Research-backed calculator guide

Cost per Resolved Support Ticket with AI: From Headline Rate to All-In Outcome Cost

Calculate AI cost per resolved support ticket by separating attempts, verified resolutions, handoffs, fixed fees, and human follow-up from the headline vendor rate.

Why headline AI rates mislead

A per-resolution quote can be useful because it ties some spend to an outcome. It becomes misleading when it is compared with a fully loaded human cost or when “resolution” means something different between vendors. Some plans add seats, minimums, platform charges, usage allowances, voice fees, or implementation services. A team can also pay for an AI step and then pay a human to finish the same customer problem.

The measurement problem comes first. “Ticket closed” is an administrative state. “Resolved issue” is an outcome. If a customer asks a follow-up question, repeats the same problem through another channel, or reopens the case, the first interaction should not remain in the successful-resolution denominator. Zendesk’s AI resolution guidance similarly says that fully resolved issues should be separated from deflection, partial answers, escalation, and repeat contact.^[1]

Pricing semantics matter just as much. Intercom’s Fin documentation, checked on 2026-11-02, lists $0.99 per outcome for a resolution and for some configured procedure handoffs, while default escalations and unsuccessful attempts are not billed under the stated rules.^[3] Zendesk says AI agents are included in Suite and Support plans and that automated resolutions are the unit used for AI-agent billing, with successful outcomes not escalated to a human.^[2] These are vendor-specific definitions, not universal accounting rules. Put each vendor’s event type into your own ledger before comparing rates.

The cost-per-resolved-ticket formula

Use a fixed period, normally one calendar month. Keep every numerator and denominator on the same scope. For a support channel, queue, product, or workflow, calculate:

All-in CPR = (AI outcome fees + AI platform allocation + model and tool usage + QA/monitoring + implementation allocation + human follow-up cost + other directly attributable overhead) ÷ verified final resolutions.

There are two useful denominators:

1. AI-only verified CPR. Divide the cost assigned to successful AI resolutions by the count of issues that AI resolved without a human finish. This is useful for judging the agent’s own unit economics.
2. AI-assisted final CPR. Add the cost of escalated conversations that humans finish, then divide by all final resolutions in the AI-touched cohort. This is the number to use when an AI attempt can create or avoid downstream human work.

Report both when possible. The first shows what a clean AI outcome costs. The second shows what customers and the support budget actually experienced.

Define “resolved” before collecting prices

Write an operational rule that your support, finance, and analytics teams can apply to the same event stream. A practical definition is: the customer’s requested task was completed or the question was answered accurately; no human intervention was required for an AI-only resolution; no repeat contact for the same issue occurs during a 24–48 hour recontact window; and the event is not a duplicate, spam, abandoned conversation, or failed workflow.

The window should reflect the work. A password reset may need 24 hours. A billing dispute or delivery problem may need several days. Do not silently change the window between vendors because it makes one resolution rate look better. Record the rule in the metric definition and version it when the business changes.

Also distinguish a handoff from a failure. A handoff can be the right customer outcome, especially for a sensitive or complex case. It is not an AI-only resolution. Some vendors charge for a configured handoff and not for a default escalation, so your ledger needs a field for the handoff reason and billable status rather than one generic “escalated” flag.^[3]

Worked cost example

Assume a support team evaluates one month of AI-assisted chat and email. These are modeling assumptions, not a vendor quote or a forecast:

• 6,000 conversations enter the AI workflow.
• 4,200 are verified AI-only resolutions.
• 1,800 are handed to a human and eventually resolved.
• The modeled outcome charge is $0.99 for each of the 4,200 AI resolutions.
• Ten human seats are allocated to the workflow at $55 per seat per month, or $550. This is a simple allocation assumption; a real contract may use a different plan or allowance. Zendesk’s public pricing page showed a Suite Team price of $55 per agent per month when checked on 2026-11-02, while its AI billing details are separate from that base plan.^[2]
• Each AI attempt uses 18,000 input tokens and 1,800 output tokens. At 6,000 attempts, that is 108 million input tokens and 10.8 million output tokens.
• For a transparent model-only estimate, use OpenAI’s displayed GPT-6 Luna rates of $0.10 per million input tokens and $0.50 per million output tokens, checked on 2026-11-02. The resulting estimate is $10.80 + $5.40 = $16.20.^[5]
• QA, monitoring, and reporting are allocated at $800 per month.
• A $12,000 implementation is amortized over 12 months, or $1,000 per month.
• A human finishes each of the 1,800 handoffs at an estimated $12 per final resolved issue. This is an internal labor assumption, not a benchmark or vendor quote.

The AI-only view is different. Exclude the human follow-up row and divide $6,524.20 by 4,200 AI-only resolutions: $1.55 per verified AI resolution. Both figures are valid for different decisions. Use $1.55 to evaluate the cost of an autonomous outcome. Use $4.69 to evaluate the AI-touched cohort after the downstream human work is included.

For comparison, a hypothetical all-human baseline at the same estimated $12 per resolved issue would be $72,000 for 6,000 resolutions. The modeled difference is $43,875.80, or about 60.9%. That is an illustration of the assumptions above, not a promised saving. Change the handoff rate, human labor rate, platform fee, recontact rule, or token profile and the result changes immediately.

For a broader monthly spend model, use the AI Agent Cost Calculator | Model Full Monthly Spend. The calculator can hold shared infrastructure and usage assumptions, while this article supplies the resolution-ledger denominator.

