Stop reporting deflection to your board

Deflection records a customer who gave up as a success. If it is the headline figure in your AI report, your board is looking at the wrong measure.

Deflection is the headline figure in most AI support reports, yet it measures only whether a customer reached a human, not whether the issue was resolved.

If deflection is the primary measure presented to the board, the reported return on AI investment may be overstated. The gap typically becomes visible only later, when repeat contacts, costs or complaints rise.

What the numbers say

A simple example shows the difference. A customer asks the AI agent about a refund, and the agent links to the refund policy. The customer closes the chat with the issue unresolved and emails two days later, less satisfied than at first contact. Deflection records the chat as a success and the email as a new ticket. Resolution records the whole interaction as a single failure, which is what it was.

Gartner’s research shows the scale of the difference. In a survey of 5,728 customers, only 14% of service issues were fully resolved in self-service, and only 36% of issues that customers described as very simple. A customer in the remaining 86% who gave up without contacting the company again would typically be recorded as deflected.

Vendor research suggests the gap is large, although I could not confirm these figures at source. Aissist, an AI support software provider, estimates that deflection runs 20 to 40 percentage points higher than true resolution on the same deployment. HappySupport, a help-centre software provider, estimates that true deflection is 30 to 40% lower than the headline figure once returning tickets are counted; on a headline of 50%, that would mean a true figure of 30 to 35%. HappySupport credits Forrester and Aberdeen for this estimate but does not link to either. It also reports that vendor presentations quote 30 to 60%, while independent surveys find 10 to 25%.

Boards are likely to scrutinise these figures more closely. Gartner found that service and support leaders allocated a median 12% of their 2025 budget to AI, more than any other function, yet only 24% of the leaders surveyed could demonstrate a positive financial return from AI.

Why deflection persists

Deflection is easy to measure, and vendors often define resolution in a way that matches it. Intercom reports that its Fin agent resolves 76% of conversations on average, counting a conversation as resolved when Fin handles it without a human. This definition measures the absence of human involvement, which is closer to deflection than to customer-confirmed resolution. Intercom is now moving its pricing metric from resolutions to “outcomes”, which also include actions such as gathering context and handing over to an agent. Neither measure confirms that the customer’s issue was solved.

Cost figures share the same weakness. Aissist’s analysis puts repeat contact at 2.3 contacts per issue, which would make the true cost of resolving an issue more than double the cost per contact. Aissist does not publish the data behind this figure, but the principle holds: a business case built on cost per contact is likely to understate the true cost of resolution.

Metrics also shape behaviour, not always in the intended direction. One company I worked at measured cases closed on first contact. Every regional team struggled to meet the target except one, which reported 100%. That team had been instructed by its leadership to close every case on first contact and open a new one if the customer needed further help. The dashboard showed the target met, while customers received a worse service. Deflection creates the same risk when it becomes the measure of success.

What to report instead

Lead with resolution rate. Count an issue as resolved only when the customer confirms it, or does not contact you again about the same issue within a set period. Choose the period, for example seven days, and keep it consistent every month so that the trend is meaningful.

Report repeat contact rate alongside it. If customers continue to return about the same issue, the AI agent is closing conversations without resolving them.

Report cost per issue, not cost per contact. This figure is less favourable, but it is the one finance teams are most likely to calculate independently.

Split CSAT by who handled the conversation. Aissist estimates that AI-handled interactions score 5 to 10 points lower than those handled by people in the same team, although it does not publish the supporting data. Measure your own gap. It is preferable for the board to hear about it from the service team than to discover it later.

Review the handover process. Confirm that agents receive the full conversation history when a customer is transferred. Zendesk’s 2026 research found that 74% of consumers are frustrated when required to repeat information, and a poor handover requires them to do so a second time.

Where to start

The main challenge is managing the change with leadership rather than the calculation itself. If the board has been seeing a deflection figure of 50%, a resolution rate of 30% will appear to show a sudden decline in performance.

Avoid replacing the figure without notice. Report both measures side by side for a quarter, and explain the difference once, clearly, using a real example such as the refund conversation above. Then remove deflection from the headline and retain it in an appendix for reference.

Reset expectations at the same time. HappySupport estimates median AI deflection at 22%, and 10 to 15% for a typical B2B SaaS team in its first year. Even as approximate guides, these are well below a 50% target. If the original business case assumed a figure of that order, raise it now rather than at the next budget round.

Put it into practice: Define “resolved” before you report it