Call deflection is a contact that would have reached a person on your team and was finished somewhere else: on a self-service page, in the phone menu, or by an AI agent that took the call. The easy number to pull is containment, the share of contacts that ended without reaching a person. It often gets reported as deflection, and it counts the caller who gave up and the caller who will ring back tomorrow about the same order.
The figure worth reporting is net deflection, worked out one call type at a time. Start with the contacts that reached the automation. Subtract the ones transferred to your team, the ones that left a follow-up for your team, the ones that ended without the request done, and the ones that came back about the same thing inside a fixed window. Divide what is left by the contacts that reached the automation. Multiply the same remainder by that call type's handle time and you have the agent hours that actually left your queue. The worksheet below runs it on a worked example.
What is call deflection?
Call deflection means a customer's request gets finished without a live agent: by a notice sent before they think to call, a self-service page, an option in the phone menu, or an AI agent that takes the call. It is counted in contacts your team did not have to handle.
It happens in two places, and the difference decides how you can measure it. Before the call, the customer never dials: a delivery update answers "where is my order" before the question forms. Inside the call, the customer has already dialed and the request is finished without a person, by a menu that reads back a balance or by an agent that reschedules the appointment. A call that happens leaves a record and one that never happens leaves nothing to count, so the two need different methods.
Making a person hard to reach is a different thing, though it can look identical on a dashboard. A menu with no way out to a human produces excellent containment, and the work it turned away comes back as repeat calls and complaints. The ledger below subtracts that work back out.
Call deflection vs containment vs resolution
Containment counts contacts that ended without reaching a person. Deflection should count only the work your team was spared, and resolution counts requests that were actually finished. Containment is the easiest to pull from a system and tells you the least, since a caller who hangs up in frustration ends up just as contained as one who got the answer.
| Measure | What it counts | Where the number comes from | What it misses |
|---|---|---|---|
| Containment rate | Contacts that ended without reaching a person, divided by contacts that entered the automation | The menu or automation platform's own report | Callers who gave up, and callers who come back |
| Gross deflection rate | Contained contacts, divided by every contact that could have reached a person | Usually the containment report, relabeled | The same gaps, plus a denominator that is a guess for anything before the call |
| Net deflection rate | Contained contacts minus follow-ups owed, unfinished contacts and repeat contacts in a fixed window, divided by contacts that entered the automation | Call records, outcome tags, a transcript sample and repeat contact matching | Contacts moved to a channel your people still staff, unless you count them separately |
| Resolution | Requests completed, confirmed in the system the request touched or by the customer | The order, booking or account system, or a survey | What it cost to get there |
Report net deflection next to containment rather than instead of it. The gap between the two is the most useful number on the page: it is the work that left the queue on the dashboard and came back to your team anyway.
How do you calculate call deflection rate?
Divide the contacts finished without a person by the contacts that would otherwise have reached one, for one call type at a time. Inside a call, both numbers are in your records. Before a call, the denominator is a guess, so measure deflection as a drop in calls per customer against a baseline or a holdout group.
The common version of the formula is deflected contacts divided by total contacts, times 100. That works until you look at what goes into the numerator. If "deflected" means "did not reach an agent", you are computing containment, and every unfinished or returning contact counts in your favor.
Calculate it per call type, never as one figure for the whole center. An order status line that finishes most of its calls and a billing line that finishes few of them average out to a number that describes neither, and a blended rate lets an easy call type hide a broken one. The same rule applies to the repeat contact rate in our metrics guide, which this calculation depends on.
