One of the reasons I was so excited to start my internship at General Legal was the chance to finally use the latest AI tools in a professional environment. I’ve spent the last year almost entirely focused on my law school studies.
Law school is an amazing place to learn, but (for many valid reasons) it takes a conservative approach to technology adoption. My legal writing professor told me the program’s entire strategy for teaching incoming first-years to use AI was a single in-class activity in their spring semester. Incoming students will not learn how to use these tools until spring 2027.
Starting at General Legal felt like entering a parallel universe. It completely reset how I think about AI for legal work and the role of humans in the knowledge economy.
The old workflow is now the slow workflow
Two projects I completed last week perfectly illustrate the dramatic changes in workplace efficiency with the advancement of AI. The first project was a data validation task: identify the discrepancies between two sources of inbound client requests.
Leaning on my background in data analytics, I cleaned up the datasets, pulled them into an Excel sheet, and started throwing everything I know about fuzzy matching at the rows to see how the sheets differed.
As I neared the finish line, I opened Slack to tell my teammates the list was almost ready. I was stunned to see that thirty minutes earlier, one of them had already posted the answer. He hadn't built a spreadsheet at all. He’d simply dropped both datasets into Claude and started prompting. My work was careful, methodical, and completely outdated. I spent over an hour producing something a colleague had produced in minutes.
The second project drove the point home. To evaluate how different AI models perform on legal tasks, I used Claude to generate a realistic set of tasks, modeled on the work our attorneys do every day. I felt like I was finally adapting to the reality of work in 2026, until I was ready to actually test the models.
Without pausing to ask whether I could spin up a process to run multiple model versions concurrently, I started manually pasting prompts and files into Claude and ChatGPT, one at a time. Once again, by defaulting to how I would have worked in the recent past, I burned time on repetitive tasks a script, easily generated by AI, could have handled.
| Task | Old (manual) way | AI-native way | Outcome |
|---|---|---|---|
| Data validation: find discrepancies between two client-request datasets | Clean the data, load into Excel, run fuzzy matching by hand | Drop both datasets into Claude and prompt | Over an hour by hand vs. minutes in Claude; the colleague posted the answer 30 minutes before the manual list was ready |
| Model evaluation: test how different AI models handle legal tasks | Paste prompts and files into Claude and ChatGPT one at a time | Generate a script to run multiple models concurrently | Time burned on repetitive work a script could have handled |
| Idea generation: surface new ideas from company projects | Wait for ideas to emerge from meetings and hallway talk | Build a Claude agent that drafts ideas in Notion each morning | Fresh ideas daily, but plausible and hollow without human context |
With access to AI, lack of imagination is often the main bottleneck to accomplishing tasks. Initially, I felt a blow to my professional pride when I was outstaged in a data validation task by an AI model. Rather than letting it affect me, however, I’m now inspired to think critically about how to best accomplish all professional tasks: can a model or an agent do this better than me? If so, let them have it.
Why “AI for legal work” is no longer optional
It would be comforting to think this is a passing trend, but the models keep getting better at exactly the kind of reasoning legal and many other types of work demand. The recent release of Claude Fable 5, Anthropic’s first publicly available Mythos-class model, is the latest marker. Reporting on the launch noted that it excels at software engineering and knowledge work while shipping with hard safety limits in high-risk domains (TechCrunch). While Anthropic withdrew access to Fable under pressure from the US Government, who cited national security concerns, the continual improvement in the models’ capabilities is undeniable (Anthropic).
For lawyers and the founders who hire them, this trajectory has a clear implication: the firms that treat generative AI as core infrastructure will simply move faster, at lower cost, with more consistency. That’s the bet General Legal is built on.
So what is the human actually for? Quality control.
If a model can draft, match, summarize, and analyze, what’s left for us? After two weeks using AI extensively, human quality control is still vital.
One of our engineers put it perfectly: "I hardly write code now, even after 20 years. I spend most of my time reviewing." That’s not a story about decline, but about where the value moved. The scarce, valuable skill is not producing the first draft, but knowing whether the draft is right.
And being right matters enormously in law, because the models still miss things they weren’t explicitly told to look for. In my own review work, Claude breezed past a section-numbering error: it let 1.2 follow 1.3. At one point labeled a new section "Exhibit A" immediately after "Exhibit 1." Neither mistake was catastrophic, but serving clients slop while pitching yourself as an “AI native law firm” is a sure way to erode trust. A model optimizes for the instructions it’s given. A lawyer is accountable for everything on the page, including the problems no one thought to flag. The model handles volume and velocity and the human owns correctness and consequences.
