Latest / Future of Work Tech with Fexingo: Remote Tools, AI Productivity, and Workplace Software / How AI Is Automating Your Internal Document Generation
Transcript
- Lucas: So you know how every company has that one document — the employment contract, the non-disclosure agreement, the vendor MSA — that everyone hates writing but everyone needs, over and over again? Luna: Oh, absolutely. At my last company, we had a 47-page master service agreement template that our legal team basically copy-pasted and then spent two weeks tweaking every single time. Lucas: Right. And that's the kind of work that feels necessary but is actually incredibly inefficient. So today I want to talk about a specific use case for AI that's quietly saving companies thousands of hours a year: internal document generation. Luna: Not external client-facing documents, but the ones we use to run the business internally? Lucas: Exactly. I'm talking about employment contracts, offer letters, compliance memos, vendor agreements, board meeting minutes — the stuff that every employee in legal, HR, and finance touches multiple times a week. And I want to anchor this in a real example: a mid-size SaaS company called CloudBridge. Luna: I've heard of them. About 800 employees, based out of Austin. They do cloud infrastructure monitoring, right? Lucas: That's the one. So last quarter, CloudBridge deployed a custom AI document generation pipeline. They took their entire template library — hundreds of documents — and fine-tuned a large language model on them. The goal was to let anyone in the company generate a first draft of any internal document just by filling out a short form. Luna: And how did it actually work in practice? Because these templates usually have a lot of conditional logic — if the employee is in California, you need this clause, if they're in the EU, you need GDPR language. Lucas: That's exactly the challenge they solved. Instead of a simple 'fill in the blank' approach, they built a structured input system. The user selects the document type, enters key variables like jurisdiction, role, start date, and the model generates a complete draft with all the right clauses inserted. The legal team then reviews, edits if needed, and it's done. Luna: So the AI is handling the boilerplate, but the lawyer is still in the loop as the final reviewer. That's a model I've seen work well in other contexts. Lucas: Yes. And the numbers here are pretty striking. Before the AI pipeline, CloudBridge's legal and HR teams were spending an estimated 1,200 hours per year just on drafting these documents. That's basically one full-time employee's entire working year. Luna: One thousand two hundred hours — that's a lot of copy-pasting. Lucas: It is. After deployment, the contract turnaround time dropped from five days to about ninety minutes. And the legal team shifted from being the primary drafters to being reviewers. They now spend their time on the complex cases and the edge cases, not on formatting. Luna: But I've got to ask — what about the risk? If the AI generates a clause that's inaccurate or hallucinates something that creates liability, the company is still on the hook. Lucas: That's exactly the problem CloudBridge hit in their first month. One of their generated contracts included a merger clause that didn't exist in any of their templates. It was completely fabricated — looked plausible, but was legally meaningless. And it got caught during the human review step, which is why they built that step in. Luna: So the human-in-the-loop isn't just a nice to have; it's a legal necessity. Lucas: Exactly. But here's the thing: the rate of hallucinated clauses was less than 1 percent after the fine-tuning. And the team at CloudBridge told me that the errors they saw were almost always in clauses that were rarely used. The common stuff — employment terms, standard indemnification — was rock solid. Luna: That makes sense. The model has seen the common language thousands of times, so it's confident. The rare edge cases are where you need a human. Lucas: Right. And CloudBridge also ran into a version control issue. They have multiple versions of some templates — one for pre ipo employees, one for post ipo, one for contractors. The AI sometimes mixed them up. Luna: That's what I'd call 'template drift.' The model doesn't know that Version 3.2 is the current one unless you explicitly tell it. Lucas: Exactly. So they added a metadata tag to each template — effective date, version number, approval status — and then fine-tuned the model to respect those tags. After that, the accuracy rate for version selection went up to 99.5 percent. Luna: So the fix was more structured data going in, not just more training data. Lucas: Precisely. And I think that's the key lesson for any company considering this: the AI is only as good as the structure you give it. If you just throw a bunch of templates at a model and say 'generate a contract,' you'll get chaos. Luna: So what's the broader trend here? Are we going to see every company with a legal department building their own document generation pipeline in the next year? Lucas: I think we'll see a lot of pilots, but the ones that succeed will be the ones that invest in the upfront work — building the template library, tagging everything, setting up the review workflow. The technology itself is almost commodity at this point. The differentiation is in the process design. Luna: And that process design is really about change management, too. You've got to get your legal team to trust the AI enough to review instead of redraft from scratch. Lucas: Yeah, that's a cultural shift. CloudBridge actually did a two-week pilot where the legal team was asked to review ai generated contracts alongside their normal workflow, and they tracked the time savings. Once the lawyers saw that they were saving three to four hours per contract, the resistance melted away. Luna: That's smart — let the data speak for itself. Lucas: So what's next? I think we're going to see document generation move from templates to truly dynamic documents that update themselves based on regulatory changes. Imagine an AI that reads new SEC rules on Monday and updates your compliance memo template by Tuesday. Luna: That would be a game-changer for public companies. Right now, every time a new regulation drops, legal teams scramble to update all their documents. Lucas: We're probably a year or two away from that being reliable, but the pieces are coming together. The same pipeline that generates the document can also scan new regulations and flag which templates need updating. Luna: Honestly, if today's conversation gave you something usable — a concrete approach to internal document generation, or a cautionary tale about template drift — that's exactly the kind of insight we try to bring every week. Lucas: Yeah, and if it saved you even a couple of hours of frustration, that's the feeling we're after. If it was worth a coffee to you, the link is buy me a coffee dot com slash fexingo. Luna: And that small gesture really does help keep this show ad-free and focused on the actual work — not on sponsored segments. Lucas: Exactly. So back to the future of document generation: I'm also watching how companies are starting to use AI to generate not just text documents, but also the data contracts that sit behind them — the structured agreements that define how data flows between systems. Luna: So like, the MSA for your API integration? Lucas: Exactly. And that's a whole new frontier. Because a data contract isn't just legal language — it's also a specification that your engineering team needs to implement. If the AI can generate both the human-readable contract and the machine-readable schema in one pass, you've just eliminated another handoff. Luna: That's the kind of integration that really starts to pay off. Not just automating a single task, but connecting two workflows that were previously siloed. Lucas: Right. And that's the longer-term vision. For now, the low-hanging fruit is still the internal document generation we talked about today. If your legal or HR team is spending more than a few hours a week on drafting, there's almost certainly a case to build a custom pipeline. Luna: And start with a pilot, track the time savings, and get your team on board with the data. Lucas: Exactly. That's the playbook from CloudBridge, and I think it's going to become the standard approach over the next 12 to 18 months. Luna: Well, I'll be keeping an eye on my own company's template library after this.