Dr. Jonah Tebaa keeps running into the same dashboard slide across companies that have nothing else in common. Deflection rate: 80 percent. To most executives in the room, that number reads as a support function that finally solved its cost problem. To Dr. Tebaa, it reads as a metric that never once asked whether the customer got what they came for. What he notes is missing from the same slide is the number that would settle the question: the churn and renewal line underneath, which he says has typically not moved at all.
A Number Built to Avoid a Question
Deflection rate measures one thing only: whether a human agent got involved in a given contact. It says nothing about whether the underlying issue was resolved. In his work advising companies on AI rollouts, Dr. Tebaa walks clients through a composite scenario that recurs often enough to be a pattern rather than an anecdote: a support operation fielding 12,000 contacts a month, at roughly $6 a contact when a human handles it end to end. Once AI takes the front line and deflects 80 percent of volume at closer to $0.80 a contact, the monthly bill drops from about $72,000 to roughly $22,000, savings north of $49,000, and a number every board wants to see again next quarter.
Public service design reached the same conclusion years ago and standardised the opposite metric. The UK Government's service manual instructs teams to report a service's completion rate, the proportion of transactions users actually finish, and states explicitly that the figure "includes transactions where the user receives support from someone to use the digital service". Human involvement is not scored as a failure there; an unfinished task is, which is exactly the inversion Dr. Tebaa asks support leaders to make.
The Cost That Shows Up Later
The trouble, in Dr. Tebaa's framing, is what the deflection number does not track. In the same illustrative operation, a meaningful share of resolved contacts, he estimates around a fifth, come back within a week, reopening the same problem on a different channel. Each of those now needs a human agent to close out properly, quietly eating a large share of the original savings. Handoffs that do reach an agent after a failed AI attempt also take longer to resolve, because the agent has to untangle what the AI already tried before solving the actual problem. None of that shows up on the deflection dashboard, because by the time it happens, the original contact has already been logged as a win.
Three Tiers, Not One Metric
Dr. Tebaa's response is not to abandon deflection rate but to stop applying it uniformly. He separates support contacts into three tiers: purely deterministic issues, like password resets and order status, where deflection and resolution genuinely track together; judgment-assisted issues, like billing disputes and account-specific troubleshooting, where resolution rate and repeat-contact rate matter far more than deflection; and relationship-sensitive issues, like churn risk and renewal-adjacent conversations, where he argues deflecting to protect a dashboard number is often the costliest decision a support team can make, since the damage lands on future revenue rather than this month's support budget.
A Different Checklist for Renewal Decisions
Before signing off on renewing or expanding an AI support deployment, Dr. Tebaa asks clients to produce resolution-adjusted cost per contact, seven-day repeat-contact rate, human cleanup time on handoffs measured against an all-human baseline, the CSAT gap between AI-only and handoff contacts, and the renewal or repeat-purchase rate among customers whose last contact was AI-only. None of these figures require new tooling; most support platforms already log the raw events, they simply are not surfaced next to the deflection percentage that gets presented upward. His broader argument is as much about sequencing as measurement: automate the deterministic tier first, prove resolution-adjusted cost per contact holds up there, and only then expand into contacts where judgment, and customer patience, carry more weight than deflection ever could.
The distinction matters most for companies currently reporting AI support wins purely through deflection percentages. Dr. Tebaa's point is not that the technology underperforms; in the deterministic tier, it frequently outperforms human agents on both cost and speed. His point is that a single blended metric hides which tier is actually driving the number, and that hidden tier is usually the one costing the company customers. For a company sitting at 50 to 500 employees running AI in support today, that reframing is the difference between a metric that flatters the board and one that predicts renewal.
