AI and People, Designed Together
Nearshore & offshore call-center teams that protect your brand — PCI DSS and HIPAA-compliant, and ready to scale with your volume.
“Human in the loop” becomes a fact when it has numbers attached. Three measures do the work: how often the AI hands off and why, how long the caller waits for a person, and whether the person already knows what the caller said. This page defines each, says where the data comes from, suggests reasonable targets, and explains how to read the numbers in the first 90 days.
Measure 1: Handoff Rate and Reason
Definition. The share of AI-answered contacts that were handed to a person, broken down by trigger: caller request, outside scope, policy rule, or signal (frustration, emergency).
Data source. The AI platform’s call log, which should tag each handoff with its trigger. If it cannot, that is a finding about the platform.
Why it matters. The rate drives the unit economics (every handoff consumes human minutes) and tells you whether the AI is answering the right calls. The reason breakdown is where the insight is. A high “outside scope” share means the knowledge base needs work. A high “caller request” share on routine call types means callers do not trust the AI, often because of the greeting or the pace. A high “policy” share is by design.
What the market shows. In one AI receptionist vendor’s analysis of 1,446,980 calls, 73.8% of the AI’s actions were transfers to a person, a share of actions rather than of calls, but a signal of how heavily the handoff path is used 1. Your rate will depend on your call sort; the AI-or-human page in this pillar explains how to set it up.
Measure 2: Time to Human
Definition. Seconds from the handoff trigger to a person speaking to the caller.
Data source. Timestamps from the AI platform (trigger) and the telephony or agent desktop (answer).
Why it matters. This is what the caller feels. Long hold times are the biggest complaint for 39% of consumers, and 49% of Gen Z (CallRail survey of 1,000 US consumers, 2025) 2. California’s AB 1609 now requires the largest businesses to make a good-faith effort to connect a requesting customer with an agent within 15 minutes or offer an appointment within one business day, and limits hold to 15 minutes at any time 3. For a front desk, that is a ceiling; the target is seconds. Report the median and the 90th percentile, because averages hide the calls that went wrong.
Measure 3: Context Carried
Definition. The share of handoffs in which the person did not ask the caller to repeat their name, reason for calling, or details already given to the AI.
Data source. QA scoring of a sample of handed-off calls, using a rubric item worded exactly as the definition above. This measure cannot be automated reliably yet; it needs a human listener, which is why it is often skipped and why it is the most honest.
Why it matters. It is the measure of Requirement 4 of the Handoff Standard and of the single most-cited friction in customer service: 74% of consumers are frustrated by repeating information, and 81% want conversations continued without backtracking (Zendesk, 2026) 4.
Reasonable Targets
Targets should be set from your first 30 days of data rather than from a vendor’s brochure, but the shape is predictable: handoff rate falling as the knowledge base and triggers are tuned, with the “policy” share stable and the “outside scope” share shrinking; time to human in seconds at the median during covered hours, with a clear fallback (callback window, confirmation text) outside them; context carried approaching 100% on QA-sampled handoffs, because anything less is a process failure you can fix.
Reading the First 90 Days
- Days 1 to 30. Expect a high handoff rate and a noisy reason mix. Review every handoff weekly. Fix the knowledge base and the greeting first.
- Days 31 to 60. The rate should fall and the reason mix should settle. Start sampling context carried in QA. Tune the call sort: move call types that fail in practice.
- Days 61 to 90. Set targets from the data. Add caller satisfaction split by AI-only and handed-off calls, and compare with the 88% versus 60% human-versus-AI satisfaction gap in Verizon Business research 5.
How Redial Fits
These three measures are on the scorecard Redial intends to instrument from the first pilot call of the AI front desk it is previewing, IzzyOps (working name), alongside answer rate, AI resolution rate, disclosure compliance, caller satisfaction, after-hours bookings captured, and Spanish-language share. Redial will publish them with the period and call count stated, and not before they exist. The same three measures govern the handoff between any AI tier and Redial’s agents in staffed programs.
Frequently Asked Questions
What is a good handoff rate?
The one your call sort predicts. A high rate on call types you expected the AI to resolve is a problem; a high rate on policy categories you routed to people by design is not. Read the reason breakdown, not the headline rate.
How should time to human be reported?
Median and 90th percentile, in seconds, for covered hours, with the fallback behavior outside covered hours described separately.
Can context carried be measured automatically?
Not reliably yet. It needs a QA listener scoring a sample of handoffs against a specific rubric item. That effort is what makes it credible.
How often should these measures be reviewed?
Weekly for the first 90 days of any new AI tier, then monthly, and whenever any of the three moves noticeably.
Related Pages
- The Handoff Standard: What Human in the Loop Should Actually Mean
- Why Customers Still Want a Person: What the 2025 and 2026 Surveys Show
- The Five Requirements of a Good AI-to-Human Handoff
- AI or Human? Deciding by Call Type With Three Tests
- Early Access: What a Pilot Looks Like and What We Measure
References
- NextPhone, AI receptionist statistics (2025 and 2026 reports) — Vendor operational data from 1,446,980 calls across 2,074 businesses: 28.5% of calls arrived outside standard business hours in 2025 (14.1% under a stricter 2026 definition), and 73.8% of the actions the AI took were transfers to a person.
- CallRail, Missed calls cost businesses more than ever (September 25, 2025) — Vendor survey of 1,000 US consumers: 78% have abandoned a business after an unanswered call, 82% say they will call a competitor if a business does not answer, 42% leave a voicemail, 21% immediately call another business, and 39% name long hold times as their biggest complaint.
- California AB 1609, enrolled text (signed September 28, 2026) — Applies to businesses with more than USD 500 million in national gross annual revenue; bars presenting a chatbot as human, requires disclosure in the medium of interaction, and requires a good-faith effort to connect a requesting customer with an agent within 15 minutes or offer an appointment within one business day; civil penalties up to USD 5,000 and USD 10,000, no private right of action.
- Zendesk, CX Trends 2026 consumer research (November 2025) — Survey of 6,182 consumers in 22 countries, fielded June 2025: 74% are frustrated when they have to repeat information, 81% want agents to continue a conversation without backtracking, 74% expect 24/7 service, and 85% of CX leaders say one unresolved issue is enough to lose a customer.
- Verizon Business and Longitude, AI and the human-first customer experience (August 2025) — Research among 5,000 consumers and 500 executives in seven countries: 88% satisfied with interactions handled mostly or fully by humans versus 60% for AI-driven interactions; 47% name the inability to reach a human as their top annoyance with automation.
See How the Three Measures Are Captured
Book a demo of IzzyOps (working name), the AI front desk Redial is previewing, and we will show where handoff reason, time to human, and context carried come from on a sample call and in the dashboard. Commercial availability is planned for late 2026 to early 2027.