The Real Cost of a Support Ticket: Human vs AI Agent, With the Full Model

A human handled support ticket costs about 12 dollars once wages, benefits, occupancy and overhead are counted. An AI resolution costs 1.50 dollars at Zendesk list price. That looks like an easy trade until you price the ticket the bot fails to resolve, which reaches a human more expensively than it would have done directly. At a 50% deflection rate, a claimed 44% saving lands at 27%.

What a cost per ticket figure actually contains

Cost per ticket is the total operating expense of a support function divided by the tickets it handled. That sounds simple. It is the reason no two published benchmarks agree.

MetricNet, one of the older benchmarking firms in this category, defines the numerator as agent salaries and benefits, indirect personnel, technology and telecom, facilities, travel, training and office supplies. Most vendor ROI calculators use one of those seven lines and ignore the rest.

The gap matters because the AI comparison is a ratio. If the human number is understated, the saving looks smaller than it is. If it is overstated, the saving looks larger. Neither error is neutral, and the second one is the commercially convenient direction.

The three layers most cost models leave out

First, benefits. A salary figure is not a cost figure. Employer costs for benefits are not optional and they are large.

Second, occupancy. Agents are paid for eight hours and spend a fraction of that on live work. Breaks, training, meetings and idle time between contacts all sit inside the same wage bill.

Third, everything that is not the agent. Supervisors, quality assurance, workforce planning, licences, floor space and the helpdesk platform itself. This layer is invisible in a per seat budget and it is real money.

Building the human number from wage data

I would rather build this from public statistics than quote a benchmark I cannot inspect. Every input below has a source or a stated assumption, and you can change any of them.

Start with fully loaded labour, not salary

The Bureau of Labor Statistics reports a median hourly wage of 20.59 dollars for customer service representatives, against 2.81 million jobs, for May 2024. That is wages only.

The same agency reports that benefits made up 31.6% of total compensation for civilian workers in March 2026, at 15.60 dollars per hour against 33.72 dollars of wages. Applying that share gives a fully loaded hourly cost of 30.10 dollars.

That is the first honest correction. A support leader quoting a 43,000 dollar salary is understating the labour line by almost a third before anything else is added.

Occupancy is the assumption that moves the answer most

Assume agents spend 70% of paid hours on tickets. The cost of a productive hour becomes 43.00 dollars. At a 12 minute average handle time, that is 8.60 dollars of direct labour per ticket.

Then apply the non labour layer. This model uses a 1.4 times loading for supervision, tooling and facilities, which lands the ticket at 12.04 dollars.

Occupancy and that 1.4 multiplier are assumptions, not measurements, and I say so again in the limitations. Move occupancy to 80% and the ticket falls to 10.54 dollars. Move it to 60% and it rises to 14.05 dollars.

What an AI agent charges per resolution

This side of the ledger is easier, because three vendors publish list prices. The numbers are small, which is precisely why they distract.

Fin, formerly Intercom, charges 0.99 dollars per outcome, capped at one charge per conversation. Zendesk charges 1.50 dollars per automated resolution. Salesforce Agentforce charges 2 dollars per conversation, or 500 dollars per 100,000 Flex Credits with a standard action consuming 20 credits.

Sierra and Decagon quote through enterprise sales and publish no rate card. Any per resolution figure you see attributed to them traces to a reseller blog rather than the vendor, so I have left them out of the model. The token cost underneath all of these is a separate question, covered in the piece on inference costs and AI margins.

What counts as billable when the bot fails

Read the billing definitions rather than the price. They differ in ways that change your bill by double digit percentages.

Fin counts an assumed resolution when a customer goes quiet for 24 hours after its last answer. A customer who got a perfect answer and a customer who gave up both bill at 0.99 dollars. Fin also bills for a Procedure handoff that you configured to end with a human, while escalations triggered by the customer asking for a person are not billed.

Zendesk uses a 72 hour quiet period and states that an escalated conversation is not an automated resolution. Agentforce bills per conversation under its Conversations model, and consumes Flex Credits per action under the other, so neither depends on the agent succeeding.

That is three different answers to the same question. The shift from per seat to per outcome pricing is the wider story here, and it is the subject of the analysis of seat compression in SaaS pricing.

The escalation premium is the line that is missing

Here is the whole argument in one sentence. A ticket the bot could not resolve does not cost the same as a ticket that went straight to a person.

The customer has already spent minutes explaining the problem. The agent picks it up cold, rereads the transcript, and often asks for information the customer already gave. Vendor case studies almost never price this, because it appears on the human side of the ledger and the vendor is selling the other side.

Four minutes of rep time, and where that number comes from

Gartner surveyed 1,492 customers in December 2022 and found that 62% of customer service channel shifts are high effort. The same research put a number on the other side of it. A transition handled well saves an average of 4 minutes of agent time per journey, because the agent is not asking the customer to repeat what they already said.

Read that backwards and it is a cost. Four minutes on top of a 12 minute handle time is a 33% longer ticket, which prices the escalated ticket at 16.05 dollars in this model.

