AI wrong information about a financial product: regulated risk

When AI gives wrong information about a financial product, insurance policy, medicine or property, the risk is higher than a normal brand error because buyers act on it with money or health at stake. You cannot edit the AI answer, but regulators expect your own claims, pages and chatbots to be accurate whatever tool produced them. The response is governance: an approved fact register, severity triage, compliance sign-off on fixes, and regular re-testing.

This guide covers what the evidence shows, what regulators have said, and a governance model that fits regulated marketing. It is general information, not legal advice; involve your compliance and legal teams.

01How often do AI tools get regulated facts wrong?

Often enough to matter. In September 2025 the UK consumer group Which? put 40 questions on finance, legal, health and consumer rights to six AI tools. Scores ranged from 55% for Meta AI to 71% for Perplexity.^[1] When a question contained a deliberately wrong ISA allowance of £25,000 instead of £20,000, ChatGPT and Copilot failed to correct it.^[1]

Which? also found ChatGPT wrongly said travel insurance was mandatory for UK residents visiting Schengen states.^[1] In health, a January 2026 Guardian investigation reported misleading AI Overview summaries on topics including liver blood test ranges, and Google then removed AI Overviews for some of those queries while saying the information was supported by high-quality sites.^[2][3]

Buyers trust these answers. Which? reported that 47% of UK adults who use AI for web searches trust the results to a great or reasonable extent, rising to 65% among frequent users.^[1]

02Why is wrong AI information riskier in regulated industries?

Because the facts AI gets wrong are the ones regulators care most about. Rates, fees, cover limits, eligibility, side effects, permits and handover dates are both what buyers ask AI and what rules govern.

03What have regulators said about AI and marketing claims?

No regulator we found makes a brand liable for what a third-party AI engine independently says. But they are consistent on two points: existing rules apply to content produced with AI, and firms answer for their own claims and channels.

Courts have applied the same logic to company-run chatbots. In Moffatt v. Air Canada (February 2024), a British Columbia tribunal held the airline responsible for a refund policy its website chatbot invented, rejecting the argument that the chatbot was a separate entity.^[8] If AI sits on your website or app, its answers are your answers.

04Where do wrong AI answers about regulated products come from?

Trace before you fix. In regulated sectors, the cause is often an old document the firm itself published: a superseded rate sheet, a past policy wording PDF, a product page for a closed product, or a press release with last year's figures.

Other common sources include comparison sites and aggregators with stale data, broker or agent pages that paraphrase terms loosely, community threads, and same-name companies. Our correction workflow covers tracing in detail; in regulated sectors, record every trace step, because compliance may need to show what you knew and when.

05How do we make our own sources harder to misread?

Much of the fix is publishing discipline. State regulated facts in plain text on a canonical page, with the effective date and the conditions next to the figure, rather than only in a PDF, an image or a table without context. A summary model can easily lift "from 4.9%" without the qualifying sentence that follows it.

Keep one current version of each document. When rates or policy wordings change, update the canonical page, redirect or remove the old URL, and do not leave superseded PDFs indexed. Archive what compliance needs to keep in a location search engines cannot reach.

Finally, align partners. Brokers, agents, aggregators and distributors often republish your terms. Give them a single source to link to, and include accuracy of AI-visible content in partner reviews. A wrong figure on a high-authority partner site can outweigh your own page.

06What governance should a regulated brand put in place?

Start from your existing financial-promotions or medical-review process, and extend it to AI answers rather than creating a parallel one. Our guide to AI visibility governance covers roles in general; these steps add the regulated layer.

1. 1
   Build an approved fact registerFoundation

   One owner-maintained list of regulated facts per product and market: rates, fees, limits, eligibility, warnings, with effective dates and approver.
2. 2
   Test the questions buyers askMonthly

   Run prompts about each product across major AI engines and markets on a fixed cadence. Log answers against the register.
3. 3
   Triage by severityWithin 24 hours

   Severity 1: wrong rate, cover, eligibility, safety or legal status. Severity 2: outdated but not harmful. Severity 3: tone or framing.
4. 4
   Fix sources with sign-offPer issue

   Corrections to pages, PDFs and partner content go through compliance review like any promotion. Retire superseded documents.
5. 5
   Prepare customer-facing linesSeverity 1

   Give service teams an approved response when customers cite a wrong AI answer, pointing to the official source.
6. 6
   Keep the evidenceAlways

   Store screenshots, prompts, dates, sources traced, fixes shipped and re-test results in one log.

07How should we handle a severity 1 error?

Treat it as an incident. Notify compliance and legal, capture evidence before anything changes, and check whether customers have acted on the wrong information through complaints, call notes or chat logs. Fix your own sources first, then file feedback with the platform and ask third parties to correct theirs; each platform's channel is covered in our guide to reporting incorrect information to AI platforms.

After the fix, keep the incident open until re-tests across engines and markets show the corrected fact, then record the closing date. The time from detection to corrected answer is a useful measure to report to the risk committee.

08Does this apply to our own AI chatbot too?

Yes, and more directly. The Air Canada ruling and FINRA's notice both point the same way: your own AI assistant is a channel you answer for.^[4][8] Ground it on the approved fact register, restrict it from generating rates or eligibility not in that register, log conversations, and include it in the same monthly test.

09Who should own AI accuracy in a regulated company?

Marketing usually sees the problem first, compliance owns the rules, and product owns the facts. One accountable owner must connect them, with a standing route into the approval chain. OwnlyReal works as that kind of embedded team: measuring what AI says across engines and markets, fixing sources at the root, and shipping changes through draft, brand, legal and regional sign-off, as described for enterprise teams.