What to figure out before you price a private ChatGPT for a law firm

A partner team I build with came to me with a request that sounded simple. They wanted a closed ChatGPT for law firms: a chat that runs on the firm's own server or private cloud and only reads that firm's documents. On day one it would redact documents, read them, analyze them and answer questions about them. Later they'd sell add-on agents on top.

What they needed from me was a price tag. Their salesperson was about to start walking into law firms, and he needed a number ready when a lawyer asked what it costs. They didn't want to spend anything on the build until the first firm signed.

That's a fair ask. It's also where these projects go sideways, because the price depends on three answers nobody had written down yet. So before I quoted anything, I asked for them.

Is this a demo or a real first version?

My first question was whether they wanted a demo or a minimum viable product. A demo is something the salesperson opens on a call: throw your document in there and watch it analyze it. A real first version has to hold up inside an actual firm, with real files, every day.

Those are two different builds with two different prices. Their answer was clear. Build the real first version with the first paying firm, tell that firm it takes about three weeks, then use it as the demo for the second firm. Once I knew that, I knew I was pricing something that has to survive real use.

What will the lawyers actually use it for?

My second question was why the firms want it in the first place. The reason a firm buys decides the infrastructure. A firm that wants to ask questions across years of case files needs a different setup than a firm that wants a few documents redacted each week.

The salesperson didn't have that yet. His plan was to go in with the pitch and pick the lawyers' brains about what they need while he was there. So I priced it off what I already knew: the workflows I've seen hold up at firms I've built similar systems for. I've built an agentic system that researches legal documents and a tool that drafts one section of a legal opinion, so I wasn't guessing at what a firm does day to day.

Template work is a different product

Then one of them asked whether the same system could pull up any template a lawyer needs. A firm they knew had decades of templates and wanted them organized and easy to fill.

I split that out right away. A lawyer chatting with their documents wants a dynamic answer, a different one each time depending on the question. Template work is fixed. Same document, same fields, filled in again. That's a job for an agent built to fill templates, priced on its own, and not something to cram into the chat.

They agreed, and added the part that made it obvious. A newer lawyer wants a library of templates because they don't have their own yet. An established firm already has its templates and wants something to fill them out. Two different buyers, two different products. Keeping it out of the base price kept the first quote about the one thing every firm needs on day one.

Who pays for the server?

The last question is the one that surprises people. A private AI has to run somewhere. If the firm hosts it, the firm carries the hardware and the bill. If you host it for them, you're paying every month for the processing power and the electricity to keep it running, whether anyone logs in or not.

So the structure has two parts. A one-time setup fee installs the private system with the basics. A subscription covers firms that want it hosted and want new agents as they ship, with those running costs built in. The premium pieces, like the template agent, get priced on their own, so nobody's surprised when a firm asks for one later.

What this means for your business

Whether you run a law firm, a recruiting firm sitting on years of candidate notes or any service business with files you'd never paste into public ChatGPT, the order is the same. Decide whether you need a demo or a real first version. Name the job it does every day. Decide who runs the server. Then price it.

And once it's live, how good the answers are depends on how the files are split up and labeled before the AI ever reads them. That's the next conversation.