TL;DR: Resolution vs deflection: one tracks whether a customer avoided an agent, the other whether their problem was solved. Optimizing for deflection alone can quietly hurt resolution, so pair it with outcome-based metrics and vendors that verify answers, not just retrieve them.
One tells you a customer left. The other tells you why. Most support dashboards only report the first.
Deflection rate and resolution rate get treated as if they mean the same thing: the customer didn’t need an agent. They don’t. Deflection measures whether a customer avoided human contact. Resolution measures whether their problem was actually resolved. A ticket can score well on one and fail completely on the other, and most support reporting has no way to tell the difference.
The gap is bigger than most teams assume. Gartner estimates that only 14% of customer service issues are fully resolved through self-service, even as deflection rates get reported as a win. In this blog, we break down what each metric actually measures, where they pull apart, and what it takes to close that gap.
Table of Contents
- What Deflection Rate Actually Measures
- What Resolution Rate Actually Measures
- Resolution Vs Deflection: Side-by-Side Comparison
- Where the Two Metrics Diverge
- Why Optimizing for Deflection Alone Backfires
- What Measuring Resolution Requires
- Metrics that Close the Gap
- What to Ask Vendors During Evaluation
- Final Thoughts
- FAQ
What Deflection Rate Actually Measures
Deflection rate is the percentage of support requests handled through self-service, such as a help center article, a chatbot, or an FAQ, instead of reaching a live agent. It’s calculated as the number of deflected cases divided by the total number of cases. It’s one of the most widely reported support metrics because it’s easy to track and directly tied to cost: fewer agent-handled tickets mean lower cost per interaction and better scalability as volume grows.
It’s a genuinely useful metric for measuring self-service reach, telling you how much volume a knowledge base or bot absorbs before it hits a human queue.
What Resolution Rate Actually Measures
Resolution rate is the percentage of cases where the customer’s actual problem was solved, such as a refund processed, an account updated, or an accurate answer confirmed against current data, regardless of whether an agent was involved. It incorporates the outcome, not just the channel, which makes it harder to calculate since it requires confirming what happened after the interaction ended.
It’s also the metric that correlates most closely with what customers actually experienced: whether they had to come back, whether they were satisfied, whether the issue is genuinely closed.
Resolution Vs Deflection: Side-by-Side Comparison
| Dimension | Deflection Rate | Resolution Rate |
| What it measures | Contact avoided | Problem solved |
| Best used for | Volume and cost forecasting | Quality and customer outcomes |
| When it’s knowable | At the moment of contact | Only after the outcome is confirmed |
| Counts abandonment as success | Yes | No |
| What it predicts | Short-term cost savings | CSAT, retention, repeat contact |
Neither metric replaces the other. Deflection shows how much self-service is absorbing. Resolution shows whether that absorption is actually working.
Where the Two Metrics Diverge
The two are meant to move together. A good self-service experience should deflect and resolve. In practice, they often don’t, because deflection counts every self-service interaction that avoided an agent identically, whether or not it worked.
Genuine resolution. The customer found the right answer, completed the task, and had no reason to come back.
Silent abandonment. The customer searched, found nothing useful, and left, not satisfied, just gone. No ticket was logged, so this failure is invisible in the deflection number.
Confident-but-wrong resolution. The customer got an answer and acted on it, but the answer was incomplete or incorrect. This is the most damaging of the three, because it doesn’t just fail to help; it can surface later as a support case, or as churn, with the customer now working from bad information.
Audits of AI-driven support deployments regularly find a meaningful share of “deflected” cases fall into this third bucket. A single deflection percentage can’t distinguish between these three outcomes, which means a team can hit an excellent deflection number while resolution quietly falls.
Why Optimizing for Deflection Alone Backfires
The risk isn’t that deflection is a bad metric. It’s that optimizing for it in isolation rewards the wrong behavior. If an AI agent is tuned purely to raise deflection, the fastest lever isn’t better answers; it’s more confident answers and fewer escalations.
If a team widens the AI agent’s triggers so it answers more questions, and makes it harder for customers to reach a human, deflection will climb. But if those extra answers aren’t actually correct, resolution falls at the same time. A rising deflection rate paired with a flat or falling resolution rate usually means a team is reporting the metric that’s easiest to produce instead of the one that reflects what customers actually experienced.
What Measuring Resolution Requires
Deflection is easy to report because it only requires knowing whether a ticket was opened. Resolution requires knowing what happened after: whether an action was completed and verified against live data, not just whether the conversation ended without escalation. That’s a different infrastructure requirement, one that can confirm outcomes rather than just retrieve and respond. This gap is why most support organizations still lead with deflection in their reporting. It’s the number their existing tools already produce.
