Over the past five years leading customer support at Dose, I’ve worked closely with subscription customers across retention, fulfillment, cancellation, product questions, and post-purchase issues. One thing I’ve learned is that fewer tickets does not always mean customers are having a better experience. Sometimes it means we "genuinely fixed" the problem. Sometimes "automation handled" the request well. And sometimes the "ticket disappeared" while the customer friction was still there. That is why I think deflection is useful as an operational metric, but not as proof of a customer outcome.
This distinction matters even more now that AI and self-service are becoming a bigger part of support operations. If someone says AI deflected 40% or 50% of contacts, that can sound immediately positive: lower workload, faster support, lower cost. But I think the more important question is what exactly disappeared. Was it the problem, the ticket, or the customer signal? Those are not the same thing.
Questions customers should never have to ask
One pattern I have seen repeatedly is customers contacting support about something that really should have been fixed elsewhere in the business. A simple example was a winback campaign where existing subscribers received messaging intended for customers who were no longer subscribed. Naturally, some customers contacted support asking why they were getting a "winback" email when they were already active subscribers. We could have treated those as repetitive support tickets and automated the response, but that would only have made us more efficient at answering a question customers should never have had to ask. The real issue was campaign segmentation. Once that was corrected, the contacts disappeared because the underlying problem disappeared.
We saw something similar with delivery exceptions too. “Where is my order?” looks like a straightforward WISMO ticket and, in many cases, it actually is. If a package is moving normally and a customer simply wants an update, there is very little value in making them wait for a human agent. But the same question can also mean the shipment is stuck, a carrier exception has occurred, or the package needs intervention. If every one of those contacts is treated as a routine tracking request, support may resolve the question without recognizing the actual failure in the journey. Proactive exception communication and clearer resolution paths helped us reduce those contacts because customers were informed earlier and given support before they had to ask.
The safe zone for automation
But that does not mean every contact needs a human. There is a large safe zone where automation makes complete sense: low-risk, predictable, transactional requests where there is usually one clear answer or action. Normal order tracking, changing a shipping address, moving a billing date, or adjusting subscription frequency are good examples. If a customer can complete those actions accurately and immediately through self-service, that is often a better experience than waiting for an agent. The question I would ask is not simply, “Can AI answer this?” but rather, “How much uncertainty, judgment, or customer risk exists in this interaction?”
Automate the cancellation, not the reason
Cancellation is where this becomes more interesting. A customer may say, “I want to cancel because I have too much product.” A self-service system can handle that transaction very efficiently: identify the cancellation intent, present the cancellation option, and complete the request. But “too much product” can point to several different underlying issues. The shipment cadence may not match the customer’s actual consumption. The customer may not be taking the recommended amount consistently. They may not fully understand the dosage or how the product fits into their routine. Or the default replenishment timing itself may not reflect how customers really use the product.
Those are very different problems, and they may require action from different teams. Retention may want to test different cadence options. Lifecycle Marketing may need to improve education after purchase. The subscription team may want to review replenishment timing. Product may want to understand whether customers are actually forming the intended usage habit. The cancellation itself can be automated, but the reason should not disappear with it.
The same is true when we hear a customer saying “I didn’t know I subscribed.” That may appear to be a support cancellation issue, but if the same reason appears repeatedly, the business may need to look at checkout messaging, offer framing, subscription disclosure, post-purchase education, or renewal communication. Support sees the complaint first, but support does not necessarily own the root cause. This is one reason I think the support inbox should be treated as more than a queue. It is also a diagnostic system.
Deflection vs. resolution
There is another very important blind spot with deflection that I think gets overlooked: customers who contact support make friction visible, while customers who give up do not. Imagine a customer tries to skip an upcoming order through self-service and something does not work. One customer contacts support, which gives the business evidence that the journey failed. Another customer has the same experience, gets frustrated, and leaves without contacting anyone. From a support dashboard, that second customer may look like a success because there is no ticket, no agent work, and no escalation. But nothing was actually resolved.
That is why I separate deflection from resolution. Deflection tells us what happened to the contact. Resolution tells us what happened to the customer. Even resolution does not tell the whole story, because a customer may successfully complete a cancellation while the business loses the reason behind it. The transaction worked, but the learning disappeared.
What falling ticket volume hides
Top-line ticket volume can create a similar problem. Overall support contacts can decline while one particular customer journey gets worse. If leadership sees total volume falling, the first conclusion may be that the experience improved. But what if shipping contacts increased while other categories declined? What if basic transactional tickets went down because of automation, but repeat contacts increased in a more complex journey? The total number will not show that. I think CX teams need to look at what disappeared, what increased, and why, rather than relying only on the headline volume.
Four buckets for support contacts
In practice here’s what I think, I find it useful to think about support contacts in four buckets. The first is to eliminate the cause. If the contact exists because of avoidable friction such as poor campaign segmentation, missing delivery communication, or unclear subscription messaging, the goal should be to fix the underlying problem rather than build a faster support workflow around it. The second is to automate the transaction. Predictable, low-risk actions such as tracking, address changes, billing-date updates, and subscription frequency changes are strong candidates for self-service.
The third bucket is to preserve the signal. Some interactions can be automated while still capturing the reason behind them. Cancellation reasons, overstock, product adoption issues, dosage confusion, subscription misunderstanding, and recurring feedback all fall into this category. The customer may not need a human interaction, but the business still needs the information. The fourth bucket is to escalate the exception. Repeated delivery failures, unusual refunds, product safety concerns, emotionally sensitive complaints, and other high-risk situations should not be treated as opportunities to maximize deflection. The goal should be to recognize the exception quickly and get the customer to the right person.
What to ask about a 50% deflection rate
If someone told me, “Our AI deflected 50% of support contacts this month,” I would want to know what was inside that 50%. What types of contacts were deflected? How do we know those customers were actually resolved? What happened afterward? Did they remain subscribed, contact us again, move to another channel, or encounter the same issue later? I would also want to know what customer signal stopped reaching us because those interactions disappeared.
That is why I do not think deflection should stand alone. I prefer to separate the measurement into three questions: efficiency, customer outcome, and business learning. Did we reduce unnecessary workload? Did the customer actually complete what they came to do? Did we preserve the reason they needed help in the first place? Depending on the journey, that may mean looking at repeat contact, successful task completion, cancellation reasons, escalation rates, CSAT, customer effort, or downstream retention.
Fewer unnecessary conversations
AI and self-service should absolutely remove unnecessary work. Customers should not have to wait for an agent to complete something that can be handled accurately in seconds. But the goal should not be to make every customer conversation disappear. Some contacts exist because the business created friction and should be fixed upstream. Some are simple transactions and should be automated. Some contain valuable customer signals that need to be preserved. And some require judgment and should be escalated quickly.
For me, that is the difference between using automation to reduce support volume and using it to improve customer experience. The goal is not simply fewer conversations. It is fewer unnecessary conversations, better resolution, and better use of the customer signals that remain.
