Key Takeaways
- Most AI errors that reach clients happen at firms that already have an AI policy. The gap is verification, not rules.
- Tier every AI-assisted output by the consequence of an error, then match the depth of checking to the tier.
- Verify by claim type: citations, numbers, quotations and regulatory statements each need a different check.
- Every client-facing output needs a named human signatory and a short evidence trail.
- Report an error escape rate to the board, not the number of AI licences.
In April 2026, one of Wall Street's most respected law firms wrote to a federal bankruptcy judge to apologise. An emergency motion it had filed contained some 42 inaccuracies generated by AI, and the errors were found not by its own reviewers but by opposing counsel.
The firm had AI policies and training. As its letter admitted, those policies "were not followed". That sentence is the real lesson for enterprise leaders in Hong Kong: an AI usage policy tells people what they may do; a verification policy decides what is allowed to leave the building.
Why do AI errors reach clients even when a firm has an AI policy?
Because most AI policies govern access, not output. They say which tools staff may use and which data they may enter, but rarely define who must check an AI-assisted deliverable, how deeply, and against what evidence. Under deadline pressure, review becomes a skim, and fluent text passes as verified text.
The pattern is now well documented. According to AFP reporting in June 2026, researcher Damien Charlotin's crowd-sourced database had logged about 1,600 court filings containing AI hallucinations across 35 countries, with cases up eightfold on the previous year. Law is simply the sector where errors become public. The same failure happens quietly in board papers, credit memos and client reports.
Professional services firms are not immune. In October 2025 Deloitte agreed to partially refund the Australian government for a A$440,000 report that contained non-existent references and a fabricated court quote, as reported by Accounting Times. The firm said its findings stood, but the reputational cost had already landed.
The deeper problem is structural. Generative AI produces text that looks finished. A manager who would question a junior analyst's rough draft tends to trust a polished paragraph. Verification has to be designed into the workflow, because instinct will not supply it.
What should an AI output verification policy include?
A workable policy has five parts: a consequence tier for every output type, a verification method for each kind of claim, a named human signatory, a lightweight evidence trail, and a quality metric reported upward. Together they turn "someone should check it" into a repeatable control that auditors and clients can inspect.
The five-part verification framework
- Tier: classify AI-assisted outputs by what happens if they are wrong.
- Method: define how each claim type is checked, and against which source of truth.
- Signatory: name the person accountable for release, not the person who ran the prompt.
- Trail: record the tool, the sources checked and the reviewer, in under two minutes.
- Metric: track how many AI errors escape to clients, and report the trend.
None of this requires new technology. It requires the discipline firms already apply to financial controls: segregation of duties, evidence of review, and a number that tells management whether the control works.
How do you decide how much checking an AI output needs?
Match the depth of review to the consequence of an error. Internal brainstorming needs almost none; a client deliverable needs source-by-source checks; a regulatory filing or court document needs an independent second review. Tiering prevents two failures at once: unchecked high-risk work and wasteful over-checking of low-risk drafts.
| Tier | Example outputs | Minimum verification |
|---|---|---|
| 1. Internal draft | Meeting summaries, brainstorm notes, first drafts | Author reads before sharing |
| 2. Internal decision | Board papers, credit memos, management reports | Author traces every number and source |
| 3. Client-facing | Advice letters, proposals, client reports | Author check plus a reviewer who signs off |
| 4. Regulatory or legal | Court filings, regulatory returns, disclosures | Independent second review against primary sources |
The tier is set by the output, not by the tool. A summary produced in a sanctioned enterprise platform becomes Tier 3 the moment it is pasted into a client email. Make that rule explicit, because staff otherwise assume an approved tool means approved output.
Keep the tiers few. Four is enough for most organisations; with more, teams argue about classification instead of checking content.
How should each type of AI claim be verified?
Split the output into claim types and check each against its own source of truth. Citations must be opened and read, numbers recalculated from the underlying data, quotations matched word for word, and regulatory statements checked against the current rule. A general read-through catches tone problems, not fabricated facts.
- Citations and references: open every source and confirm it exists and says what the text claims. Fabricated citations look exactly like real ones.
- Numbers: recalculate from the source system or spreadsheet. AI tools can transpose figures or silently round them.
- Quotations: match word for word against the original. Paraphrase presented as quotation is a common failure.
- Regulatory and legal statements: check against the current version of the rule, since models may describe superseded guidance.
- Names, dates and entities: confirm against your own records, especially client and counterparty details.
A second AI model can help find passages worth reviewing, but it is not verification. Two models can share the same blind spot, and a reviewer who reads "checked by AI" tends to relax. Use tools to prioritise human checking, not to replace it.
