The vendor's quarterly report lands in your inbox with a number in a big green circle: 82% deflection. Two tabs over, your phone system tells a different story. Tuesday's caller ID list looks suspiciously like Monday's. The same account number shows up in the chat log at 9:14 PM and in the call queue at 8:03 the next morning. Somewhere between those two tabs is the question the green circle never answers: did anyone actually get what they came for?
Deflection is the easiest number in customer support to make go up, which is exactly why you should distrust it. This piece is about what deflection actually counts, what resolution means when you define it strictly, and how to audit the difference on your own lines in about an hour a week.
What deflection actually counts
Deflection rate is the percentage of conversations that touched your bot or IVR and never reached a human. That's the whole definition. Nothing in it asks whether the customer's problem got solved.
Run three visitors through the same widget. The first asks about your refund window, gets an accurate answer, and leaves satisfied. The second asks the same question, gets three buttons that don't match, types "agent" twice, gives up, and closes the tab. The third gives up and buys from your competitor. All three count as deflected. The metric scores them identically because it only measures one thing: your payroll was not disturbed.
Deflection counts silence as success, and silence is where abandoned customers hide. The gap between the two numbers is not small. Gartner surveyed 5,728 customers in late 2023 and found that while 73% used self-service somewhere in their journey, only 14% of service issues were fully resolved there. Even for issues customers themselves called "very simple," the full self-service resolution rate was 36%. Any vendor quoting an 80% deflection rate against a backdrop like that is describing how many people their software got rid of, not how many it helped.
What resolution means when you define it strictly
Resolution needs a definition tight enough that you can't flatter yourself with it. Here's the one we'd argue for: a conversation is resolved when it ends in one of three states - the customer got a verified, accurate answer to what they asked; an action was completed in your systems (a booking made, a payment taken, a ticket filed, an order located); or the conversation was handed to a human who received the full context and continued it without making the customer start over.
Everything else is not resolution. A customer who stopped replying is not resolved. A "did that answer your question?" clicked out of fatigue is not resolved. A handoff that dumps the customer into a general queue to re-explain from the beginning is barely resolution at all - it's deflection with extra steps.
The honest test of resolution is whether the customer comes back about the same issue. That's why the strongest companion metric is recontact rate: of the conversations you scored as resolved, how many produced another contact about the same issue within seven days? A high deflection rate with a high recontact rate isn't efficiency. It's the same work done twice, with an annoyed customer in the middle - and it's invisible if deflection is the only number on your dashboard.
The numbers that keep you honest
If you're running an AI agent on your phones or your website - ours or anyone's - these are the four numbers worth a weekly look. None of them come from a vendor's brochure; all of them come from your own logs.
- Verified resolution rate. Not the bot's self-reported score - the percentage of conversations that ended in an accurate answer, a completed action, or a context-preserving handoff, verified by sampling transcripts yourself.
- Seven-day recontact rate. Of the conversations marked resolved, how many customers came back about the same issue within a week? This is the metric that catches inflated resolution claims.
- Abandonment inside the bot. How many conversations simply stopped - no answer delivered, no action taken, no handoff? Every one of these was counted as a win under deflection. Under resolution, they're your defect log.
- Handoff quality. When conversations escalated, did the human receive the transcript and the gathered details, or did the customer repeat everything? Count the repeats. Each one is a customer being told their time is worthless.
Deflection can only ever measure a cost you avoided; resolution measures work that actually got done. That difference shows up in revenue, not just support tickets. The 9 PM pre-sales question that gets a real answer becomes an order. The one that gets deflected becomes your competitor's order. A cost-avoidance metric will never show you that loss, because the customer left quietly.
Run your own numbers
You don't need an analytics suite to find out which side of this line your current setup falls on. You need one hour and your own transcripts.
Pull 50 recent bot conversations - chat, phone, or both. Read each one and put it in one of four piles: finished (answer or action, and it held up), abandoned (customer went silent before getting anything), handed off warm (human got the context), handed off cold (customer started over). Then take the finished pile and search your logs for the same customer in the following seven days. Every match moves that conversation out of the finished pile.
Fifty transcripts is enough to see the shape of it. Your true resolution rate is the finished pile after recontacts, divided by fifty. Compare that to whatever deflection number you've been reporting upstairs. The gap between the two is the work your customers are doing to route around your automation - calling back, emailing, or leaving.
Do the same exercise on the phone side, because the same illusion lives there under different names. "Contained in IVR" is deflection wearing a headset: a caller who navigates four menus and hangs up counts as contained. A voice agent should be held to the identical standard as a chat widget - the caller got the answer, the action happened, or a human picked up the thread with full context. And if your chat agent and phone line run on separate systems that can't see each other, your recontact rate is hiding in the seam between them: the 9 PM chat abandoner who calls at 8 AM looks like two clean interactions instead of one failure.
Why pricing shapes the metric you get
Here's an uncomfortable pattern worth watching for: how a vendor charges you tends to predict which number they'll show you. Per-conversation pricing rewards short, shallow interactions - the faster the bot ends the chat, the better the margin, and deflection is the metric that makes that look like a win. Opaque bundles hide the mechanics entirely, so the big green circle is all you ever see.
The alternative isn't a magic pricing model - it's transparency plus the raw material to check the numbers yourself. Verlingo charges a flat monthly plan with a shared allowance, and beyond it a published per-minute rate for voice and a flat rate per chat reply - every rate printed on the pricing page, no negotiation, no bundle to decode. Per-minute pricing has a useful side effect for this audit: there's no incentive to inflate a resolution count, because we don't bill on one. What you're paying for is time, so the number that matters to both sides is whether that time produced an answer, an action, or a warm handoff - which is exactly what the four-pile exercise measures.
Two design choices make the audit easier to run, too. Every plan includes call recordings, transcripts, and summaries, so the raw material for your 50-conversation sample is already sitting in your dashboard. And voice and chat run as the same agent on one system - same knowledge base, same tools, same logs - so the 9 PM chat abandoner who calls at 8 AM shows up as one customer with one unresolved issue, not two clean interactions. When a conversation does escalate, the human gets the transcript and full context, and the moment a caller connects to your team the AI rate stops - transferred time bills at a nominal line rate, so there's no meter running that rewards keeping the bot talking.
Start small, measure it
You don't have to re-platform anything to act on this. Start with the transcript audit above on whatever you run today - it costs an hour and tells you your real number. If the gap between deflection and resolution turns out to be wide, run a contained experiment: put an agent on one line or one page, let it run for a month, and score it with the same four metrics. Same transcripts, same recontact check, same four piles.
That's the standard we ask to be measured by. Setup takes minutes, the trial is free with usage included, and the first month gives you a clean before-and-after: your verified resolution rate, your recontact rate, your warm-handoff rate - pulled from transcripts we hand you by default, at per-minute rates published on the page. If the numbers don't beat what you're running now, you'll have the transcripts to prove it. Either way, you'll never look at a big green deflection circle the same way again.
