Wellington council's Deloitte report was mostly written by AI

A massive chunk of the Deloitte report for the Wellington City Council was actually generated by an LLM. This isn't just a rumor or a small tweak to a few sentences; the Mayor has explicitly pointed out that significant sections of the document were produced using artificial intelligence. It’s a perfect real-world example of how the lines between human consultancy and automated content generation are blurring faster than most organizations can regulate.

We often talk about prompt engineering and AI workflows in the context of boosting productivity, but this situation highlights a major transparency crisis in professional services. When a firm like Deloitte delivers a high-stakes report to a government body, the expectation is a level of human oversight and analytical rigor that an LLM, no matter how advanced, cannot guarantee on its own. If a "large chunk" is AI-generated, it raises immediate questions about the verification process. Did the consultants actually fact-check the hallucinations? Or did they simply use the AI to draft the prose and then slap their logo on it?

This is a cautionary tale for anyone looking at the deployment of AI in professional environments. While using an LLM agent to summarize data or structure a report is efficient, the "human-in-the-loop" part of the workflow seems to have failed here. In a professional setting, the value isn't just in the text; it's in the accountability behind that text.

The risks of unverified AI deployment

When we look at this through a technical lens, we can identify several points where the workflow likely broke down:

• Lack of attribution: There was no clear indication in the document that AI was used for the primary drafting of key sections.
• The Hallucination Gap: If the AI was tasked with synthesizing complex council data, the risk of it creating plausible-sounding but factually incorrect conclusions is extremely high.
• Superficial Review: It seems the "review" phase was likely a surface-level check for grammar rather than a deep dive into the logic and data integrity of the generated content.

For those of us working on building better AI workflows, this is a reminder that "AI-assisted" shouldn't be a mask for "AI-generated without oversight." If you are implementing these tools in a corporate or governmental capacity, you need a strict protocol for how much of the output is automated and how every single claim is validated.

If you're building an AI-driven consultancy or even just using LLMs to speed up your own documentation, you have to treat the output as a raw draft that requires a complete, ground-up audit. Using Claude Code or other advanced agents to assist in coding or drafting is one thing, but when the output becomes the final product without heavy human intervention, you aren't just using a tool—you're outsourcing your professional credibility to a probability engine.