What should organisations consider about retrieval-augmented generation?

Imagine using an AI tool to draft a report at work, with substantiated and verified research behind it. A market brief. A policy note. A client update. What could go wrong?

Plenty of teams across APAC already use copilots this way, and for good reason. Drafting moves faster. Internal documents get summarised. A first pass that used to take an afternoon can land in minutes. The productivity case is real.

The mechanism behind much of that workflow is retrieval-augmented generation, or RAG. In plain terms, RAG means the model searches a knowledge base or the web, pulls in relevant passages, and uses them to shape its answer. Enterprise tools often present the retrieved files as footnotes or a source panel. That is supposed to ground the output in evidence rather than invention.

However, it can still produce work that looks verified when it is not. A few scenarios are worth keeping in mind:

• The citation links to a real PDF, but the sentence beside it overstates what the page says.

• The answer is drafted first and sources are attached afterwards to fit, a pattern sometimes called post-hoc citation.

• The retrieved material is real but wrong for your market, such as a US dataset cited in an Asia-Pacific strategy note, or an English-only source applied to a multilingual audience without adjustment.

• The web pages the model can reach are not the same as the web a human researcher would trust, because many credible publishers block AI crawlers.

What happened

In a typical RAG workflow, the system embeds your question, searches connected documents or indexed web pages, retrieves the closest matches and generates a response that weaves them together. On paper, every claim should trace back to something on file.

In practice, retrieval finds related text, not necessarily supporting text. The model still summarises, compresses and connects ideas. It can attach a genuine regulator circular, vendor white paper or news clip and write a line that goes further than any paragraph in the source. The footnote looks like diligence though there may be a gap in the synthesis.

Generative search tools follow a similar logic when they browse the open web. Researchers audited ChatGPT, Copilot, Gemini and Perplexity on 712 real-world queries in politics, health and the environment. Roughly 16% of cited sources showed evidence of AI-generated content, about one in six footnotes pointing at synthetic web material rather than human-authored pages.

That connects to a broader shift in what models can actually read. Cloudflare reported in 2026 that automated agents generated more than 57% of web traffic, ahead of humans at about 42%. AI systems are scanning and requesting pages at scale as part of how they source information. Many high-credibility publishers, newsrooms and research sites restrict AI crawlers in their robots.txt files, partly because bot traffic has grown so quickly, partly to protect subscription and licensing models.

What it means

A browsing copilot is not necessarily reading the full credible web. It reads the crawl-accessible slice: pages that still allow automated access. What comes back can be thinner, older or machine-generated, even when the citation panel looks authoritative. Teams may send work outward believing the footnotes did the checking for them.

Across APAC, teams from Tokyo to Jakarta can receive fluent drafts with sources attached while missing local regulator releases, statistics or language-specific reporting that never entered the index. In financial services, public policy, education, healthcare and corporate communications, the risk is often misplaced confidence: citations present, paragraphs unread. A source may be real but stale, or useful globally but wrong for your market.

The check is not whether a link exists. It is whether the paragraph supports the claim for the audience you are writing for.

In writing this article, we found that Singapore University of Social Sciences advises students that ideas taken from generative AI tools must be appropriately attributed, and that they should review and verify generated content using an evaluation framework such as CRAAP (currency, relevance, authority, accuracy and purpose). Assessment submissions that use AI are expected to include discipline-specific citation and a record of the tool, prompts and output used.

Some organisations are starting to govern the output, not only the tool. In August 2026, US go-to-market platform Clay made an internal AI writing policy company-wide after engineering teams wrote it for their own use and other departments asked to adopt it. The policy is not a ban on AI drafting. It requires people to stand behind every sentence they send, labels raw model output as such, and treats "AI wrote that, just ignore it" as an unacceptable answer in review.

With this example in mind, APAC firms should consider deciding what "verified" means when a copilot attached sources automatically. Minimum useful steps: spot-check retrieved paragraphs before external send; require named ownership of claims in client-facing drafts; and treat citation panels as a starting point for review, not a substitute for reading the source.

While the habit remains boring, it is perhaps the most effective: open the page, find the paragraph, and verify whether it says what the draft from AI claims. As with Clay's policy, "AI wrote that, just ignore it" should not be an acceptable answer in review.

What to watch

• RAG literacy: Whether workplaces, universities and schools teach what retrieval-augmented generation actually does, at a level managers, teachers and students can use.

• AI writing policies: Whether companies and organisations in APAC and beyond add AI citation guidelines, fact-checking steps and line-level accountability to their AI use policies, not only data-handling rules and tool bans.

• Crawler blocking: Whether more credible websites and online resources restrict AI access, and what that means for the quality of AI-generated responses over time. If high-trust publishers stay outside the crawl path, copilots may rely more heavily on whatever remains indexable, including synthetic or SEO-oriented pages.