Human innovation: bolstered not replaced
I tested this idea one more way. I built a Claude agent that generates new ideas in Notion based on our open projects and other company resources. It’s exhilarating to wake up to a fresh batch of ideas in my database every morning.
While exhilarating to see in action, the ideas aren’t especially compelling. They’re plausible, well-formed, and oddly hollow—because they lack the context that comes from actually talking to colleagues, clients, and people wrestling with real challenges. The agent has the company’s documents. It can’t have the hallway conversation, the frustrated client call, the offhand comment that sparks a genuinely good idea. Human innovation is safe for now, but can be greatly bolstered with the power of AI models.
The future of legal work isn’t a world where lawyers are replaced. It’s a world where the lawyers and firms who tactfully embrace AI will be rewarded. Those who reach for the agent before the spreadsheet, and who treat their own judgment as the quality bar, will pull decisively ahead of those who don’t.
Working with us
If you’re a founder or operator looking for legal counsel built for the speed of the AI era, General Legal delivers high-quality, predictable work backed by experienced attorneys focused on your business goals. Ready to see the difference an AI-native approach makes? Sign up here to get started.
Sources
Jessica Rosberger, AI in Law: The Real Risks Beyond Hallucinated Cases, Cornell J.L. & Pub. Pol'y (Apr. 9, 2025), https://publications.lawschool.cornell.edu/jlpp/2025/04/09/ai-in-law-the-real-risks-beyond-hallucinated-cases/.
Anthropic, Claude Fable 5, Anthropic (June 9, 2026), https://www.anthropic.com/news/claude-fable-5-mythos-5.
Jessica Rosberger, TechCrunch: Anthropic’s Claude Fable 5 is a version of Mythos the public can access today, TechCrunch (June 9, 2026), https://techcrunch.com/2026/06/09/anthropics-claude-fable-5-is-a-version-of-mythos-the-public-can-access-today/.
- A summer intern from a law school that barely teaches AI describes arriving at an AI-native firm as entering a parallel universe.
- Two tasks made the shift concrete: a data-validation job the intern did by hand in over an hour, a colleague did in minutes by dropping the datasets into Claude; and a model-evaluation task the intern ran manually instead of scripting it.
- The real bottleneck with AI is not capability but imagination, the habit of reaching for the spreadsheet before the agent.
- As models keep improving, the human role shifts from producing the first draft to quality control: knowing whether the draft is right.
- AI still misses what it was not told to look for (it let section 1.2 follow 1.3, and labeled a section "Exhibit A" right after "Exhibit 1"), so the lawyer stays accountable for everything on the page.
- AI can bolster human innovation but not replace it: an idea-generating agent produced plausible but hollow ideas, because it cannot have the hallway conversation or the frustrated client call.
| What this is | An intern's field notes on moving from law school to an AI-native firm. |
|---|---|
| Why it matters | AI now does routine production faster than careful manual work, so the old workflow is the slow one. |
| The wake-up call | A task done by hand in over an hour took a colleague minutes in Claude; a manual test run should have been a script. |
| The real bottleneck | Imagination, not capability. Reach for the agent before the spreadsheet. |
| What the human is for | Quality control and accountability. AI misses what it was not told to check. |
| On innovation | AI bolsters human ideas but cannot replace them. It cannot have the hallway conversation. |
| Bottom line | Lawyers are not replaced. Those who treat their judgment as the quality bar pull decisively ahead. |
What is an agentic office?
An agentic office is a workplace where AI agents handle routine production (matching, drafting, summarizing, analysis) while people direct the work, review the output, and own the result. The article's title refers to returning to work and finding that this is now the default way tasks get done.
What is an AI agent?
An AI agent is software that carries out multi-step tasks on its own toward a goal, rather than answering one prompt at a time. In the post, the author builds a Claude agent that reads the company's open projects and generates fresh ideas in Notion each morning.
What is fuzzy matching?
Fuzzy matching is a technique for linking records that are similar but not identical (for example, "Acme Inc." and "Acme, Inc."). The author used it in Excel to compare two datasets, a task a colleague completed far faster by prompting Claude directly.
If AI can draft, match, summarize, and analyze, what is the human lawyer for?
Quality control and accountability. As one engineer put it, after 20 years he now spends most of his time reviewing rather than writing. The scarce skill is not producing the first draft but knowing whether it is right, and owning the problems no one thought to flag.