Notes on AI for insurance distribution
Short explainers on doing AI right on an insurance floor, grounded in published research and standards. Each note is sourced, hedged where the evidence is thin, and written for people who run a floor, not a lab.
Three things decide whether AI helps a floor
Plain notes on each, with the figure that sizes it.
- What the paperwork actually costs. Sized from your own calls, not a benchmark.
- How the call reaches a system. Live media, a recording, or neither.
- What has to happen before a record moves. A person approving it.
No vendor hype. Every figure is sourced at the foot of the page, and anything we could not confirm directly is hedged.
Explainers
Six topics worth getting right before you put AI near a call.
What after-call work really costs an insurance floor
After every call, licensed agents stop selling and start typing: CRM records, applications, notes, and dispositions. Industry benchmarks put this after-call work at roughly 6 to 12% of an agent's paid time, and compliance-heavy insurance lines tend toward the high end.
Size it on your own floor ›SIPREC vs recording-based capture
SIPREC (IETF RFC 7866) is a standard way for a phone system to stream live call media to a recording server, separate from working off a finished recording. The practical difference is timing: live capture can support in-call help, while recording-based capture runs after the call ends, and which one a floor can use depends on what its dialer exposes.
The three capture tiers, compared ›Why most floors will not let AI touch a record
In McKinsey's 2025 survey, inaccuracy was among the most cited AI risks, and risk concerns ranked as a top barrier to scaling. On a regulated insurance record a wrong field is not a typo, so most floors will not let software write to a record unsupervised.
The questions that surface this on a demo call ›Provenance over generation: safe AI on regulated data
Generation invents plausible text; provenance ties every field back to a moment in the call and flags anything unsure. Since manual entry already carries reported error rates around roughly 0.3% to 6.6%, the safer design is one that shows its source and waits for a person to approve before anything moves.
What provenance means in practice ›After-call automation vs full agent replacement
A field study of more than five thousand support agents found that an AI assistant raised issues resolved per hour by about 14% on average, with the largest gains for newer agents. That is the augmentation case: AI handles the after-call work while the licensed agent stays on the phone, rather than replacing the agent.
How to evaluate that claim in a demo ›How to measure your own after-call cost
You do not need our numbers; you can size your own. Time a sample of calls from hangup to "ready for the next call," multiply by call volume and loaded agent cost, then compare against published after-call-work benchmarks to see where your floor sits before you change anything.
The measurement, step by step ›Put these notes to work on your floor
We start with a paid diagnosis sized on your floor's own calls and numbers. You leave with a plan whether or not we build.
Sources
- Voiso, "Average After-Call Work Time." After-call work as 6 to 12% of agent time. voiso.com
- SQM Group, "Industry Standards for Top Call Center KPIs." Call-center KPI benchmarks. sqmgroup.com
- IETF, "RFC 7866: Session Recording Protocol" (SIPREC), Standards Track, May 2016. Standard for streaming live call media to a recording server. rfc-editor.org
- McKinsey, "The state of AI in 2025." Inaccuracy among the most cited AI risks; risk and security concerns a top barrier to scaling. mckinsey.com
- Brynjolfsson, Li & Raymond, "Generative AI at Work" (NBER Working Paper 31161; QJE). AI assistant raised resolutions per hour ~14%, largest for novices. nber.org
- Garza et al., "Error Rates of Data Processing Methods in Clinical Research" (systematic review, NCBI/PMC). Manual data-entry error roughly 0.3% (keyed) to 6.6% (record abstraction). ncbi.nlm.nih.gov