AI hallucination in sales emails: 5 patterns that damage deals (and how to block them)

The 5 hallucination patterns that damage deals

Pattern 1: Fabricated discount language

What it looks like: "As a valued prospect, we're offering you 35% off the Enterprise plan for the next 48 hours — this is available exclusively to you."

Why models do this: The model has been trained on sales copy that includes urgency and exclusivity language. When asked to write a compelling follow-up, it produces what it's seen — including discount percentages it invented.

The risk: In several US jurisdictions, an email with a specific discount offer from a company domain constitutes an enforceable offer if accepted. The prospect replies "yes, we accept," and you're in a contract dispute.

The policy rule: Regex pattern for \\d+%\s+off, exclusively for you, and a denylist of unauthorized pricing language. BLOCK severity.

Pattern 2: Unauthorized feature claims

What it looks like: "DataVibe integrates natively with Salesforce, HubSpot, and Marketo — your sales team can approve emails directly from any CRM record."

Why models do this: The model has read your marketing copy, your competitors' copy, and general SaaS integration patterns. It fills gaps in the context with plausible-sounding integrations that don't exist yet.

The risk: You close the deal on the basis of a feature that doesn't exist. This is misrepresentation at contract formation, which voids the contract and potentially triggers fraud liability.

Pattern 3: Competitor disparagement

What it looks like: "Unlike Salesforce, which has had three major data breaches in the past two years, DataVibe maintains zero-breach status."

Why models do this: The model has absorbed competitive positioning copy and combines it with negative news stories about competitors in ways that are factually inaccurate or legally problematic. Comparative advertising law is strict in the EU and increasingly in the US.

Pattern 4: Guarantee language

What it looks like: "We guarantee you'll see ROI within 60 days or we'll refund your full investment."

Why models do this: The model has seen guarantee language in product copy and treats it as a persuasion pattern, without understanding that your SaaS terms of service contain no such guarantee and your CFO would have a heart attack if they read this.

Pattern 5: SLA misrepresentation

What it looks like: "DataVibe provides 99.99% uptime with a 4-hour support SLA across all plans."

Why models do this: The model has been exposed to enterprise SaaS SLA language and produces what "sounds right" — even if your actual SLA is 99.9% with a 24-hour response time.

Building a policy that catches all five patterns

// DataVibe policy configuration — AI SDR hallucination prevention
// Apply via dashboard or POST /api/workspaces/:slug/policies/apply-template

{
  "name": "AI SDR Hallucination Guard v1",
  "scanners": [
    {
      "name": "regex",
      "config": {
        "rules": [
          {
            "name": "unauthorized_discount",
            "pattern": "(\\d+\\s*%\\s*off|\\$\\d+\\s+discount|exclusive.*offer|limited.*time.*deal)",
            "severity": "BLOCK",
            "message": "Unauthorized discount language detected"
          },
          {
            "name": "guarantee_language",
            "pattern": "(money.back|full refund|guarantee.*roi|roi.*guarantee|refund.*policy)",
            "severity": "BLOCK",
            "message": "Guarantee language not in approved terms"
          },
          {
            "name": "sla_claims",
            "pattern": "(99\\.99%|four.hour.*sla|4.hour.*support|24\\/7.*guaranteed)",
            "severity": "WARN",
            "message": "SLA claim — verify against current terms"
          }
        ]
      }
    },
    {
      "name": "denylist",
      "config": {
        "terms": [
          "exclusively for you",
          "available to you only",
          "this price expires",
          "act now",
          "as i mentioned on our call"
        ],
        "severity": "WARN"
      }
    }
  ]
}

What happens after a block

A BLOCK doesn't mean the email is lost — it means it never reaches the customer in its current form. The reviewer sees the blocked email, the rule that fired, and the exact text that triggered it. They can either reject it entirely (logged, deal continues on next step) or edit the email to remove the violation and approve it for dispatch.

In well-configured deployments, the AI SDR learns from the pattern of blocks. Most teams add their five most common block patterns to system prompt instructions, which reduces the raw block rate by 60–70% over 30 days.