AI Contract Review—How a One-Person Company Reviews a 30-Page Cross-Border Bilingual Contract

When people think about using AI, the first question they ask is “which AI tool should I use.” The question worth spending real time on is a different one—what should I use AI tools to do.

I’ve roughly sorted my own AI use into categories:

1. High-frequency things I do every day, which AI does instead of me
2. Things I would frequently outsource—finance and legal for a one-person company, for instance
3. High-value things only possible when AI and I work together

Today is about the second category—contract review with AI.

1. Just how painful contract review actually is

Very few people ever meet a lawyer who speaks plainly.

Back when I was at a big company, or anywhere with a large headcount, every contract review went like this: explain the business to the lawyer → lawyer reviews → lawyer picks apart everything (and it never feels like they’re entirely on my side) → talk to the counterparty’s business team → their business team goes to their legal team → several layers of people in both directions, and every layer with different communication habits.

Running a one-person company made it worse. My cross-border business keeps growing, and a 30-page Chinese–English bilingual contract is now routine. A lawyer with the right specialty typically runs over ¥10,000 per contract, with at least a few hours per round of feedback—and your whole tempo gets stuck there.

The worst part isn’t the money. It’s the tempo.

2. Why build your own AI legal agent

I built and polished an agent dedicated to contract review. A few key design decisions:

Drafting and reviewing are split into two agents. Not mixed together, each with a clear role.

No need for an in-house legal template library. The model already carries a huge amount of legal knowledge. Each time I do this work, I send it to the right sources for the latest information and top up the context with what it finds. That updates faster than a static library ever will.

Clause-by-clause review, with every finding linked back to the source text. It is not allowed to say “there’s a problem with one of the clauses” and leave it there.

Advice only, never the final decision. For small contracts I read the review and decide myself; for large-value contracts, the AI’s issue list goes to a lawyer for a second pass. The lawyer is more efficient too—no reading from scratch, just professional opinions on the issues the AI has already organized.

3. The seven-step review flow

The whole review breaks into seven steps:

1. Role anchoring: it asks whether I’m Party A or Party B, then loads all my company details and payment information into context so agreements can be auto-filled
2. Context and concerns: I hand it meeting notes or a full transcript of my conversations with the client, plus “what I’m most worried about”—for a first-time engagement, that’s usually getting paid and protecting commercial reputation
3. Scan for blanks and typos: surface unfilled fields, numbering errors, and typos
4. Clause-by-clause review: every recommendation links back to the source text
5. Consolidated report: the whole thing rewritten in plain language, for me
6. Version numbering: which version, negotiated on which date, with every change tracked
7. Generate the negotiation email: subject, recipients, cc, attachments, Chinese–English side by side, the clauses I want changed and why—copy it and send

4. Actually running a 30-page cross-border contract

This was a bilingual agreement I signed with a major US brand for one of my projects.

The agent’s read: this is a template issued by US headquarters, clearly slanted toward the client, and signing it unchanged carries substantial risk.

The web page it produced told me five things I needed to decide on:

• Payment terms: it digs up standard industry practice under comparable conditions as a reference. It gives you both the terms to aim for and the ones you shouldn’t accept
• Copyright ownership: the original contract demanded “permanent assignment of all rights in all materials”—meaning I couldn’t even show the work in my portfolio later. Its suggestion was to change this to “low-resolution versions may be displayed in portfolios, on the company website, and in case studies, with no resale or sublicensing.” The reasoning: Fortune 500 companies (P&G, Unilever) and 4A agencies all keep this carve-out too—this is standard industry practice
• Termination clauses
• Liability and indemnity: the original contract had no cap—a ¥500,000 project where a likeness-rights claim comes in at ¥10 million could leave you paying far more than ¥10 million. The recommendation is a cap set at “the contract value, or the payments received in the 12 months preceding the incident,” also standard international practice
• Insurance currency: US dollars or renminbi

Then there were highlighted items worth fighting for—for instance, “services guaranteed free of any defect or error.” The creative industry can’t deliver that, especially when AI-generated ad assets can come with small logo artifacts or e-commerce images at resolutions too low to read. You need a defined remediation window.

5. The efficiency comparison

This used to take me several days of back-and-forth.

Now: drop the contract in → 15 minutes for the AI to finish the review → 5-10 minutes for me to review its recommendations → copy the email and send it.

The second email, covering detailed revisions, is already drafted too—send the first one to test which points the counterparty accepts in principle, then send the second for the detail work.

The lawyer is happy too. What I forward is professional, and I’ve already worked out what I can and can’t accept—which makes the whole exchange easy and genuinely pleasant.

Wrapping up

AI doesn’t replace lawyers. It front-loads the things I can’t do—so I have the standing to act like the client instead of a legal novice.

