6 metrics to track after deploying your AI support agent

Updated on October 9, 2026

What metrics should I track after deploying an AI support agent? It’s the question leadership asks within weeks of launch, and most teams don’t have a clean answer. The metrics exist; the problem is knowing which ones signal genuine health versus which ones just look good in a slide deck. This guide covers six metrics that matter most after deployment, how to calculate each one, what good looks like, and exactly what to do when a number trends the wrong way.

The gap between “we deployed an AI agent” and “we know it’s working” is almost always a measurement gap. Teams track conversation volume or glance at a single automation percentage, then miss the signals that reveal whether the agent is actually solving problems or just containing complaints. Getting this right requires tracking a small, interconnected set of KPIs that together tell the full story, covering everything from chatbot metrics like fallback rate to broader conversational AI analytics like cost-per-ticket trends.

Some platforms surface these metrics through real-time dashboards and transparent reasoning logs, so you can start measuring from day one instead of piecing together data across five separate tools. Capabilities vary by vendor, so it’s worth confirming what your platform provides before you’re deep into deployment. What’s consistent across high-performing teams is that the ones who improve fastest measure consistently and act on what they see, not the ones with the most sophisticated technology.

1. Automation rate: the number every stakeholder asks for first

This is the headline metric after any AI support deployment. It tells you what percentage of incoming support conversations your AI agent resolved end-to-end, without a human ever touching the conversation. Get this number wrong and every business case you’ve built falls apart.

How to calculate it correctly

The formula is straightforward: AI-only resolved conversations divided by total qualifying conversations, multiplied by 100. The word “resolved” is where most teams get tripped up. A conversation should count as resolved only when the customer’s issue was genuinely completed, no handoff occurred, and there was no same-topic follow-up within 24 hours. Silent abandons, where the customer leaves because the AI failed, should never count as successful containment.

Also be precise about your denominator. Does it include contacts the AI never handled, or only conversations where the AI was the first responder? That distinction can shift your reported rate significantly, so document your definition and apply it consistently across every reporting period.

What a healthy automation rate looks like

A realistic first-month range sits between 40 and 60% for most teams. Mature deployments with well-trained agents typically reach 70 to 85% on high-volume, transactional intents like order status or password resets. Chasing 90%-plus containment across all intents often means the AI is blocking escalations rather than genuinely resolving issues. Always read this metric alongside CSAT and recontact rate, not in isolation.

What to do when automation rate is low

The most common culprits behind low automation are narrow intent training and an overly conservative confidence threshold, though sometimes incoming queries simply fall outside what the agent was built to handle. Pair automation rate with fallback rate to diagnose. A high fallback rate, often above roughly 15%, can indicate intent-coverage or routing gaps worth investigating through intent-level analysis. Expanding the agent’s knowledge base with real customer queries from your queue is a practical first step, but confirm the root cause before acting.

2. CSAT for bot interactions: measuring satisfaction without bias

CSAT tells you whether automation is actually helping customers or just deflecting them. A bot that contains 80% of conversations but leaves customers frustrated is worse than one that handles 50% well. Volume without quality is a liability, not a win.

How to collect CSAT specifically for AI interactions

Trigger a micro-survey immediately at the end of every bot-only conversation, separate from your post-agent CSAT collection. This keeps signals clean and prevents human-agent performance from masking AI problems. Use a simple 1, 5 scale or a thumbs-up/down with an optional free-text field. Keep in mind that the minimum response rate needed for statistical confidence depends on your traffic volume and desired confidence interval, as a rule of thumb, many teams require at least 10 to 15% response rates before treating the aggregate number as meaningful, though higher-traffic deployments can set a lower percentage threshold and still achieve adequate sample sizes.

Industry benchmarks by vertical

AI CSAT of 80 to 87% is broadly healthy across verticals; 88% and above is excellent. The practical guardrail: your AI CSAT should not fall more than 5 to 10 percentage points below your human-agent CSAT. If it does, the gap usually points to poor intent recognition, weak answer grounding, or a broken escalation path. Regulated industries like insurance and financial services naturally have lower ceilings, around 72 to 82%, because the queries themselves are more complex and the stakes are higher.

When CSAT drops: diagnosing the cause

Filter CSAT by intent type and channel before drawing conclusions. A single poorly performing intent can pull down the aggregate average significantly, making the overall number look worse than it actually is. Reasoning logs, where your platform provides them, let you see exactly what the agent cited when it answered poorly. Low-confidence responses and weak answer grounding are frequent culprits behind unexplained CSAT drops, and intent-level filtering is usually the fastest way to isolate them.

3. First-contact resolution: separating genuine fixes from clever deflection

FCR is the most honest measure of whether your AI agent is actually solving problems. A conversation can be contained, meaning no human was involved, and still leave the customer’s issue completely unresolved. FCR corrects for that gap and holds the agent accountable to actual outcomes.

How to calculate FCR for an AI agent

Issues resolved on the first contact divided by total qualifying first contacts, multiplied by 100. Confirm resolution with a follow-up window of 7, 14, or 30 days depending on your support context. Seven days works well as a general default because shorter windows like 24 hours tend to overstate FCR, capturing customers who initially accept an answer but return later with the same issue. Report three variants where possible: AI-only FCR, overall FCR including clean AI-to-human handoffs, and first-contact automation resolution.

The recontact rate: FCR’s companion metric

Recontact rate measures how many customers come back about the same issue after a conversation marked as resolved. Track recontact rate alongside FCR because a high FCR combined with a high recontact rate signals that your resolution criteria are too loose. A recontact rate below 8 to 12% is a reasonable starting target for AI-resolved conversations, treat it as a heuristic to tune against your specific intent mix rather than a universal rule. Anything higher typically means customers are accepting answers in the moment but discovering those answers didn’t actually solve the problem.

