I Used Claude to Draft a Singapore SHA. Here's What It Got Wrong.

I've been experimenting with AI tools for legal drafting for the past year. Not as a replacement for legal work, but as a way to accelerate the parts of the process that don't require judgment — first drafts, clause research, consistency checking.

Last month, I decided to test Claude on something more ambitious: drafting a complete shareholders' agreement for a Singapore-incorporated investment SPV. Not a template. A bespoke SHA for a specific (anonymised) deal structure.

The results were instructive.

The Setup

The deal was relatively straightforward: three investors co-investing through a Singapore SPV into a Series A round in an Indonesian fintech company. Total commitment of USD 2 million. One lead investor with a board seat, two passive investors.

I gave Claude a detailed prompt: the commercial terms, the parties, the capital structure, the governance requirements, and the key protective provisions the lead investor needed.

What I got back was a 28-page document that looked, at first glance, remarkably competent.

What It Got Right

Credit where it's due. Claude produced a structurally sound document with:

A coherent and well-organised table of contents
Proper defined terms used consistently throughout
Reasonable boilerplate provisions that were correctly adapted for Singapore law
A sensible approach to information rights and reporting obligations
Clean language that was, in places, better than some human-drafted SHAs I've reviewed

The formatting was professional. The internal cross-references were mostly correct. The general architecture of the agreement was sound.

What It Got Wrong

The problems were subtle, which made them more dangerous than obvious errors would have been.

1. The anti-dilution provisions were mechanically correct but commercially wrong. Claude drafted a broad-based weighted average anti-dilution clause. Technically fine. But the commercial context called for a narrow-based formula because of how the option pool was structured. This is the kind of judgment call that requires understanding the deal, not just the law.

2. The drag-along threshold was too low. Claude set the drag-along at a simple majority (51%). For a three-investor SPV where one investor holds 45%, this effectively gives the lead investor unilateral drag-along rights. A competent lawyer would have flagged this and suggested a higher threshold or additional conditions.

3. Singapore-specific requirements were incomplete. The document referenced the Companies Act but missed several practical requirements:

No mention of the requirement for directors to disclose interests under Section 156
The share transfer provisions didn't account for the right of first refusal mechanics that ACRA expects to see
The deadlock resolution mechanism defaulted to arbitration in London, which is unusual for a Singapore-incorporated company with Southeast Asian investors

The most dangerous errors in AI-drafted legal documents are the ones that look correct to a non-specialist reviewer.

4. Tax provisions were generic. The withholding tax and distribution provisions didn't account for the specific tax treaty positions of the investors' home jurisdictions. This isn't something you can template — it requires analysis of the actual treaty network.

The Real Lesson

The issue isn't that Claude can't draft legal documents. It can, and it's getting better. The issue is that the gap between "structurally correct" and "commercially appropriate" is where legal judgment lives.

A good shareholders' agreement isn't just a collection of correctly drafted clauses. It's a document that reflects the commercial bargain between the parties, anticipates the specific risks of the deal, and provides workable mechanisms for the situations that actually arise in practice.

AI tools are excellent at the first part. They're unreliable at the second. And they're currently incapable of the third without significant human input.

How I Use AI Now

My current workflow uses AI as an accelerant, not a replacement:

I draft the term sheet and key commercial terms manually
I use Claude to produce a first draft based on those terms
I review and revise the draft with the same rigour I'd apply to any junior lawyer's work
I use AI again for consistency checking and cross-reference validation

This cuts my drafting time by roughly 40% while maintaining the quality standard my clients expect. That's a meaningful efficiency gain, but it's not the revolution some people are selling.

The lawyers who will thrive are the ones who learn to use these tools effectively while maintaining the judgment and expertise that makes legal advice valuable. The ones who will struggle are those who either ignore the technology entirely or trust it too much.