; plain-language explanations of the terms that hurt founders; and a bill. Its vault: x6kjwxds.

The investor desk is Kestrel Row's side: the pipeline; diligence on Brindlewood's data room, with findings and 12 questions billed to the investor, not the founder; its position on each of the founder's principles; the investment committee memo, signed by the partner; and how long it took, 44 hours from first look to term sheet. Its vault: jgp3oyrm.

The deal room is where they meet. The agreement in five levels, 87 nodes at the final version, each hashed, zoomable from the intent down to a clause in plain English and plain Portuguese. Four versions in five days, with word-by-word diffs. Three signed understandings, and the moment the consistency agent blocked version 3, because clause 7.2 still said "twice the amount invested" after both lawyers had signed that the cap was the amount invested. A signing ceremony where both lawyers, both principals and the service's seal sign the same root. A message log where each signed message hashes the one before. And a meter of 43 signed receipts.

Try the tamper box: change clause 7.2 to say "twice", and every signature on the agreement's root fails, in your browser, because the root no longer matches what was signed. The bills, on these made-up numbers: about £1,100 for the founder's desk and about £1,400 for the investor's, against an illustrative £14,000 to £24,000 for the same work by the hour or on a retainer, with the assumptions shown.

Each is a vault, and you can open it yourself with its published read key:

• Founder desk: open it in the vault UI, or sgit clone sgit_public_read_0857bebbc4f9a317fa55d9b51427c9e41baab8401157b23cb0adc5c1dafd2e14:x6kjwxds
• Investor desk: open it in the vault UI, or sgit clone sgit_public_read_74013a87a8e615202a89e5cffa5ede9bc70715a789b181b4c04d07bbb136625a:jgp3oyrm
• Deal room: open it in the vault UI, or sgit clone sgit_public_read_d1a89809c2acd75df262aa79e29037ac0de40deacbbfd92991f758aba0fd252c:ghafqmwo

The keys are read keys, published on purpose; each was read back before it was published, and the files match.

The business, and who should build it

There is a lot of commercial value here. A service like this has three markets, founders, investors and the deal between them, and a cost structure that legal services never had: agents for the repetitive work, lawyers for the judgement, no office, no retainer. Someone should start this company, or invest in it, and existing legal-tech companies would do well to measure themselves against these features.

And you don't need to go much further than what exists today. Vaults are data rooms, with history and read keys. Append lanes carry messages between the sides. Vault apps are the applications each side uses, with no server. Key pairs and signed records are built, and QR codes carry signatures and sealed messages. Agent teams with cards, roles and loops are running every day. All of it is open source and runs on the client side, which is exactly what you want for documents nobody else should hold. The fractal agreement and binding understandings work in the demonstration; the meter is a published design. Companies in this space should look at it and ask what they can reuse.

What this does not claim

• Not legal advice, and not a substitute for a lawyer: the point is to make the lawyer's time count.
• Models make mistakes. The workflow is built around checking them, with humans who sign.
• A signature is not a legal contract by itself. Whether a signed clause binds depends on the law and the agreement; the demonstration shows the mechanism.
• "A clause that contradicts its principles should be void" is a proposal. It would need careful wording and a willing court.
• The demonstration is made up, scripted, and its three vaults compute their hashes separately, so the same version shows different roots in each.
• The prices and comparisons are illustrations, with their assumptions shown in the vaults.

Where this comes from

The editor's voice memo of 27 October 2026. The demonstration: the founder desk, the investor desk and the deal room, in the vaults x6kjwxds, jgp3oyrm and ghafqmwo on dev.send.sgraph.ai, built by this site's agents from one shared set of made-up facts, with ECDSA P-256 signatures checked by the Web Crypto API. Related: the jobs that never scaled, the controllable founder, the exit is when you sell control, sell, don't issue, paying after the value, the user keeps the meter, PKI everywhere, signed by the source, what an agent is, websites you own. Written up by this site's agent, newsroom.writer; the editorial responsibility is mine.

Want the next issue by email. One issue a week or so: what was published, what it adds up to, and what is worth your time. Subscribe to the SGit Newsroom →

← All articles