Small contracts: review and decide myself. Large contracts: lawyer does a second pass. AI has lifted a one-person company’s legal capability to within striking distance of a professional team’s.

Source: EP0006_audio.mp3 · ASR model gemini-2.5-pro · full text of the 12-minute original recording

[00:00] Friends, have you ever thought about this question? A lot of people using AI, the first question they think of is which AI tool to use. But what you should really spend time thinking about is—what to use AI tools for. I’ve roughly sorted it out: the things I use AI tools for day to day fall into a few buckets. The first bucket is the high-frequency, repetitive stuff in daily life. It may not be very hard work, but I need AI to do it in my place. The second bucket is the things I frequently pay to outsource to other people—for instance, as a one-person company I’m probably not a professional when it comes to finance and legal, so I need to find lawyers and finance people to help me solve those problems. Those are things I can’t do and have to pay someone else for, and those are exactly the things AI can help with. And there’s another bucket, the high-value things—things that maybe only I plus AI could pull off at all.

[00:48] So what we’re sharing with you today is how I use AI to do contract review in my day-to-day work. I don’t know if you’ve been through this. Back when I was at a big tech company or at companies with a lot of headcount—reviewing contracts was a genuine pain. First, you’re not a professional on the legal side yourself, so when you get a contract you have to walk the lawyer through the business, and that goes back and forth for many rounds. And our lawyer, sometimes it feels like they aren’t entirely on our side either—it feels like they’re always picking holes in everything. Then when you talk to the other side’s business people, their business people have to go talk to their legal team—so there are several layers of people in between, and different people have different communication habits, right.

[01:32] You rarely meet a lawyer who talks like a human being—or rather, let’s not call it talking like a human being, let’s call it a lawyer who can explain legal clauses in plain language—and who also gives you very quick feedback. So running my own company, a lot of the time the real headache is revising these contracts. So over a stretch of time I polished up an agent of my own, specifically to review contracts for me with AI. Let me share how that was built.

[02:05] The whole thing obviously starts from a situation like this. When you’re running your own company, a lot of the time it’s someone else—a client throws a template at you, or the client says “this kind of business is a first for us too, we don’t have a matching template either.” So at that point what you get may be a template, or we draft a template ourselves. And here, if it’s the client holding the cards, especially a template from a big platform company or a big corporation, it’s usually a contract written entirely in their own favor. There are a lot of clauses in there where, if you don’t revise them and don’t argue them, you as the vendor get a bad deal.

[02:38] So the way I’ve set the whole thing up, I need this agent to do a detailed contract review in my place, find the places in it that might carry real risk, and assess each risk point from my position, which may be the vendor side. And after that, I want it to draft an email to the other side’s legal team to push for those clause changes.

[03:02] So the original process, as you can see—hiring a lawyer, especially a specialist in the field, and lately with everyone going global there are a lot of bilingual contracts—usually runs to tens of thousands of yuan, and every round of feedback takes at least several hours.

[03:16] So based on a need like that, I built out this agent framework. The first thing is that in the whole process, drafting a contract and reviewing a contract are two different jobs, so I split them into two agents. And when I first started thinking about it, I thought I’d go find the up-to-date template library for the industry—but that turns out not to be very necessary. The reason is that most models already have a lot of the relevant legal knowledge in them; we don’t need to build that kind of library for them by hand. And on top of that, each time we do this work we can have the model go find newer knowledge from the right channels and add it into the boilerplate, so it lines up better with how China’s legal system and market regulators work right now.

[04:07] Then it also has to go through all the clauses in the contract and review them one by one. And it has to tell me, for every revision it suggests, where in the original text that links back to. And when it’s a small-value contract—plus I know the business better in those cases—in that situation it’ll lay out some options for me to decide between, along with the risks and possibilities behind each one. But it doesn’t give a final legal opinion. Especially once the amount goes past a certain threshold, I’d still recommend having a lawyer do a final round of review. And that review isn’t a back-and-forth process anymore—this legal agent of ours can lay out every issue, so the lawyer just needs to skim the AI’s write-up quickly and give their professional feedback, which makes it a lot more efficient too.

[05:03] So for the whole process, its review runs in seven steps.

Step one is anchoring the role—even though I’m the vendor most of the time now, there are a few occasions where I sign the agreement as the client. So right at the start I have a conversation with the agent about whether I’m Party A or Party B. It’ll drop all my relevant company info, payment info and so on into context, and it can even fill the agreement out completely on its own.

Second is background and concerns—I usually send the agent the meeting notes or the full transcript of my conversation with the client, and have it think through what the background of this collaboration is, and I’ll add what I’m most worried about. Like, it’s a first-time collaboration and I’m worried about them paying, or I’m worried about their commercial reputation—it’ll fold everything I care about most into the work.