Metrics that Close the Gap
Support teams that want a fuller picture usually track these alongside deflection:
- First Contact Resolution, whether the issue was fully closed on the first interaction with no follow-up needed
- 30-day repeat contact rate, since a customer coming back about the same issue means the earlier resolution wasn’t real
- Post-resolution CSAT or CES, measured after the outcome is known rather than at the point of deflection
- Escalation quality, how much of the prior self-service interaction actually helped once a case reaches a human
None of these are harder to justify to leadership than deflection. Tracked alongside it, they turn a volume metric into a quality signal.
Suggested Read: 8 Essential CX metrics to measure customer self-service success
What to Ask Vendors During Evaluation
Most vendor pitches lead with a resolution rate. Very few explain how that number was built. A handful of direct questions during evaluation will tell you more than any demo:
How exactly do you define resolution?
Ask for the specific criteria. Is it graded by the AI itself, does it require a completed action, and is customer confirmation part of the calculation?
Does that number include deflections?
Some vendors fold article views and FAQ hits into their resolution figure. A reported 70% resolution rate that’s really 30 points of self-service traffic and 40 points of actual fixes is a very different story than it sounds.
What percentage of AI-resolved cases turn out to be wrong or incomplete?
This requires checking outcomes after the conversation ends, not just whether the customer came back. Most vendors have never measured it, which is itself useful information.
Is the AI optimized to reduce escalations, or to increase verified outcomes?
These reward different behavior. A system tuned to avoid hand-offs can post strong numbers just by answering more confidently, correct or not.
Does a self-service conversation carry context into a human escalation, or does the agent start from zero?
Systems built around resolution treat escalation as a continuation. Systems built around deflection often treat it as a disconnected event, which slows the case down.
How do you measure customers who abandon self-service without ever opening a ticket?
This group is invisible in most dashboards; a customer who searches, finds nothing, and leaves looks identical to a successful deflection unless a vendor is specifically tracking it.
Does resolution get measured differently for simple questions versus real actions?
Answering a policy question and actually processing a return are not the same task, and a vendor’s resolution rate should be broken down by that difference, not blended into one number.
Can you audit individual conversations marked as resolved?
You should be able to read the transcript, see the reasoning, and judge for yourself whether the case was genuinely closed. If that access isn’t available, treat the resolution number with caution.
Suggested Read: 4 Questions You Must Ask Your Potential AI Vendor
Currently Evaluating an AI Support Vendor?
How SearchUnify Closes the Gap
Most of the vendor questions above come down to one thing: can the system prove a case was actually resolved, not just that the conversation ended?
SearchUnify’s AI Support Agent is built around that distinction. It doesn’t stop at retrieving the closest-matching article. It reasons through the query, pulls real-time data from a company’s own systems, and generates a grounded answer through SearchUnifyFRAG™, which combines large language models with an organization’s indexed knowledge base rather than relying on the model’s general training alone.
That grounding is what separates a verified answer from a confident guess, which is exactly the distinction between resolution and deflection this blog has been making. The same logic carries into escalation. When a case is too complex for self-service, SearchUnify’s agents hand off verified, context-rich cases to human or higher-tier agents instead of starting the conversation over.
The customer’s prior attempts, the system’s reasoning, and the data already gathered move with the case, so escalation becomes a continuation of the resolution process rather than a reset.
See How SearchUnify's AI Support Agent Verifies Resolution in Your Own Support Data
Final Thoughts
Deflection and resolution answer different questions. Deflection asks whether the customer left. Resolution asks whether they were actually helped. Tracking deflection alone tells half the story and lets the other half hide in plain sight. Teams that get this right don’t abandon deflection. They stop trusting it on its own.
FAQ
1. What’s the difference between deflection rate and resolution rate?
Deflection rate measures whether a customer avoided contacting a human agent. Resolution rate measures whether their problem was actually solved. A ticket can be deflected without ever being resolved.
2. Is a high deflection rate good or bad?
Neither on its own. High deflection is good if resolution is also high, since it means self-service is genuinely working. High deflection with flat or falling resolution usually means customers are giving up rather than getting helped.
3. How do you measure resolution rate accurately?
By confirming that an action was actually completed, such as a refund processed or an account updated, rather than simply noting that a conversation ended without escalation. This typically requires connecting self-service tools to backend systems, not just a knowledge base.
4. Can AI improve both deflection and resolution at the same time?
Yes, when it’s built to verify outcomes rather than just retrieve the closest matching article. AI that understands intent and confirms answers against live data can raise resolution without sacrificing deflection. The two metrics are only in tension when self-service quality is weak.
