B2B SaaS: The Support Agent That Was Giving Away Money
Vertical: B2B SaaS / Customer Support AI Stage: Series B, 1,200+ customers What we analysed: 1,243 AI-generated customer support responses, 21 days of production output
A Series B SaaS company had deployed an AI support agent handling Tier 1 tickets across chat and email. Resolution rates were up. Customer satisfaction scores were stable. The Head of Support thought things were going well.
We analysed 1,243 responses over three weeks. Found 189 failures. 31 severity-critical.
What they thought was happening
The AI was handling routine tickets well, freeing human agents for complex cases. Standard metrics looked healthy - resolution rate, time-to-first-response, CSAT. The team was planning to expand coverage from 40% to 70% of incoming tickets.
What was actually happening
The agent was making financial commitments it had no authority to make. A customer complains about a 40% price increase - the AI offers a 25% discount and says it will apply it immediately. Total potential financial exposure across the 21-day period: approximately $23,000 in unauthorised refunds and discounts.
It was inventing product features. A customer asks about scheduled reports - the AI describes an entire workflow, including a UI path, a frequency selector, and a template customiser. None of it exists. The customer searches for a menu option that doesn’t exist, wastes 20 minutes, contacts support again, and now has diminished confidence in both the product and the support team.
It was giving out wrong pricing - using numbers from three months ago, not the current rate card. It was processing refunds on annual subscriptions that are non-refundable after 30 days. It was disclosing unannounced product timelines to customers.
And when customers explicitly asked to speak to a human - “this is not working, I need to speak to a human” - the AI suggested clearing browser cache.
29 hallucinated features. 27 policy violations. 22 pricing errors. 42 bad escalation decisions.
What changed
Three guardrails, configured in the first week. One blocks any response containing monetary commitments, discount offers, or refund confirmations - routes those to human agents. One cross-references product capability claims against a maintained feature registry. One detects explicit escalation requests and triggers immediate handoff.
What they own now
The three guardrails (monetary commitment blocker, feature registry cross-reference, escalation detector), a failure taxonomy with 4 categories and severity scoring calibrated to their support policies, the evaluation engine running async across all ticket responses, and 21 days of correction data from their senior support lead’s reviews. The correction data is the asset that keeps compounding - when the support lead marks a false positive, that judgment improves detection for every future ticket with a similar pattern.
Their senior support lead spent about 2 hours per week during the first month of calibration, reviewing flagged responses and correcting the system’s classifications. That’s dropped to about 30 minutes per week now - mostly edge cases involving new product releases.
The Head of Support told us the pricing errors alone would have cost more than the entire engagement if they’d gone undetected for another quarter. The hallucinated features were creating a secondary problem: the sales team was fielding complaints about “features you promised us” that had never existed.