AI agent unit economics in 2026: the true cost per resolved task (and the human break-even)

Summary. An AI agent resolves a support task for about $0.62 on average in 2026, against roughly $7.40 for a human, per the McKinsey AI in Customer Service sample cited by vendor benchmarks. That 12x gap is the headline every AI vendor sells. The real unit economics are more sober. Agentic systems burn 5 to 30 times more tokens per task than a plain chatbot, per Gartner's March 2026 analysis; median tier-1 deflection across enterprise programs is only 41.2%, with the top quartile at 58.7%; and every task the agent fails escalates to a human who still costs $7.40. Gartner expects half of generative AI projects to overrun their budgets by 2028. This is a cost-per-resolved-task model you can apply before you sign a per-resolution contract, with the two numbers that actually decide whether an agent beats a human.

Why "cost per task" is the only number that matters

Most AI agent business cases are built on the wrong figure. A vendor quotes a per-token price of a few dollars per million, or a per-resolution fee near $0.99, and the buyer multiplies it by expected volume. That math is clean and almost always wrong, because it ignores the tasks the agent does not resolve and the human who picks them up.

The honest unit is the fully-loaded cost of one resolved task: the direct AI cost of every attempt, plus the cost of the human who handles the fraction the agent could not. Priced that way, an agent's economics hinge on two numbers and nothing else: its resolution rate, and the cost of the human it is replacing. Get those two right and the rest of the model follows. This piece builds that model from sourced 2026 inputs, and it sits alongside our wider work on enterprise AI agents in production.

The two human baselines: contact and resolution

Start with what a human costs, because that is the number the agent has to beat. Human-handled support averages about $13.50 per contact across assisted channels at SQM and Forrester benchmarks in 2026. By channel, phone runs $17 or more per contact, live chat $6 to $12 per conversation, and email looks cheap at $8 to $15 but reaches $24 to $60 per actual resolution once back-and-forth is counted. Complex healthcare and B2B tickets reach $60. On a resolution basis, the McKinsey 2026 sample puts a human resolution near $7.40.

Two words matter here: contact and resolution. A contact is one interaction. A resolution is a solved problem, which can take several contacts. Agents are usually sold against the lower per-contact figure and measured against the easier containment metric, which flatters the comparison. Price the model against real resolutions and the picture is truthful.

The AI side: what a resolved task actually costs

Now the agent. Its cost per attempt is small but not the trivial figure the token price suggests, because an agentic workflow is not one model call. It is a loop: read the ticket, retrieve context, call a tool, check the result, retry, respond. Gartner's March 2026 analysis found agentic models consume 5 to 30 times more tokens per task than a standard chatbot for exactly this reason.

"Organizations are rapidly moving from experimentation to scaled deployment of AI coding agents, but many are underestimating the financial impact of rising token consumption," said Nitish Tyagi, Senior Principal Analyst at Gartner, in a June 2026 forecast. He was describing coding agents, but the token-multiplier dynamic is identical for support and operations agents.

Work a realistic attempt. A five-step support agent that sends about 3,000 input and 500 output tokens per step spends roughly 15,000 input and 2,500 output tokens per task. On a mid-tier model near $2 per million input and $10 per million output as of July 2026, that is about $0.03 plus $0.025, near $0.055 in raw tokens. Add embeddings, vector retrieval, orchestration, and a slice of evaluation and monitoring, and a realistic loaded figure is $0.15 to $0.30 per attempt. If you buy instead of build, the market prices a resolved ticket at $0.49 to $2.00, with Intercom's Fin at a reference $0.99.

Read the table for the punchline: the token line is small, and the escalation line dominates. To keep even that small token line from drifting as volume grows, LLM hybrid routing to control token spend sends the easy majority of tasks to a cheap model and reserves the expensive model for the hard minority.

The model: fully-loaded cost per resolved task

Put the two sides together. For every inbound task, the agent attempts it at cost C. A fraction R is resolved by the agent. The remaining fraction (1 minus R) escalates to a human at cost H. Because both paths end in a solved problem, the blended cost per resolved task is simply:

Cost per resolved task = C + (1 − R) × H

Take C at $0.20 per attempt, H at $7.40 per human resolution, and run it across the resolution rates that real programs actually hit in 2026.

The lesson is stark. At the 41.2% median deflection, an agent is already 39% cheaper per resolved task than an all-human queue, and the saving grows fast with the resolution rate. The single highest-use move is not shaving token cost from $0.20 to $0.10; it is lifting the resolution rate, because that shrinks the $7.40 escalation line that dominates the total.

When the agent beats a human, algebraically

The break-even is worth stating plainly. The agent wins on marginal cost when its blended cost is below the all-human cost:

C + (1 − R) × H < H, which simplifies to R > C ÷ H

With C at $0.20 and H at $7.40, the agent wins on pure marginal cost once its resolution rate clears about 2.7%. That is a very low bar, which is why almost any working agent looks cheap on a spreadsheet. The catch is that R must be genuine resolution, not containment, and the model above counts only marginal cost. Add the fixed build and integration cost, amortized over volume, and the honest break-even moves up. At low volume the fixed cost per task is large; at high volume it approaches zero. Volume, not cleverness, is what makes agent economics work.

