By Ayan Chakraborty, Founder, SBA Innovation Pvt Ltd, Kolkata. Last updated . Reviewed by Ayan Chakraborty.
Marketing summary, not legal advice. Rules and platform policies change; confirm every point against the official source and your compliance team before you act.
How do I fix wrong rates or fees that AI assistants state about my lender?
Treat a wrong fee in an AI answer as a compliance problem first and a marketing problem second. Record the exact prompt, assistant, date and answer, then work out where the figure came from, because assistants repeat what their sources say. Usually one or two stale pages, such as an aggregator listing or an old product page, feed many answers.
Fix it at the source: ask the aggregator to update the page, update or redirect your own outdated pages, and publish the current rate, fee and rule on a dated page in plain HTML that states what changed. Then re-run the prompt on a schedule and log whether the answer corrected. You cannot edit an assistant's answer directly, and corrections take weeks to propagate, so keep a remediation queue and re-check until each error clears.
- Log the prompt, assistant, date and the wrong statement.
- Find the pages the assistant cites or likely reads.
- Request corrections from third-party sources and update your own old pages.
- Publish the current fee, rate and rule in plain HTML with a date.
- Re-run the prompt three times and log whether the error cleared.
- Keep unresolved errors in an open remediation queue.
From our client work: a fee scrapped 14 months earlier
For a consumer lending NBFC, 21 AI answers stated a processing fee and a foreclosure rule the lender had discontinued 14 months earlier. Fourteen of the 18 corrected errors traced to two aggregator pages carrying the old numbers, and correcting those sources changed the answers within four to nine weeks. Answers containing an error fell from 21 to 3.
Method: Engagement 031: errors defined as contradictions of the client's current published fee, rate range, tenure or eligibility terms, found across a 300-prompt set run three times per assistant.
What did not work: Three errors remain on low-refresh surfaces and sit in the open remediation queue; some corrections are slow and cannot be forced.
Read case study 031 Prompt set, baseline, assumptions and footnotes.