1. Suppliers are competing over the work surrounding the call
HighLevel’s public Voice AI product-management vacancy describes responsibilities spanning telephony, CRM, compliance, analytics and workflow integration. A vacancy is evidence of stated organisational priorities—not proof that a product has achieved those priorities or that customers are obtaining a particular return. (jobs.lever.co)
Planning implication: In a vendor comparison, score record updates, exception ownership and recovery alongside speech quality. Ask who maintains each integration when a business system changes.
2. Reliability changes deserve their own release gates
Retell’s 24 August changelog describes fixes involving transfer handling and duplicate tool actions. Its July entry describes a process that proposes changes for review rather than silently applying them to live agents. These are supplier disclosures, not an independent estimate of remaining defect rates. (retellai.com)
Planning implication: Require a versioned test pack and approval process for changes to models, prompts, tools and routing. Our preferred rule is to rerun the affected tests before releasing a change—not wait for a customer to expose the regression.
3. Benchmark performance is not deployment performance
The August 2026 DuplexWorld preprint evaluates voice agents across 156 scenarios and separates task, conversational and speech-naturalness performance. Its results concern that benchmark and tested systems; they are not a universal failure rate for deployed AI receptionists. (arxiv.org)
Planning implication: Use external research to improve test coverage. Do not substitute a benchmark leaderboard for a trial against your appointment rules, caller mix and telephone infrastructure.
4. Customer experience needs an independent check
Deloitte’s June 2026 survey release reports that more than half of surveyed consumers thought service quality stayed the same or worsened in 2025, while leaders’ assessments improved. Its findings concern broader contact-centre operations, not voice agents alone, and do not establish that AI caused either outcome. (deloittedigital.com)
Planning implication: Keep customer effort, complaints and repeat contacts in the pilot scorecard. An internal automation dashboard should not be the only judge of service quality.