Build a ledger that finance can audit

The minimum event-level fields are conversation ID, customer or account ID, channel, workflow version, AI attempt timestamp, input and output usage, tool calls, outcome type, resolution timestamp, escalation reason, human minutes, recontact flag, duplicate flag, and billing status. Add vendor invoice row or usage-export IDs wherever possible. That lets finance reconcile the vendor’s count to the product analytics count instead of treating a dashboard total as proof.

At month end, reconcile in this order:

1. Start with every conversation that entered the AI workflow.
2. Remove spam, duplicates, test traffic, and events outside the reporting period.
3. Label the terminal outcome: verified AI resolution, human-resolved handoff, unresolved, abandoned, reopened, or other.
4. Apply the vendor’s billing rule to the terminal outcome. Do not infer it from the word “resolved” in a ticketing system.
5. Add direct human minutes for handoffs and repeat contacts.
6. Allocate fixed platform, QA, monitoring, security, and implementation costs with the same allocation key each month.
7. Divide by final verified resolutions, and publish the resolution count beside the dollar figure.

A model that cannot show its denominator is not yet a cost-per-resolution model. It is a rate-card calculation.

Compare pricing models on the same outcome basis

Outcome pricing is easiest to model when the vendor bills only for successful outcomes. Even then, ask what counts as success, what happens after a later recontact, and whether a configured handoff is billable. Intercom’s current documentation says it charges once per conversation for a counted outcome, and it distinguishes billable procedure handoffs from non-billable default escalations.^[3]

Conversation pricing is simpler to forecast but does not automatically mean one conversation equals one solved issue. Salesforce’s Agentforce pricing page, checked on 2026-11-02, lists $2 per conversation and separately lists Flex Credits at $500 per 100,000 credits.^[4] A conversation can contain multiple actions, a partial answer, or a handoff. If your system resolves 4,200 of 6,000 AI conversations, a $2 conversation rate would be $12,000 before fixed fees and human follow-up. Its effective cost per final resolved issue is therefore $12,000 divided by 6,000 only if every remaining conversation is resolved by humans and you include those human costs separately. The unit price alone does not establish the outcome cost.

Token billing is another layer. OpenAI’s published API page lists separate input, cached-input, and output rates by model.^[5] Multiply actual usage by the correct meter, including tool or retrieval charges where applicable. Do not use a generic “tokens per ticket” estimate in a finance report without labeling it as an estimate and replacing it with usage exports when available.

What to exclude and what to disclose

Exclude taxes, unrelated corporate overhead, and product-development work unless you explicitly allocate them to the support workflow. Do include costs that move with the outcome or are necessary to operate it: agent labor for handoffs, QA sampling, model and retrieval usage, monitoring, support tooling, integration maintenance, and a defensible share of implementation. State whether prices are USD, whether vendor prices are monthly or annualized, and whether the example uses list prices or a negotiated contract.

Do not mix a per-contact human benchmark with an AI per-resolution quote. If the human number is cost per contact, convert it to cost per final issue using the team’s contact-to-issue ratio. If a customer needs 1.4 contacts per issue, $8 per contact is not $8 per resolved issue; the comparable issue cost is $11.20 before other overhead. Keep channel, complexity, geography, and recontact windows aligned.

Finally, publish a range. A sensitivity table with 50%, 70%, and 90% AI resolution rates is more useful than a single decimal. The AI agent budget guardrail and P95 calculator can help test volume and tail assumptions, while the AI API invoice reconciliation calculator can compare modeled usage with actual invoice rows.

FAQ

What is the formula for AI cost per resolved support ticket?

Add AI outcome fees, platform allocation, model and tool usage, QA, monitoring, implementation amortization, and human follow-up for handoffs. Divide that total by final verified resolutions. If you want the autonomous unit cost only, exclude human finishes and divide by verified AI-only resolutions, but label that denominator clearly.

Should escalated tickets count as AI resolutions?

No. An escalated ticket is not an AI-only resolution. It can still be a final resolved issue for the broader AI-assisted cohort after a human completes it. Track both states so the metric does not hide downstream labor or treat a handoff as autonomous success.

Is a $0.99 per-resolution price the all-in AI cost?

Usually not. It is one vendor-metered component. You may also pay for seats, minimums, platform access, model usage, retrieval, monitoring, implementation, and human work after handoff. Intercom’s $0.99 outcome rate is an example of a vendor price and definition, not a universal market rate.^[3]

How long should the recontact window be?

Use a window that matches the support problem and keep it stable for comparison. A 24–48 hour window is a workable starting assumption for many conversational workflows, but billing, delivery, security, or regulated issues may need longer. Record the choice in the metric definition and test whether repeat contacts are being missed across channels.

What is the difference between resolution rate and deflection rate?

Resolution rate asks whether the issue was fully solved. Deflection rate usually asks whether the customer avoided a human queue or agent touch. A customer who abandons a chat after an incomplete answer may be deflected but not resolved. For cost accounting, use verified final outcomes and treat deflection as a separate diagnostic metric.

Which internal cost should I add for a human handoff?

Use loaded labor cost per resolved handoff, not only wage. Multiply productive human minutes by an hourly rate that includes benefits and relevant support overhead. If the handoff also creates QA, recontact, or specialist-review work, include those costs or disclose that the result is understated.