I have signed contracts I should not have signed. Not because the deal was bad, but because I read the first page, felt good about the person on the other side, and skimmed the rest. That skimming has cost me more than any bad project I have ever taken on.
I still do not have a lawyer on retainer. What I have is a process, and most of it now runs through AI.
The expensive clauses are never on page one
MSAs are written by the client's legal team, and they are written to protect the client. That is not evil. It just means the default version of every agreement you get handed is tilted away from you.
The clauses that actually hurt are boring ones. Net 60 payment terms that quietly become net 90 when their AP department gets busy. Termination for convenience that only works in one direction. IP assignment language that hands over everything you make, including the framework you reuse on every project. Unlimited revisions with no definition of done. Indemnification that has you covering their mistakes. Auto-renewal that locks you in for another year if you miss a 30 day window.
None of those are exotic. I have seen every one of them in real MSAs from real companies, and I have signed more than a few without noticing.
The reason is not laziness. Reading a 14 page agreement carefully takes an hour and a half, and you have client work waiting. So you skim. And the skim is what costs you.
What I actually paste into the AI now
I run every MSA and every SOW through Claude or ChatGPT before I sign anything. Not instead of a lawyer. Before one, so that when I do pay for legal time, I am paying for judgment on three specific clauses instead of paying $400 an hour for someone to read.
The setup matters more than the tool. Here is what I do.
First, redact. Replace the client name with Client and my company name with Vendor. Keep the dollar amounts and the dates, because those are the parts that matter. In ChatGPT, turn off model training in settings. In Claude, turn off the option that lets your chats improve the model. That takes 30 seconds and removes the only real confidentiality problem.
Second, work clause by clause, not the whole document at once. Whole document reviews get vague. Clause by clause reviews get specific.
Third, use a saved prompt. Mine lives in a Claude Project called Contract Review, and I have run something like 60 agreements through it. Here it is almost verbatim:
"You are reviewing a contract for a small service business with no in-house legal team. For this clause, give me: a plain English summary in two sentences, who it favors (Client, Vendor, or neutral), the specific risk if I sign it as written, and a counter I can send back in an email. Do not soften anything. If the clause is standard and fine, say standard and move on. Flag anything involving payment timing, termination, IP ownership, indemnification, liability caps, exclusivity, or auto-renewal as high priority."
That last sentence is doing a lot of work. It tells the model what to care about, which is the difference between a useful review and a wall of text about defined terms.
If you want a tool built specifically for this, Spellbook and LegalOn both do AI contract review and both are aimed at exactly this problem. They cost money. A saved prompt costs nothing and gets you most of the way there at the size most of us are operating.
My client contract review checklist, in the order I check it
I keep this list in a Google Doc and paste it into every review. Order matters, because the first four items kill more deals than the rest combined.
- Payment terms and late fees. When do I get paid, what triggers the invoice, and is there a late fee that actually means something? Net 60 with no late fee is a free loan you are giving them.
- Termination for convenience, both directions. If they can walk with 30 days notice, I want the same right. One sided termination is a red flag every time.
- IP ownership and license back. They can own the final deliverable. They cannot own my process, my templates, or my framework. Get a license-back clause or a carve-out for pre-existing materials.
- Liability cap. Is it capped at fees paid, or is it unlimited? Unlimited liability on a $30,000 project is how a small business dies.
- Scope and change orders. What happens when they ask for one more thing? If the answer is not written down, the answer is that you do it for free.
- Revision limits. Reasonable revisions is not a number. Three rounds is a number.
- Indemnification. Make sure it is mutual, or at minimum limited to your own negligence.
- Exclusivity and non-solicit. Are you blocked from working with other companies in their category, and for how long?
- Auto-renewal and notice windows. Put the renewal date and the notice deadline in your calendar the day you sign.
- Governing law and venue. If a dispute means flying across the country, that clause is worth more than the fight.
SOWs are where the real bleeding happens
The MSA gets all the attention and the SOW does the damage.
A weak statement of work says ongoing marketing support and as needed. A strong one says what you deliver, how many, by when, what the client owes you in return, and what counts as done.
Three things I now require in every SOW. An assumptions section that says out loud what has to be true for the timeline to hold. A client responsibilities section that names who on their side is giving approvals, and within how many business days. And an acceptance criteria line, so I do not like it is not the same as it is not finished.
If they push back on those three, that is information. Usually useful information.
Where the AI is wrong, and where I still call a human
The model does not know your state law. It will be confident about things it should not be confident about. It flags standard clauses as risky because it does not know what your business can absorb. And it misses how two clauses interact with each other, which is where the worst problems actually live.
So the AI is a filter, not a decision. I use it to find the five clauses worth arguing about, then I argue about them. If a deal is over a certain size, or the indemnification language is ugly, or equity is involved, I pay a lawyer for one hour to look at those five clauses specifically. That is a $500 conversation instead of a $4,000 review.
The piece that makes this stick
A review process you run once is a habit. A review process tied to a system is infrastructure.
Mine looks like this. Every contract lives in a Drive folder named with the client and the date. A Notion database pulls five fields out of each one: payment terms, termination notice, renewal date, liability cap, and who owns the IP. The renewal date field is the one that has saved me the most money, because auto-renewals are silent and they always show up at the worst possible time.
I walk founders through systems like this all the time. If you want to see how it fits together, there is a comparison of the group format and the one on one format at https://www.mastermindshq.business/mastermindvsl, and I build live versions of these workflows at https://workshop.mastermindshq.business.
That is the whole thing. Redact, run the clauses through a rubric, check the list, fix the SOW, and keep the terms somewhere you will actually see them again.
I have trained more than 90,000 people on systems like this, and the pattern I see over and over is that the founders who survive the messy middle are not the ones who work harder. They are the ones who build the boring infrastructure before they need it.
If this resonates and you want to build this kind of infrastructure in your own business, the Mastermind is where we do the work live. You can learn more at mastermindshq.business.