AI customer support ROI for a high-volume team is the loaded cost of the tickets an AI agent resolves end to end, minus what you pay per resolution and in fixed fees, and it only becomes cash where the freed hours stop a hire. In the worked model below, a 20,000-ticket-a-month team on published per-resolution pricing saves about 38% of monthly support cost at a 50% resolution share, and about 26% once the tickets left for people are costed as the harder ones. Seven inputs decide which of those numbers you get, and you can check every one of them.

Key takeaways

• Start from payroll, not a vendor slide. The BLS median wage of $21.53 an hour becomes about $55 per productive hour once benefits, 30% shrinkage and a 1.25 overhead factor are added.

• Budget on a 50% resolution share, with 35% and 65% as the downside and upside. Treat vendor averages as upside until your own tickets confirm them.

• Convert every quote to an effective price per resolution. $2 per conversation at a 50% resolution share is $4 per resolution.

• Stress the residual. Costing the tickets humans keep at a 25% premium cuts the 50% scenario's saving from about $64,700 to about $43,500 a month.

• Count savings as avoided hires unless you have a plan to reduce headcount. Gartner expects half of the organizations that planned big service workforce cuts to abandon those plans by 2027.

Which tools price AI support per resolution?

Several AI agent vendors now publish an outcome price, which is what makes an ROI model checkable before you sign. Lorikeet charges $0.99 or $0.90 per chat, email or SMS resolution on its Start and Scale plans and does not charge for escalations to a person or unresolved tickets. Alternatives with published prices:

• Fin (now part of Salesforce) lists $0.99 per outcome with a 50-outcome monthly minimum and no setup or platform fees.

• Zendesk includes AI agents in every Suite and Support plan and bills per automated resolution alongside per-agent seats.

• Salesforce Agentforce sells Conversations at $2 each, Flex Credits, or per-user licences.

How we built and checked this model

We built the model for teams handling 5,000 to 20,000 or more tickets a month across digital channels and voice, because that is where per-ticket savings start to outweigh fixed fees. Every labour input comes from a public statistical source, every price comes from the vendor's own pricing page as published on 30 October 2026, and every assumption is labelled as one so you can replace it. Customer results are quoted only as printed on the customer's published story. Lorikeet publishes this guide and is one of the tools listed. We have tried to describe every product, including our own, using only what each vendor publishes, and we link to the source for each claim. If you would rather run your own numbers first, Lorikeet's free CX ROI calculator compares pay-per-ticket and pay-per-resolution pricing on your volume.

1. Loaded cost per hour: build it from payroll data

The loaded cost of a support hour is wage plus benefits plus the paid time agents spend away from the queue plus overhead. The Bureau of Labor Statistics puts the median hourly wage for customer service representatives at $21.53 in May 2025, across about 2.7 million jobs. In the BLS Employer Costs for Employee Compensation release for June 2026, wages were 70.0% of private-industry employer costs and benefits the remaining 30.0%, so dividing the wage by 0.70 gives about $30.76 per paid hour.

Paid hours are not productive hours. Breaks, training, meetings and absence make up shrinkage, and Call Centre Helper reports that most contact centre professionals put it between 30 and 35%. Divide by 0.70 again and a productive hour costs about $43.94. Then add what compensation misses: helpdesk and telephony seats, team leads, QA, recruiting and ramp time. That multiplier is yours to take from the P&L; this model uses 1.25 as a placeholder, which gives $54.92 per productive hour.

The most common error is comparing an AI price with the bare wage. On these inputs the bare wage understates the true hourly cost by more than half.

2. Tickets per productive hour: the denominator that moves most

Cost per ticket is the loaded hourly cost divided by tickets handled per productive hour, and the denominator varies more between teams than any other input. Pull it from your workforce data by channel. The table shows the shape at $54.92 an hour.

One external reference point for voice and chat: Klarna reported that before its AI assistant a customer errand took 11 minutes to resolve, roughly five an hour. If your handle times are longer because tickets involve several systems, your cost per ticket is higher and so is the return on resolving those tickets end to end. There is a longer treatment of the components in customer service cost per ticket.

3. Resolution share: model 35, 50 and 65 percent

Resolution share is the percentage of all tickets the AI closes with no human touch and no reopen inside a window you choose, and it is the input to treat most sceptically. Published numbers vary widely because they count different things. Klarna's assistant handled two-thirds of customer service chats in its first month and cut repeat inquiries by 25%, by its own account. The Salesforce release on closing the Fin acquisition cites an average resolution rate of 76%, a vendor figure. Looking forward, Gartner predicts agentic AI will autonomously resolve 80% of common customer service issues by 2029.

None of those is your month-one number. A deflected ticket, where the customer read an answer and left, is not a resolved one, where the refund was issued or the account updated; the difference is set out in resolution rate vs deflection rate. Published customer results show how wide the range is even for one product: Chexy moved its AI agent's share of inbound from about 24% to about 72% in one week, and Hnry automated roughly 58% across its tax-season period and around 70% in the peak week. Three scenarios keep the model honest. Compare published averages in AI resolution rate benchmarks.

4. Effective price per resolution: convert every quote

The effective price per resolution is what you pay the vendor in a month divided by the tickets it actually resolved, and it is the only unit that lets you compare quotes. Pricing units differ. Fin bills $0.99 per outcome, and its outcomes include procedure handoffs and disqualifications as well as resolutions, though conversations simply passed to your team without an outcome are free. Zendesk bills per automated resolution; since 18 May 2026 its resolution tiers exclude assisted escalations and contained resolutions from the allowance, and an LLM verifies the conversation before it counts as a verified resolution. Agentforce Conversations are billed at $2 each, per conversation rather than per resolution.