Retrieval-based systems that answer only from your approved documents reduce the error rate, and they make checking faster because each claim links to a source. They do not remove the need to look.
What does AI verification look like in Hong Kong industries?
The framework stays the same; the high-risk claims change by sector. Financial services firms focus on figures and regulatory statements, professional services on citations and advice wording, and logistics or property firms on contract terms, dates and counterparty details that flow straight into daily operations.
Financial services
A wealth management team uses AI to draft client portfolio reviews. Under the framework these are Tier 3: every performance figure is recalculated from the portfolio system, and any statement about product risk is checked against approved product documentation before a licensed representative signs off. Institutions supervised by the HKMA or the SFC will also want the evidence trail available to internal audit.
Professional services
An accounting or consulting firm drafts reports with AI assistance. Citations to standards, case law or research are the highest-risk claims, so they get a mandatory open-and-read check. The engagement partner, not the analyst who wrote the prompt, is the named signatory.
Logistics and property management
Operations teams use AI to summarise service contracts and tenancy agreements. A wrong notice period or rent-review date travels directly into scheduling and billing. Here verification centres on dates, amounts and party names, checked against the signed document.
What mistakes make AI verification fail?
The common failures are treating policy training as a control, letting the prompt author approve their own output, checking only the parts that look wrong, and accepting a second AI's confirmation as proof. Each leaves a gap that deadline pressure will eventually find, usually in the most important document of the quarter.
- Training without enforcement: staff complete the AI course, but nothing in the workflow requires a check before release.
- Self-approval: the person who generated the draft is the only reviewer, so their assumptions are never challenged.
- Spot-checking by instinct: reviewers check what looks suspicious, but fabricated content looks normal by design.
- AI checking AI: a second model's "looks correct" is recorded as verification.
- No feedback loop: errors caught internally are fixed silently, so nobody learns which tasks or tools produce them.
The last point matters most for improvement. Every caught error is free data. Log it by task type and tool, and within a quarter you will know where AI helps and where it quietly adds risk.
How should you report AI output quality to your board?
Report three numbers each quarter: verification coverage, meaning the share of Tier 3 and 4 outputs with a recorded check; errors caught before release; and errors that escaped to clients or regulators. Boards need evidence that AI-assisted work is controlled as tightly as any other client deliverable.
Frame it as risk management, not technology. Directors are increasingly alert to AI liability: according to Global Legal Post, Gartner has advised general counsel to consider direct AI insurance cover as insurers add AI exclusions to traditional policies. A falling escape rate is a credible answer when insurers, auditors or clients ask how AI use is governed.
Pair the numbers with time saved. If verification adds 20 minutes to a report that AI drafted in place of four hours of work, say so. The point is not to slow AI down; it is to make the productivity gain safe enough to scale.
Verification is what lets AI scale
The firms hurt by AI errors are rarely the ones without policies. They are the ones whose policies stop at access and never define release. A tiered, claim-by-claim verification policy with a named signatory closes that gap without slowing every task to a crawl.
Start with your Tier 3 and Tier 4 outputs this month. Name the signatories, write the checks, and start counting what you catch. By the next board meeting, you will have a control you can show, not just a policy you can quote.
Empowering AI. Centered on you.
Put verified AI to work in one role first
Start with a free AI staff configuration assessment: we map one workflow, set the review points and agree who signs off before anything reaches a client. We'll walk you through every step, from the first role to measured results.
Frequently Asked Questions
Can a second AI model verify the first model's output?
Not reliably. A second model can flag passages worth checking, but models share blind spots and may even confirm fabricated sources. Treat AI cross-checks as triage that directs human review, never as the verification record.
Does an enterprise AI platform remove hallucination risk?
No. Enterprise platforms strengthen data protection and can ground answers in approved documents, which lowers error rates. They do not guarantee accuracy, so client-facing outputs still need human verification.
Does the PDPO apply to AI-generated statements about individuals?
Yes. Data Protection Principle 2 of Hong Kong's Personal Data (Privacy) Ordinance requires data users to take all practicable steps to keep personal data accurate, and that includes AI-generated statements about clients or staff.
Who should sign off AI-assisted client work?
The person who would sign the deliverable if AI had not been used: normally the partner, manager or licensed representative responsible for the client relationship, not the staff member who wrote the prompt.
Does verification cancel out the time AI saves?
Usually not. Internal drafts need minutes; client and regulatory outputs need a structured check. When checks target the high-risk claim types, the time saved by AI drafting normally still exceeds the review time.