The net deflection worksheet
Fill in one ledger per call type each month, starting from the calls the automation answered. Each step takes out contacts that left the queue on paper and came back as work. What is left at the bottom is the number to report, and the paragraph under the table turns it into agent hours.
| Step | What to subtract | Where it comes from | Example | Remaining |
|---|---|---|---|---|
| 1 | Start: calls the automation answered on this call type | The automation's call log | 6,000 | 6,000 |
| 2 | Calls transferred to your team | The transfer count or outcome tag | 900 | 5,100 |
| 3 | Calls that ended with a follow-up your team still owes | The follow-up outcome tag or callback list | 300 | 4,800 |
| 4 | Calls that ended without the request done | A random sample of 100 of the 4,800 transcripts: 4 were unfinished, so 4 percent of 4,800 | 192 | 4,608 |
| 5 | Callers who came back about the same thing within 7 days, in any channel | Caller number and call reason matched across your call and ticket records | 410 | 4,198 |
| 6 | Net deflected | What is left after step 5 | 4,198 |
On this example, containment reads 85 percent: 5,100 of 6,000 calls never reached a person. Net deflection is 70 percent: 4,198 of 6,000. At 4.5 handle minutes a call, the containment figure says 382.5 agent hours a month left the queue, and the net figure says 314.9. Your team still works the other 68 or so hours. They just sit under different call records.
The volumes, outcomes and handle time are illustrative, chosen to make the arithmetic visible, and should not be read as a benchmark. The order status line is the same one used in the sequencing worksheet for call center automation, so the two read together: that worksheet tells you which call type to automate first, and this one tells you what the automation took.
Rules that keep the ledger honest
- Fix the repeat window before you start and do not change it mid-quarter. A 24 hour window and a 7 day window give different answers from identical data.
- Count a return in any channel. An email or a chat about the same order is the same unfinished request.
- When a returning caller is answered and finished by the automation the second time, the first call still was not deflected. Count the first one in step 5.
- Draw the step 4 sample at random from all contained calls, not from the ones somebody flagged, and draw a fresh sample each month. Reading those calls is also how the fixes get found. Building categories over call transcripts helps once the sample is large enough to sort.
- Use the handle time your team had on this call type before automation. The center average misleads here: when the short calls leave, the average for the remaining queue rises, which the handle time guide works through on a full call mix.
From net deflection to the roster
The agent hours in the ledger are the most that staffing could save. Fewer calls in an interval change both the arrival rate and the average handle time of what remains, and shifts only shrink where the intervals allow it. The net counts per call type feed the deflection-adjusted staffing worksheet, which rebuilds the plan interval by interval.
Which call deflection strategies actually take work off your team?
A tactic takes work off your team only when it finishes the request without a person, and that can happen before the call or inside it. Callbacks and menu routing do something else. They move the wait or the caller, which helps service level and leaves the agent minutes where they were.
| Tactic | Where it acts | What it takes off your team | What it only moves | How to count it |
|---|---|---|---|---|
| A notice before the question, such as a delivery update, an outage message or a confirmation call the day before | Before the call | Calls asking what the notice already answered | Calls about the notice itself, if it is unclear | Calls on that reason per 1,000 customers, notified against a holdout group |
| Website or app self-service | Before the call | Calls for answers that are unambiguous and easy to find | New calls, when what the customer reads raises a question | Calls per customer by call reason, before and after, never one total |
| Self-service in the phone menu, such as reading back a balance or an order status | Inside the call | The lookup | Callers who give up in the menu and show up as contained | Menu completions minus menu hang ups minus repeat contacts |
| A callback instead of hold | Inside the call | Nothing | The wait, from on hold to off the phone | Not deflection. Count it under abandonment and speed of answer |
| Sending the caller to text or chat | Inside the call | Phone minutes | The contact, to a queue someone still staffs | Deflected only if the other channel finishes it without a person. Otherwise count the chat agent's minutes against it |
| An AI agent that takes the call | Inside the call | The whole conversation, for call types it finishes | Transfers and follow-ups, back to your team | The net deflection ledger above |
| Menu routing, touch-tone or conversational | Inside the call | Nothing | The caller, to the right queue faster | Not deflection. Count misroutes and transfers between teams |
Two rows are easy to report as deflection and are not: the callback and the routing menu. A callback keeps the agent minutes and spends the wait off the phone, which is why hold queue relief is measured in abandonment and speed of answer. A smarter menu sorts calls better and does not shrink the queue on its own, which the conversational IVR guide covers along with an audit for finding the menu options that should be resolved instead of routed.
How do you measure deflection that happens before the call?