The conversion is simple: price per conversation divided by resolution share. At a 50% share, $2 per conversation is $4 per resolution. Then add the fixed fees in step 5. Ask every vendor who defines a resolution and whether you can mark a bad one as unbillable, because the party that defines the metric controls your ROI.

5. Fixed fees and QA: the lines that set break-even

Fixed fees decide how many resolutions you need before the AI pays for itself. Lorikeet's published pricing has Start at $2,100 a month for teams under 5,000 monthly tickets and Scale at $5,100 a month for 5,000 to 20,000, both paid annually with the platform and forward-deployed support included. Above 20,000 tickets, Signature pricing is custom with a $75K platform fee. Automated QA is optional at $0.30 or $0.27 per ticket.

Break-even resolutions per month equal the fixed fee divided by the human cost avoided per ticket minus the price per resolution. On Scale at $7.85 per digital ticket that is $5,100 divided by $6.95, about 734 resolutions a month. Fixed fees also explain a real limit: at low volume the effective price climbs. Lorikeet's own pricing examples work out to about $2.39 per resolution at 1,500 chat resolutions a month on Start, against about $1.92 at 5,000 on Scale. A team resolving a few hundred tickets a month will usually get a better return from a pure per-outcome tool with no plan fee.

6. The residual premium: cost the tickets humans keep

The tickets an AI agent leaves behind are, by design, the harder ones, so costing them at the blended average overstates the saving. When the simple password resets and order-status questions leave the human queue, average handle time on the remaining tickets rises. That is expected, not a failure, but it belongs in the model. This one costs the residual at a 25% premium as a stress case. Replace that with your own handle-time data split by intent once you have a month of production tickets.

Field evidence supports the direction rather than a single number. In the NBER study of 5,179 support agents, an AI assistant raised issues resolved per hour by 14% on average and 34% for novices, which tells you that tooling for the humans who remain is its own lever, separate from autonomous resolution.

7. Cash conversion: when freed hours become savings

Freed hours only become savings when they stop a hire, a contractor or overtime. In the 50% scenario below, the AI takes about 1,543 productive hours a month off the human queue, roughly 14 full-time agents at 160 paid hours each and 30% shrinkage. If you are growing, that is the headcount you do not add. If you are not, the saving is theoretical until you plan attrition or redeploy people. Gartner predicts that by 2027, 50% of organizations that expected to significantly reduce their service workforce will abandon those plans, and its poll found 95% of service leaders plan to retain human agents. Price cost per resolution into the contract too: Gartner also predicts GenAI cost per resolution will exceed $3 by 2030, so lock rates for the term.

What does AI customer support ROI look like for a 20,000-ticket team?

Take a team handling 16,000 chat and email tickets and 4,000 calls a month on the Scale plan: $5,100 a month, $0.90 per digital resolution, $1.20 per voice resolution on a three-minute average, and automated QA on every ticket at $0.27. Human cost is $7.85 per digital ticket and $10.98 per call, so the all-human baseline is $169,520 a month. Escalations to a person are not charged, so the resolution line carries only closed tickets.

Read the stressed row first. At 35% the case is thin, which is why the resolution share deserves the most diligence. For payback, assume $40,000 of one-time internal effort for integration, workflow building and pre-launch testing, and a ramp of 20% resolved in month one, 35% in month two and 50% from month three. Unstressed, that effort is recovered during month two; stressed, during month three. The stressed month one is negative, about minus $14,300, because QA and the plan fee arrive before resolutions do. A deeper payback treatment is in payback period for AI customer support.

Tools that help

These vendors publish enough to run the model, or say plainly that they do not. Prices are as listed on each vendor's page on 30 October 2026.

Where Zendesk or Fin is the better pick: if your volume is mostly simple questions and you already run Zendesk or Intercom, the native agent avoids a second contract, and on Fin there is no plan fee to clear. Lorikeet earns its fee on tickets that need actions in your systems, such as refunds, identity checks and account changes, which it runs through workflows and deterministic code described on its how it works page.

What still needs a human

Judgement calls with money, safety or legal consequences, and any customer in distress. Keep these on the human side of the model from day one. Dispute and chargeback decisions: the agent can gather evidence and start the clock, but a person decides. Hardship and vulnerability: a customer who says they cannot pay should reach a person quickly, with context already collected. Clinical questions in healthtech: the AI escalates to clinicians and never advises. Novel issues with no workflow yet: a signal to build, not something to improvise. The residual 35 to 65 percent of tickets is where your CSAT and compliance exposure now sit, so design the escalation path as carefully as the automation.

Questions to ask before the number goes in a budget

• What is our loaded cost per productive hour, and which overhead multiplier did finance agree?

• Which intents can be closed with a system action, and what share of volume do they cover?

• Who defines a resolution in the contract, and can we mark a bad one as unbillable?

• Are escalations, handoffs or conversations billed, and at what rate?

• Which freed hours turn into avoided hires, and when?

• Will we measure reopen rate and CSAT on AI-handled tickets separately from human ones?

To test the model on your own tickets, start a 30-day free trial; Lorikeet's average time to first live ticket is under 29 days.