Deflection vs Resolution: Formulas, a Worked Example, and the Metric That Hides Failure

Deflection is the metric most support tools lead with, and the easiest one to inflate. Resolution is harder to measure and much harder to fake. This guide gives both formulas, works through an example with the same data, and shows how to tell whether your deflection number describes anything real.

The short version

What deflection actually counts

Deflection is any action that prevents a customer request becoming a human-handled ticket. Help centre articles read instead of tickets filed. Bot answers accepted. In-product hints that answered a question.

It also counts a customer who opened chat, could not get an answer, and closed the window. That conversation is recorded as deflected, identically to one where the customer got exactly what they needed.

A customer counts as deflected when they:

• Got a clear answer and left happy. (Good.)
• Got a vague answer, gave up, and stewed. (Bad.)
• Could not find the “talk to a human” option. (Worse.)
• Closed the tab in frustration and churned a month later. (Worst.)

All four look identical in a deflection report. That is the whole problem.

Deflection took off because it mapped cleanly to cost and matched the technology of the time. Searchable knowledge bases and scripted chatbots were good at one thing: stopping a customer before they reached an agent. The metric described what the tools could do.

The formulas

Deflection rate formula

Deflection rate = (contacts that did not reach a human ÷ total contacts) × 100

For a self-service help centre, the common variant is:

Self-service deflection rate = (help centre sessions not followed by a ticket within N hours ÷ total help centre sessions) × 100

For an AI agent or chatbot, “AI deflection rate” or “containment rate” usually means conversations that ended without a handoff, divided by all AI conversations. Note what goes in the numerator: every conversation that ended without a human, including the ones where the customer simply left.

Resolution rate formula

Resolution rate = (contacts genuinely solved ÷ total contacts) × 100

Where “genuinely solved” means all three of:

1. The customer received a correct answer or completed action.
2. No handoff to a human was needed.
3. No repeat contact about the same issue, on any channel, within a set window (seven to fourteen days is typical).

When the AI does this end to end, it is usually called autonomous resolution rate. The exclusions are the point: abandoned chats, silent timeouts, and escalations are removed from the numerator.

A worked example

Here is one month of illustrative numbers for an AI chat agent. They are made up for the arithmetic, not taken from any customer.

• AI conversations started: 10,000
• Handed off to a human: 2,000
• Ended without a handoff: 8,000

On those numbers alone, the deflection rate is 8,000 ÷ 10,000 = 80%.

Now break down the 8,000 that ended without a human:

• Customer confirmed the answer or completed the action: 5,500
• Customer went silent or closed the window mid-conversation: 1,700
• Of the 5,500, contacted again about the same issue within fourteen days (by email, phone, or a new chat): 600
• Of the 1,700 who went silent, contacted again within fourteen days: 900

Genuine resolutions are the 5,500 confirmed, minus the 600 who came back: 4,900. Anyone who went silent is excluded whether or not they came back.

The resolution rate is 4,900 ÷ 10,000 = 49%.

Same month, same agent, same data. One number says 80%. The other says 49%. The 31-point gap is made of 800 conversations that went silent and never returned (possibly fine, possibly churn, unknowable), 900 customers who gave up and tried another channel, and 600 who were told something that did not hold.

If you were paying $2 per conversation handled by the AI, the bill would cover all 10,000 conversations. If you were paying per genuine resolution, it would cover 4,900. That difference is why the definition matters commercially as well as operationally.

The three ways deflection lies

Abandonment counts as success. The formula is avoided contacts divided by contact attempts. A customer giving up reduces human contacts and therefore improves the ratio.

Cross-channel leakage is invisible. A customer deflected from chat who emails an hour later has not been deflected, only moved. Chat deflection improves and email volume rises, and unless someone is matching those, both numbers look fine independently.

Definitions vary. Two vendors quoting AI deflection rates are frequently measuring different things. Neither is comparable to the other, and neither is comparable to yours.

The failure signature

Watch for this combination:

• Deflection rose sharply, often after a configuration change.
• Satisfaction did not move, or fell.
• Volume on another channel rose within a fortnight.
• Repeat contact rate is not being tracked at all.

That pattern means customers are giving up. The dashboard is recording a triumph.

The most common version: someone made the human option harder to find. Deflection improves immediately, and nothing else does. This can happen without anyone deciding to game the number. When a team is told to raise deflection, the cheapest path is often to make leaving harder, not to make answers better.

Why resolution is the healthier metric

A resolution measures presence, not absence. Something good happened, and you can point to it. That changes behaviour:

• For your team, the goal becomes “did we actually help”, not “did we avoid a ticket”.
• For your customers, the path to a human stays open. A handoff is not a failure. Some questions need a person, and pretending otherwise frustrates people.
• For your tooling, the pressure points the right way. Better answers grounded in your real help content beat clever ways to make people stop asking.

What to measure alongside it

Genuine resolution rate. As defined above. Write the definition down before measuring anything, and exclude abandonment and handoffs explicitly.

Reopen rate across channels. Match by customer identity and topic, not by ticket thread. A resolved chat followed by an email about the same issue is a reopen even though the two systems treat them as unrelated.

Satisfaction, segmented. By topic, by channel, and by whether the issue was resolved or escalated. Survey the AI path too: plenty of teams only survey human-handled tickets, which leaves a blind spot the size of the whole automation.

Those three together cannot be gamed by making support harder to reach.

How to audit your current numbers

Five questions, and you can answer all of them this week:

1. How is our deflection calculated? Ask your vendor whether abandonment and timeouts count. If they do, the number is not measuring what you assumed.
2. What is our repeat contact rate? If nobody tracks it, that is the first thing to fix. It is the honesty check on every other metric.
3. Did volume move to another channel? Compare chat and email volume over the period deflection improved.
4. What did satisfaction do? Segment it. If it fell in the topics where deflection rose most, you have your answer.
5. How many steps to a human? Open your own chat as a customer and count. If it takes effort, your deflection number is partly a measure of friction.

Why this matters commercially

Teams act on these numbers. A support leader reporting rising deflection gets budget and credibility. If the underlying reality is customers abandoning, the consequence arrives two quarters later as churn, unattributed, in a different report.

It also shapes your bill. If a vendor bills per conversation, they earn identically whether the customer was helped or gave up. If they report deflection as the headline outcome, the metric they are optimising counts abandonment as a win. Neither is an accusation of bad faith. Both are incentive structures, and incentive structures produce outcomes regardless of intent. Some vendors bill per “automated resolution” with their own definitions, so read those definitions closely too.

Fidiora bills only on genuine resolutions, at $0.69 each, dropping to $0.45 for every resolution above 1,000 a month, and does not report or price on deflection. Abandoned chats, silent timeouts, and handoffs to a human are never billed. The full resolution definition is published, and the pricing page shows the spend cap. We would make the same argument if we sold nothing: the metric you pay on is the metric you will get more of.

The one change worth making this week

Put repeat contact rate next to whatever automation metric you currently report.

That single addition makes the other number honest, and it costs nothing except the effort of matching contacts by customer and topic within a fortnight window. Every team that does this learns something in the first month. For the wider set of numbers worth tracking, see the support metrics that matter.