We're deflecting 4,200 tickets monthly with support AI. Here's how to calculate it properly

Last month we reviewed a customer service AI deployment for an e-commerce company. Their dashboard displayed a prominent 68% AI resolution rate, and management was satisfied believing the bot was deflecting most tickets. We extracted three months of raw conversations and recalculated. The actual figure (tickets that never reached a human agent and where the customer's problem was genuinely resolved) came to 31%. That 37-percentage-point gap represented the bot responding without serving the customer. Some customers followed up with another question. Others called the support line. Some abandoned the interaction.

That's why ticket deflection requires clarity about what you're measuring before reporting any numbers.

Start with a definition you can actually cite

Ticket deflection is the percentage of conversations that would have become human-handled tickets but were resolved by the AI during the self-service stage and never entered the human queue. Key phrases: 'would have' and 'resolved.' Without either one, it doesn't count.

Deflection differs from containment, another metric often confused with it. Containment asks whether the conversation reached a human at all. If a customer abandoned the interaction, got frustrated by the bot, or switched channels, it still counts as 'successfully contained.' Vendor dashboards usually default to displaying containment because the numbers look better. Deflection requires more: no handoff to a human and the problem actually got closed. The two metrics often differ by 20 to 40 percentage points.

Four calculation methods; numbers can differ substantially

We only cite the third row, Net Deflection, externally. Only that one filters out false deflections. If a customer asks the bot, and 40 minutes later opens a new ticket asking the same thing, the first interaction doesn't count as deflected. We subtract it from the numerator.

How we got to 4,200 tickets

Before deploying AI, this e-commerce client's human team handled roughly 14,000 tickets monthly. After launching an AI copilot and routing three high-frequency intents (order status checks, return status updates, and invoice reissues) to an agentic workflow, human tickets dropped to 9,800. That's a 4,200-ticket difference.

Raw difference doesn't equal deflection. Seasonal swings are normal in e-commerce. To normalize, we look at human tickets per thousand sessions. Pre-deployment: 118 per 1,000 sessions. Post-deployment: 82 per 1,000 sessions. That's a 36-ticket improvement per thousand. Multiplied by that month's 116,000 sessions, it comes to roughly 4,180 tickets, which rounds to 4,200. This approach eliminates traffic variation, which is why it's reported this way.

Breaking down deflection by intent level is more informative. Order status deflects at 79%, returns at 61%, but payment disputes only at 12% because those require human judgment. This intent breakdown shows where to invest engineering next quarter, rather than chasing a single overall percentage.

Common practices that artificially inflate numbers

Three common practices artificially inflate deflection numbers, and we've encountered and helped unwind all of them. First: treating 'message read, no reply' as resolved. Most of these are abandonments. The correct approach uses a satisfaction survey or a 72-hour follow-up window to verify actual resolution. Second: including bot-initiated proactive conversations in the denominator, which dilutes the true help-seeking rate. The denominator should contain only customer-initiated conversations.

Third: cross-channel leakage. A customer gets 'deflected' online, then calls the support line. That phone call doesn't get attributed back, and the numbers become inflated. Connect phone, email, and chat through a unified customer ID to see the actual deflection picture.

Acceptance benchmarks you can use as your standard

A useful reference point: automatable transactional intents should achieve 55% or higher net deflection within three months of launch, which indicates healthy performance. Overall net deflection in the 25-35% range is normal. Anything over 50% should trigger a review of whether containment is being used instead of deflection in the calculation.

When implementing customer service AI, the value isn't in the 68% dashboard metric. It's in this method: deflection broken down to the intent level, false deflections filtered out, plus a breakdown updated weekly. What matters is whether the human queue is genuinely 4,200 tickets lighter three months later and customers haven't switched to calling instead.