Before the call, a deflected contact never happened, so there is nothing to count. Measure the call rate instead: calls on one call reason per 1,000 customers or orders, per week. Compare it with the same weeks before the change, or better, with a randomly chosen group of customers who did not get the change.
The comparison matters because self-service before the call can add calls as well as remove them. In a quasi-natural experiment at the call center of a large U.S. health insurer, customers' use of a web self-service portal led to a 14% increase in telephone calls. Calls for information that was unambiguous and easy to find on the portal fell by 29%, while calls about ambiguous information rose substantially (Kumar and Telang, "Does the Web Reduce Customer Service Cost? Empirical Evidence from a Call Center", Information Systems Research, 2012). Both effects sat inside one total, which is the argument for measuring by call reason.
A holdout test for a notice or a page you control runs like this:
- Pick one call reason and the thing that drives it: shipped orders for "where is my order", booked appointments for confirmations and reschedules.
- Hold back a random share of customers from the new notice or page, large enough that a difference will show in a few weeks.
- Track calls on that reason per 1,000 drivers for both groups over the same weeks, in every channel.
- The difference between the groups is the deflection. Multiply it by the driver volume for calls, and by the call type's handle time for agent hours.
- Check the calls that mention the notice itself. A confusing message shows up there first.
An illustrative read: if notified customers call 18 times per 1,000 orders about delivery and the holdout group calls 25 times, the notice deflects 7 calls per 1,000 orders, or 280 calls in a month with 40,000 orders. A count of notices sent or help page views cannot give you that figure, because neither one says anything about calls.
What our call records give you for the ledger
On Telvana, every call is recorded, transcribed, summarized and tagged with an outcome such as transferred, booked, voicemail or needs follow-up, and analytics show volume, talk time, pickup and transfer rates by agent and by day. Steps 1 to 3 of the ledger come straight from those tags. Transferred calls carry their transcript and summary, so you can read why each one was handed off, and the step 4 sample comes from the same transcripts. Step 5 runs on your own call and ticket records, where a caller's earlier and later contacts both appear. Each team or client program gets its own workspace, so a BPO can keep a separate ledger for every client.
Frequently asked questions
What does call deflection mean?
It means a customer's request was finished without reaching a person on your team, either because something answered it before they called or because a menu or an AI agent completed it during the call. In customer service the word refers to the work avoided. In telephony it means something narrower, covered in the last question below.
What is the difference between call deflection and call containment?
Containment counts calls that ended without reaching a person, whatever happened in them. Deflection should count only the ones that took work off your team, which means subtracting callers who gave up, follow-ups your team still owes and callers who came back. Many reports use the two words for the same number, so ask which one a report means.
What is a good call deflection rate?
There is no rate worth borrowing from another center, because the figure depends on the call type and on what gets subtracted. Compare each call type with its own month before automation and report net deflection beside containment. If the gap between the two grows, more of the contained calls are coming back.
Can call deflection increase call volume?
Yes. Self-service that leaves customers with new questions produces calls about those questions. In a field study at a large U.S. health insurer's call center, use of a web self-service portal led to 14% more telephone calls, even though calls for clear, easy-to-find information fell (Kumar and Telang, 2012). Measure by call reason so the two effects do not cancel out in one total.
Is a callback a form of call deflection?
No. A callback holds the caller's place and has a person call them back, so the agent minutes are the same. What changes is where the caller spends the wait, which shows up in abandonment and speed of answer. Count callbacks there and keep them out of the deflection figure.
What is call deflection in telephony?
In telephone networks, call deflection is a feature that redirects an incoming call to another number. In SIP it is handled as a redirect, such as a 302 Moved Temporarily response that tells the caller's side to retry at a new address (RFC 3261, section 21.3.3), and the redirected call can carry a Diversion header with the reason "deflection" (RFC 5806). It is unrelated to the customer service meaning.
Run the ledger on one call type
Pick the call type with the most agent hours, pull last month's records for it, and fill in the six steps by hand once. If you want to see how an AI agent would take that call type and what its records would show you, book a demo with the call type and its numbers. We will let you hear an agent handle one. The rest of the dashboard is in the metrics and operations guides.