What Is First Call Resolution (FCR)?

What Is First Call Resolution (FCR)?

First Call Resolution (FCR) is a contact-center and AI-agent quality metric that measures whether a customer issue is fully resolved in the first interaction, without a callback, transfer, or follow-up. In production voice and chat systems, FCR appears in CRM outcomes, support traces, and evaluator runs that compare the agent’s claimed resolution with later repeat contacts. FutureAGI treats FCR as an evaluator target because fluent answers can still leave the user’s task unfinished.

Why First Call Resolution Matters in Production AI Agents

A voice or chat AI that scores 92% on transcript accuracy can still drop FCR through the floor. The agent may answer fluently, the user may say “thanks,” and the case may close — only for the same user to call back the next day with the same issue because the agent confidently delivered a wrong or incomplete resolution. Pure quality signals like word error rate, fluency, and even sentiment do not catch this. FCR does.

The pain spans roles. A CX leader watches operating cost rise after deploying an AI agent because the deflection rate looked fine but callback rate quietly doubled. A platform engineer sees no errors in traces because no traces failed — they all “completed.” A compliance owner asked whether the AI agent meets the same FCR floor as humans cannot answer without a measurement layer. End users churn quietly.

In 2026 voice and chat agent stacks built on Pipecat, LiveKit, or OpenAI Realtime, FCR is the cleanest cross-channel KPI. Multi-step agents complicate it: a single call may trigger a tool query, a knowledge-base lookup, a clarification turn, and a transfer; the question “was this resolved on first contact” requires evaluating the trajectory, not just the final response.

How FutureAGI Handles FCR Evaluation

FutureAGI’s approach is to model FCR as an evaluator-target on top of the trace, not a raw post-call survey field. The anchor surfaces are TaskCompletion, ConversationResolution, and CustomerAgentConversationQuality, plus the LiveKitEngine simulation surface for pre-production testing.

Concretely: a voice-AI team running a support agent on Pipecat instruments the call with traceAI; every conversation lands as a parent span with nested LLM, ASR, and tool spans. After the call closes, TaskCompletion evaluates whether the user’s stated goal was met given the full transcript; ConversationResolution scores whether the conversation reached a clean close versus an ambiguous handoff; CustomerAgentConversationQuality adds rubric grading for clarity and helpfulness. The team then joins this evaluator data with 7-day callback records from the CRM to produce eval-predicted FCR and measured FCR per route, language, and customer cohort. When evaluator-FCR is high but measured-FCR drops, the gap is the silent failure mode — the agent thought it resolved the issue; the customer disagreed.

For pre-production, the team uses Persona and Scenario with LiveKitEngine to simulate hard cases (angry customer, ambiguous symptom, multi-issue call) and predicts FCR on each before shipping. FutureAGI’s approach treats FCR as a regression target, not just a quarterly board metric.

How to Measure or Detect First Call Resolution

FCR for AI agents is a join of evaluator scores and downstream contact data:

• TaskCompletion — returns 0–1 plus a reason for whether the agent finished the user’s actual goal across the call trajectory.
• ConversationResolution — scores whether the conversation reached a clean resolution rather than a transfer or hand-wave.
• CustomerAgentConversationQuality — rubric score for clarity and helpfulness; predicts callback risk.
• Repeat-contact rate (downstream signal) — the share of calls followed by another contact within 7 or 14 days; the ground-truth FCR signal.
• Transfer-rate — how often the AI hands off to a human; high transfer rate caps FCR at the handoff point.
• Eval-predicted vs. measured FCR delta — the gap between evaluator-predicted resolution and CRM-measured resolution; the tuning target.

from fi.evals import TaskCompletion, ConversationResolution

result = TaskCompletion().evaluate(
    input="I can't log into my account.",
    output="I've reset your password — check your email.",
    context=trace_transcript,
)
print(result.score, result.reason)

Common mistakes

• Confusing deflection with resolution. A bot that closes a chat without answering deflects the call from a human queue; that is not the same as resolving the issue.
• Counting “no callback in 24 hours” as resolved. Customers often retry the next business day; use 7 or 14-day windows.
• Excluding transfers from the denominator. A high transfer rate makes raw FCR look great; report transfer-adjusted FCR.
• Trusting transcript-level evaluators alone. Evaluator-FCR predicts measured FCR but isn’t it; track the delta and tune.
• Ignoring per-cohort FCR. Aggregate FCR hides regressions in narrow but high-value customer segments.