FCR targets and what poor FCR signals

As operational benchmarks, healthy AI FCR for transactional intents often falls between 75 and 85%. For troubleshooting or multi-step issues, 55 to 70% is realistic. These ranges should be validated against your own intent complexity and traffic mix. If FCR consistently falls below 50%, the agent is likely answering adjacent questions rather than the actual query, or the knowledge base it pulls from is outdated and needs a systematic review against recent ticket data.

4. Escalation rate and handoff quality: reading the signal correctly

Escalation rate is the inverse of automation rate, but it tells a different story. A well-calibrated AI agent should escalate confidently on complex or high-stakes queries, not fail silently and loop the customer through unhelpful responses until they give up.

What your escalation rate is actually telling you

Calculate escalation rate by dividing conversations handed off by total qualifying AI conversations. A rate above 35 to 40% often means the agent’s training coverage is too narrow. A rate below 5% is a red flag: it typically means the agent is failing to route appropriately, not that it’s genuinely resolving everything. The ideal range for most teams is 15 to 25%, with escalations happening because of complexity, sensitivity, or user preference rather than AI uncertainty.

Handoff quality: the metric most teams miss

Measure transfer quality by tracking how often the receiving agent has to ask the customer to repeat information already shared with the bot. A well-designed AI-to-human handoff preserves full conversation context and detected intent, so agents can pick up mid-conversation without starting over. Poor handoff quality drives a secondary CSAT drop that often gets attributed to human agents when the real cause is an incomplete or missing context transfer during the handoff.

Kommunicate’s handoff architecture is built around this principle, context and detected intent travel with the conversation so agents aren’t starting cold. Confirm that whichever platform you use handles context transfer explicitly, because many don’t by default.

Corrective actions when escalation rate trends poorly

If escalation rate climbs week over week, run a breakdown by intent. Identify which topics are escalating disproportionately and build out the agent’s knowledge for those intents first. If escalation quality is poor, measured by a high repeat-question rate from receiving agents, audit the context package being passed during handoff and verify that every relevant piece of information is included before the transfer completes.

5. Average handle time and cost-per-ticket: making the business case

These two metrics translate AI support performance into the language finance and operations teams understand. They also reveal whether automation is creating real efficiency or just shifting work from one column to another without reducing total cost.

Average handle time after AI deployment

AHT measures total time to resolve a case, including bot interaction, agent time, and any post-contact work. After AI deployment, AHT for agent-handled escalations should decrease because agents receive pre-filled context and don’t spend time diagnosing the problem from scratch. Teams with effective AI assistance commonly report human-agent AHT reductions in the 30 to 40% range. If AHT increases post-deployment, the handoff process is adding friction rather than removing it, and that handoff design needs an immediate review.

How to calculate cost-per-ticket

Total support operating cost divided by total resolved contacts gives you cost-per-ticket. A fully AI-resolved conversation typically costs between $0.50 and $2.00 depending on model and infrastructure; a human-handled ticket averages $6 to $14 across most mid-market support teams. Track cost-per-ticket monthly and segment by AI-resolved versus agent-resolved to show the financial impact of improving your automation rate by even five percentage points. That incremental improvement compounds quickly at scale.

6. Building your monitoring dashboard and alert thresholds

Tracking these metrics once is useful. Having them update in real time and alert you when something breaks is what separates teams that consistently improve from teams that only react when a problem becomes visible to customers or leadership.

The minimum viable dashboard setup

Your AI support dashboard should surface six panels at minimum: automation rate, CSAT by channel, FCR with recontact trend, escalation rate by intent, AHT comparison between AI-resolved and agent-resolved contacts, and cost-per-ticket. Kommunicate provides out-of-the-box dashboards that surface these metrics alongside reasoning logs for individual conversations, so teams can drill from aggregate trends down to specific failure cases without switching between tools. If you’re evaluating platforms, this kind of integrated view, combining conversational AI analytics with individual reasoning logs, is worth prioritizing in your requirements.

Alert thresholds that actually trigger action

Set relative alerts, not just absolute ones. Alert when automation rate drops more than 5 percentage points week over week, when CSAT falls below your defined floor, when escalation rate exceeds 30%, or when fallback rate crosses 15%. Treat these as suggested starting points to tune against your traffic volume and noise levels rather than universal rules. Relative changes catch regressions faster because they flag directional movement before it becomes a crisis that affects real customers, absolute thresholds alone are easy to miss during gradual drift.

Review cadence for sustainable improvement

Check volume, CSAT, and fallback rate daily. Review FCR, escalation breakdown by intent, and recontact rate weekly. Run a full cost-per-ticket analysis and compare AHT trends monthly. After every knowledge base update or model change, run a comparison against your prior baseline to confirm you haven’t introduced regressions. Teams that treat metric review as a weekly ritual, rather than a quarterly audit, consistently outperform those that only look at data when something breaks.

What metrics should I track after deploying an AI support agent? Start here.

Deploying your AI support agent is the starting point, not the finish line. Automation rate tells you how much the agent is doing. CSAT and FCR reveal how well it does it, escalation rate surfaces where it needs reinforcement, and AHT with cost-per-ticket confirm whether it’s delivering the projected business value.

Track all six together, set alert thresholds before problems become crises, and review intent-level breakdowns regularly. The teams that get the most out of AI support aren’t the ones with the most sophisticated technology. They’re the ones who measure consistently, act on signals quickly, and treat their metrics as a feedback loop rather than a report card.

To move from guessing to measuring, the Kommunicate team can show you how real-time dashboards and reasoning logs give your team full visibility into these KPIs from day one, so you always have a clear answer when leadership asks whether your AI support agent is working.