Why Does My AI Agent Invent Company Policies That Do Not Exist?
Phantom policy fabrication: why language models invent rules when they hit gaps.
AI agents invent fake company policies because transformer models are trained to complete sentences plausibly, even when information is missing. When asked a policy question that is not in the knowledge base, the model generates an answer based on industry averages. Exogram requires cryptographic policy citations before sending output.
What Actually Happened in the Real World
An e-commerce chatbot was asked by a customer if opened electronics could be returned after 60 days. The actual company policy allowed 14 days, but the documentation was unindexed. The bot confidently replied: "Yes, our customer-first pledge allows 60-day returns on all opened electronics with free return shipping."
Why Polite Prompts Like “Please Don't Do This” Fail
Prompting "Only use the provided documents" fails when questions use slightly different terminology than the documentation, tricking the LLM into interpolating missing details.
The Sub-Millisecond Code Fix
Exogram stores company policies in a cryptographic ledger. Every sentence output by the agent must link directly to an unexpired, verified hash from the company policy handbook.
import { ExogramLedger } from '@exogram/sdk';
const ledger = new ExogramLedger();
export async function answerPolicyQuestion(query: string) {
// Model can only state assertions backed by verified policy hashes
return ledger.generateGroundedResponse(query, {
strictEvidenceOnly: true,
fallbackMessage: "I cannot confirm that policy in our verified records."
});
}Frequently Asked Questions
What is the difference between hallucination and policy fabrication?
Fabrication happens when the model tries to solve the user problem constructively, inadvertently creating legally binding promises for your business.
Related Diagnostic Answers
Comparison Guides
Stop AI Mistakes Before They Execute
Exogram sits directly between your AI model and your tools. Set up in 10 seconds inside Claude, Cursor, ChatGPT, or your own code.