Step three scans the whole contract for fields that weren’t filled in and fills them, or finds typos and numbering errors in there, to tighten the whole thing up.

Right after that it does the clause-by-clause review I just mentioned, pulls together a report, and turns that report into something I can actually read. And it also gives the contract a version number—which version from which date’s discussion—and logs what we changed each time.

[06:22] And here I can show you an actual case from back then. All I have to do is throw the Word document of the agreement at this agent, and in about 15 minutes it gives me a review report.

OK, this is a bilingual agreement where we’re the vendor, an agency, signing with a big American brand. Looking at the agreement as a whole, the agent’s read is that it’s a template handed down from the US headquarters, clearly slanted toward the client, and signing it unchanged would carry a lot of risk. The first thing is that this is an annual framework agreement, not a single order. And then maybe the nastiest part is that some places in it still have earlier tracked changes left in, which is a real problem, and it hasn’t been filled out as a final version.

[07:10] The web page the agent outputs tells me there are five things I need to decide.

The first is the payment cycle, the payment cycle and the late-payment penalty—it’ll dig up what the industry does under the same conditions so I’ve got something to go on, meaning how the industry usually sets this clause, what I’d want the clause to be in that case, and what I can’t accept.

The second point is the copyright question in here—whether the images we shoot can be shown in my portfolio. That’s also something I need to decide.

The third is the termination clause.

The fourth is the liability for payouts if something goes wrong.

The fifth is about insurance—basically, if a dispute in here escalates to an insurance payout, whether that’s on a US dollar basis or a renminbi basis. Those are the key questions.

[08:14] Next it spells out in detail what each of those points I mentioned actually means. Take the second one: if we don’t change it, then all rights to all these materials are assigned permanently. And later, if I want to take on clients from other brands, I couldn’t even show it as part of my portfolio. And actually the big ones—including Fortune 500 brands, P&G, Unilever, that tier—when they sign with 4A agencies, they do reserve portfolio and case-study rights for the creative shop. It isn’t because 4A agencies are so powerful; it’s just standard industry practice. So it suggests I revise it to: I can show a low-resolution version on my portfolio site and in case studies, I can’t resell or sublicense it to anyone else, but I do get to keep an internal archive copy. So that’s the point it suggests I adjust.

[09:12] Or take the cap on liability, right. If we don’t change it, a 500,000-yuan shoot project—the total amount is that 500,000—and it gets sued for 10 million over portrait rights, then under the contract as written my payout could be even more than 10 million. So it suggests capping it at the contract amount, or even at the payments received in the 12 months before the incident—which is also standard international practice, not something we’re changing specially for our own interests. And it explains every single one of these like that.

[09:45] Then there are the ones highlighted in yellow, the ones worth pushing for. Like this warranty of zero defects, zero errors—that just isn’t achievable in the creative industry, especially when we’re generating ad assets with AI, where there’s a good chance a logo has small flaws that a person has to adjust by hand. Or because e-commerce images run at lower resolution, there are spots that wouldn’t even be clear in a real photo. Those are things we head off first with multiple layers of review, but you can never fully guarantee zero defects. So agreeing on a window for fixes, and a clause putting some review responsibility on the other side too, is also worth pushing for.

[10:27] Further down there’s some odds-and-ends cleanup—like some inaccuracies in layout and formatting, including some blank pages, right, and filling those blanks in is all part of what it suggests. And it’ll also—this agent will also ask me for some information before it sends the email. Some of that information it can read straight out of my database, of course, but it also lays out a negotiation rhythm for me—like send the first email to feel out how willing the other side is, see what they accept in principle and what they don’t. And after that exchange, we can go do a detailed round of revisions.

[11:09] And it’s already drafted both emails. Take the second email, the detailed revisions—what its subject line is, then who it goes to, who’s cc’d, what goes in the attachment, including the full text inside, the bilingual version—which clause we want changed how, what the reason for the change is, it lays it all out. And then I can just copy that email and send it out. Very convenient.

[11:34] When I used to do this, I’d need days of back-and-forth plus reading the contract myself. But now, getting all of this done—the first full pass and the recommendations, including sending my email—is about 15 minutes of AI work, plus maybe 5 to 10 minutes of me reviewing whether these really are the issues. This is the kind of thing I’m not good at and that comes up often—maybe not that often, but it’s the kind of thing I have to pay an outside expert to do.

[12:06] And now I get it all done with AI agents helping out. Small contracts I can read and revise myself; the big, important agreements all go to a legal team for a second pass. The lawyers are happy too—the feedback I give them is professional, and I hand over my thinking up front, what I can and can’t accept, so the whole process is easy and pleasant.

[12:28] So if there’s anything you’ve thought about maybe using AI for in your daily life, or you have questions about AI, drop a comment any time. And I’ll go a step deeper for you in the daily updates. Thanks everyone.