The trap: containment is not resolution

Here is where most agent business cases quietly fail. Containment counts a task as handled whenever the interaction stayed inside the bot and never escalated, even if the customer gave up unsolved. It is the easiest metric to inflate and the least connected to whether anyone's problem was solved. Optimize for containment and the per-resolution invoice looks wonderful while the real work leaks into repeat contacts, worse satisfaction, and churn that never appears on the AI bill.

Use resolution rate, the share of tasks where the customer's problem was actually solved, and the model stays honest. A contained-but-unsolved task is not a saving; it is a deferred cost that returns as a second contact at full human price, or as a lost customer. When you evaluate a vendor, the pricing model matters less than two numbers: the true resolution rate and how fast it improves in the first 90 days. Instrumenting that honestly is the same discipline as catching AI agent evals that surface silent failures before they reach production.

The scaling trap founders miss

The other surprise is that unit cost is not constant. As usage grows, the token line grows with it, and agentic loops multiply the effect. One fintech fraud agent went from $5,000 a month at 50 users to $15,000 a month at 500 users inside a few months, because every retry, tool call, and context reload spent more tokens as real traffic hit the system. Per-token prices keep falling, and Gartner still expects roughly a 90% drop in inference cost per token by 2030, yet monthly bills rise because usage outpaces price cuts. Model the cost at your projected volume, not your pilot volume, or the unit economics that looked healthy at 50 users will not survive 5,000.

Build versus buy for the agent

The per-resolution vendor model and the build model price the same work differently. A vendor fee near $0.99 per resolution is predictable and needs no platform team, but it grows linearly with every resolution forever. Building on your own stack carries a larger upfront cost and real engineering, but the marginal cost per task can fall to the token figure at scale.

The crossover is volume. Below a few tens of thousands of resolutions a month, buying is usually cheaper all-in once you count the platform team a build needs. Above that, the build's lower marginal cost starts to win. The same reasoning drives our detailed build vs buy for an AI customer support agent and the platform-level Agentforce vs Copilot Studio agent cost comparison.

India-specific considerations

The model shifts where human labor is cheaper. India's support and BPO wage base is well below the United States, so the human cost per contact that an agent has to beat is lower than the $7.40 to $13.50 benchmarks above. Mechanically, a cheaper human baseline raises the resolution rate an agent needs before it wins on pure cost, and it lengthens the payback on any build. The agent case is still strong at high volume and high resolution, but the margin is thinner, and containment-inflated numbers are more dangerous because the human alternative was already inexpensive.

Data protection also enters the unit cost. An agent that processes the personal data of Indian residents falls under the Digital Personal Data Protection Act 2023, which adds consent, retention, and data-handling engineering to the build and, therefore, to the amortized cost per task. eCorpIT designs agent systems aligned with DPDP Act requirements so those controls are part of the architecture rather than a retrofit.

A checklist before you sign

Before committing to an agent, confirm five things. Measure the true resolution rate, not containment, on your own tickets. Price the human baseline you are actually replacing, per resolution, in your market. Model token cost at projected volume with the agentic multiplier applied. Decide build versus buy on volume, not vendor enthusiasm. And instrument cost and quality from day one, in staging, so the unit economics are visible before real traffic locks them in. The programs that overrun, and Gartner expects half of generative AI projects to do so by 2028, are the ones that skipped these and trusted the headline number.

FAQ

How eCorpIT can help

eCorpIT is a Gurugram technology company, founded in 2021 and assessed at CMMI Level 5, MSME certified, and ISO 27001:2022 certified. We build and evaluate enterprise AI agents with a unit-economics discipline: measuring true resolution rate, metering token cost per task, and modelling the human break-even before a line of production code ships. We design agent systems aligned with DPDP Act requirements. If you want an honest cost-per-resolved-task model for your use case, talk to our team.

References

1. AI Support Resolution Rates and ROI, Ranked 2026 (Lorikeet)

1. The State of AI Support Pricing 2026 (Drag)

1. ROI of AI Customer Service: 2026 Benchmarks (Fin)

1. Customer Support Cost Per Ticket Benchmarks 2026 (Stealth Agents)

1. AI Agent vs Human Agent Cost, 2026 (eesel AI)

1. 42 Customer Service Cost Statistics 2026 (Ringly)

1. Gartner Predicts AI Coding Costs Will Surpass Average Developer's Salary by 2028 as Token Consumption Surges

1. Gartner Forecasts 90% Drop in LLM Inference Costs by 2030 (AIwire)

1. Gartner: Half of Gen AI Projects Could Exceed Budget by 2028 (Campus Technology)

1. The Inference Cost Trap: Why AI Agent Economics Break at Scale (TechAhead)

1. LLM API Pricing, July 2026 (TLDL)

1. AI Customer Support Resolution Rate Benchmarks 2026 (Notch)

1. AI Agent Pricing Models: Per-Resolution vs Per-Seat (Quickchat AI)

1. AI Agent Adoption 2026: Enterprise Data Points (DigitalApplied)

_Last updated: 31 July 2026._