That is a conservative reading. It counts only agent time and ignores the customer's own wasted effort, the repeat contact, and the churn risk that Gartner attaches to high effort journeys.

The full model, with the assumptions exposed

Every input is below. Three of them are assumptions, marked as such, and they are the three you should replace first.

The break even deflection rate is 28%

Run the blend at a range of deflection rates and two lines appear. One is the arithmetic a vendor calculator produces. The other prices the escalation.

The 76% row is the resolution rate Salesforce published for Fin when it announced the 3.6 billion dollar acquisition in June 2026. It is a vendor figure across 30,000 customers, not a measurement of your queue.

On 1,000 tickets a month at a 50% deflection rate, the real saving is 3,260 dollars and the claimed saving is 5,270 dollars. The 2,010 dollar monthly gap is not a rounding error. It is 24,120 dollars a year of business case that does not exist.

The quality risk that no model prices

Everything above assumes a deflected ticket is a solved ticket. The public evidence says that assumption is generous.

Gartner surveyed 5,728 customers in December 2023 and found that only 14% of customer service issues fully resolve in self-service, while 73% of customers use self-service at some point. Even for issues customers described as very simple, only 36% resolved there. That research predates the current generation of agents, and it is the closest thing to an independent baseline that exists.

There is also a cost to the disclosure itself. A field experiment on more than 6,200 customers, published in Marketing Science in 2019, found that undisclosed chatbots matched proficient human workers, and that disclosing the bot before the conversation cut purchase rates by 79.7%. Competence was not the variable. Perception was.

What Klarna's reversal actually shows

Klarna is the case study everyone cites and almost nobody finishes. In early 2024 its assistant handled 2.3 million conversations in a month, did work equivalent to 700 agents, cut resolution time from 11 minutes to under 2, and was credited with 40 million dollars of annual profit.

By 2025 the company was rebuilding human capacity. Chief executive Sebastian Siemiatkowski has said publicly that the company cut too far and lost expertise it needed for complex and emotionally charged cases. Klarna now runs a hybrid model with a small specialist human team behind the automation.

My reading is that Klarna did not disprove the economics. It proved that the deflection rate you can sustain is set by your ticket mix, and that you find the ceiling by overshooting it. That is an expensive way to learn a number you could have modelled.

What the peer reviewed evidence actually supports

Strip out the vendor material and the strongest evidence in this category points at assistance rather than replacement.

Brynjolfsson, Li and Raymond studied the staggered rollout of a generative AI assistant across 5,172 customer support agents, published in the Quarterly Journal of Economics in 2025. Access to the assistant raised issues resolved per hour by 15%. The gain was concentrated among less experienced and lower skilled workers, while the most experienced agents saw small speed gains and small quality declines. Customers were also more polite and less likely to ask for a manager.

That is a 15% productivity gain on the human line, not a 76% removal of it. It is also the only study here with a credible identification strategy and a five figure sample. I weight it more heavily than any vendor case study, and I would build the first year of a business case on it rather than on a deflection promise. The broader pattern of where measurable AI return has shown up is covered in the piece on who is actually making money from generative AI.

Where this model is weakest

Four things, in order of how much damage they do.

The overhead multiplier is an assumption, not a measurement

The 1.4 times loading and the 70% occupancy figure carry no public source. MetricNet and HDI both keep their benchmark values behind paid membership, so the two inputs that move the answer most are the two least sourced in this post. If your finance team can produce real numbers for either, use theirs and ignore mine.

The benefits share is also a civilian workforce average rather than a contact centre figure. Support roles skew toward lower benefit loads than the economy wide mean, so 31.6% may be slightly generous.

The 4 minute escalation premium comes from research conducted in December 2022, before current agents were deployed. A modern handoff that passes full transcript context to the agent should cost less than 4 minutes. I have kept the figure because it is the only measured number available, and because the version that passes clean context is not the version most teams have shipped.

Finally, the strongest argument against everything here is that deflection rates are not static. A 30% rate at launch that reaches 60% after a year of tuning changes the verdict completely, and my table treats each rate as a fixed state rather than a path. Anyone running this model should run it as a 24 month curve, not a single row. The related question of when to build this capability rather than buy it is worked through in the build versus buy analysis for coding agents, and the same test applies here.

Frequently asked questions

Where to start this week

Pull three numbers out of your own systems before you read another vendor calculator. Your fully loaded hourly cost per agent, your average handle time, and your occupancy. Twenty minutes with payroll and your workforce tool gets you all three.

Then measure one thing you almost certainly are not measuring. Take last month's escalated conversations and compare their handle time against tickets that went straight to a person. If the gap is under 4 minutes, your handoff is better than the benchmark and the model above is too harsh on you. If it is over 4 minutes, that difference is the price of your current deflection strategy, and it belongs in the business case before anyone signs anything.

Put both into a contract question. Ask the vendor to state, in writing, which outcomes they bill for when the agent does not resolve the issue. The answer varies by vendor, it is written down, and almost nobody asks. How this reshapes the customer service software market itself is the subject of the piece on Salesforce in the agent era.

Related reading