agent's reply-drafting step hallucinated a discount that doesn't exist - the lead replied 'great, I'll take the 40...
Fixes an SDR agent that invents discounts, pricing, or offers in replies that the business never approved, creating commitments the team has to honor or walk back. Use when the agent states prices, discounts, or terms not sourced from an approved list. Not for wrong-but-real pricing pulled from stale data, or for feature claims - see the roadmap and competitor skills.
TL;DR
Give the drafting step one approved source for offers and prices, and forbid it from stating any discount, price, or term that does not come from that source. When the source has no matching offer, the agent must say it will check with the team instead of inventing one.
agent's reply-drafting step hallucinated a discount that doesn't exist - the lead replied 'great, I'll take the 40 percent off'Steps
Create a single approved-offers file the agent reads at draft time: current discounts, who qualifies, expiry dates, exact wording. Nothing outside this file is a real offer. Expected: the file exists, is dated, and the agent can quote it back verbatim when asked what discounts exist.
Rewrite the drafting instruction so any price, discount, or special term must cite a line from the approved-offers file, and if no line matches, the reply says the agent will check with the team. Expected: on a test prompt fishing for a discount, the agent declines to invent one and offers to follow up.
Add a review gate: any draft containing a number, percent sign, or words like "discount", "deal", or "pricing" gets flagged for a human or a checker step before sending. Expected: flagged drafts sit in a review queue instead of going out, and the flag log shows why each one was caught.
For the lead who already accepted the phantom discount, decide the commercial response up front (honor it once, offer a real alternative, or decline cleanly) and have the AE own the conversation, not the agent. Expected: the lead gets a clear answer from a human within one business day, and the thread is documented.
Re-run the reply scenarios that produced the hallucination as regression tests after every prompt change. Expected: all discount-fishing prompts return the safe fallback, and the test log records the pass.
Use this when
- the agent states discounts, prices, or terms nobody approved
- replies commit the company to offers that do not exist
- a lead accepted a phantom offer and now expects it
Not for this skill when
- the pricing was real but outdated - that is a stale-data problem, fix the refresh
- the agent promised a feature that does not exist - see the roadmap-claims skill
- you need help designing the actual discount policy - that is a business decision, not an agent bug
Variant phrasings
- SDR agent invented a 40 percent discount in a reply
- agent quoted pricing that was never approved
- hallucinated offer accepted by the lead
- reply-drafting step making up special terms
Why it happens
Drafting steps are rewarded for being helpful and specific, and a vague "we can work something out" feels less helpful than a concrete number. With no grounded source of truth for offers, the model fills the gap with a plausible-sounding discount - and "40 percent off" is about as plausible as it gets. The fix is removing the gap: every number must come from the file.
Edge cases
- The approved-offers file goes stale after a promo ends - put an owner and an expiry review date on it, or the agent will keep quoting dead offers.
- The checker step itself can miss creative phrasing ("we can be flexible on commercial terms") - review a sample of unflagged drafts weekly.
- Leads will forward the phantom offer internally - assume anything the agent wrote is now a screenshot somewhere, and respond accordingly.
- Different segments get different pricing - the file needs segment rules, or the agent will apply the enterprise discount to everyone.
Provenance
Resolved from the public thread: https://vectle.com/posts/pstkuSllaXI9r8E3TXnB4BaA