AI customer service costs in 2026: what vendors actually publish

Summary. Intercom lists Fin at $0.99 per resolution, Salesforce Agentforce at $2.00 per conversation or roughly $0.10 per action, Freshdesk Freddy at $0.10 per session, and Zendesk at about $1.20 to $1.50 per Verified Resolution on committed volume. Those are published rates, checkable today. The number the market actually quotes, $0.62 for an AI resolution against $7.40 for a human one, is not. It traces to a single statistics roundup dated 22 April 2026 that credits "the McKinsey AI in Customer Service 2026 sample" without a link, and no McKinsey publication of that name is findable. McKinsey's own customer-service research reports percentages, never per-ticket dollars. Gartner's 91% figure for executive pressure on service leaders is real and dated 18 February 2026. Salesforce's 66% agentic-AI adoption figure is real and dated 20 May 2026. This page previously carried the $0.62 comparison and a 340% first-year ROI figure. Both have been removed, and what replaces them is the set of numbers a buyer can verify.

If you are building the business case, the honest starting position is narrower than the marketing suggests. Vendors publish what they charge. Almost nobody publishes what a human ticket costs, which means every headline "AI is 12 times cheaper" claim rests on a denominator that no analyst firm stands behind. That does not make AI support uneconomic. It means the savings case has to be built from your own cost base, not from a borrowed multiplier. For the wider operating model, pair this with our guide to AI customer-experience use cases and ROI.

What changed on this page

This article was published on 30 June 2026 with a cost table attributed to a McKinsey 2026 customer-service sample. Re-verification on 21 August 2026 could not locate that publication, and the figures did not appear in the article's own reference list. The affected numbers were the $0.62 blended AI resolution cost, the $7.40 human comparator, the $0.41 chat and $1.18 voice splits, the 340% first-year ROI, the 41.2% median deflection rate and the 4.2-month payback. All have been withdrawn. The sections below carry only figures traceable to a named publisher with a date.

That correction is worth stating plainly because the same numbers circulate widely. If your board deck cites $0.62 against $7.40, it is almost certainly downstream of the same unsourced roundup.

The rates vendors actually publish

Six vendors publish a unit rate you can read without a sales call. The rates are not comparable to each other, because each vendor bills a different event.

These figures come from Intercom's own AI agent pricing comparison, dated 8 July 2026. Read it as what it is: a vendor page that positions Fin favourably against its competitors. The list prices in it are still the most complete published set available, and the competitor rates are the ones a buyer can check against each vendor's own pages.

The billing event matters more than the rate. A $2.00 per-conversation model charges for interactions the AI failed to resolve. At a 60% resolution rate, two fifths of that spend buys nothing. A $0.10 per-session model looks cheap until one customer issue takes five sessions across three days. Comparing sticker prices across these models produces a number that means nothing.

Ada does not publish pricing at all, and neither does Decagon, so neither can be modelled before a sales call. Zendesk announced outcome-based pricing on 28 August 2024 without naming a price, and its public pricing page still lists only per-seat rates of $19, $55 and $115 per agent per month. The per-resolution overage rates circulating for Zendesk come from third-party analysis, not from Zendesk.

Why the $7.40 human comparator does not exist

The AI side of the comparison is well documented. The human side is not. No analyst firm, vendor benchmark or paper publishes a defensible 2026 dollar cost for a human-handled support ticket that others can reproduce.

The clearest evidence for this sits inside a vendor's own page. Intercom's pricing comparison, which publishes precise AI rates down to the cent, sources its human-agent comparator of $6.00 to $8.00 to a statistics listicle rather than to an analyst report. When the vendor with the most transparent AI pricing in the market has to borrow its denominator from a blog roundup, the denominator is not established.

Gartner does publish the structural fact that anchors the case, in its 31 August 2022 release: agent labour "can represent up to 95% of contact center costs", across roughly 17 million contact-centre agents worldwide. That is a ratio, not a price. It tells you where the money is without telling you what a ticket costs.

Four cited sources, four incompatible answers

This page previously cited several statistics roundups. Reading them side by side is the fastest way to see that no consensus figure exists.

The AI figures span a factor of six. The human figures span a factor of two and a quarter. Three different research houses are credited for numbers that do not agree. A cost model built on any one of these rows is a cost model built on a coin flip, which is the practical reason to build yours from your own contact volume and fully loaded agent cost instead.

What McKinsey actually publishes

McKinsey's customer-service research reports proportional change, not unit prices. The 2023 study The next frontier of customer engagement describes a transformation that produced "a doubling to tripling of self-service channel use, a 40 to 50 percent reduction in service interactions, and a more than 20 percent reduction in cost-to-serve". That is a single institution's result, not an industry benchmark, and McKinsey presents it as a case study.

The firm's February 2026 study, Building trust: How customer care leaders pull ahead with AI, surveyed 440 customer care leaders and executives. Its headline result is a spread between cohorts: 40% of leaders, defined as the top 10% of respondents, reported significantly improved customer experience scores over the previous 12 months, against 12% of laggards. Again, no dollar figure per ticket.

There is one per-call number in McKinsey's corpus, and it belongs to a vendor rather than to McKinsey. In the March 2025 article on the contact-centre operating model, Gadi Shamia, CEO of Replicant, is quoted saying: "Implementing AI agents into our customers' contact centers has driven a 50 percent reduction in cost per call." McKinsey attaches a disclaimer that interviewee comments are their own and do not reflect the firm's positions. Treat it as a supplier's claim about its own product, which is what it is.

Deflection, containment and resolution

The distinction that most often breaks a business case is now enforced in a major vendor's billing, which makes it concrete rather than theoretical. Zendesk's May 2026 restructure splits AI outcomes into three tiers, and only one of them is billable.

A vendor that only bills for confirmed resolutions has priced in the gap between containment and resolution. That gap is the reason a dashboard can show high deflection while customers quietly go unhelped, and it is why containment rate is the wrong number to put in front of a board. Measure resolution and first-contact resolution. Our breakdown of conversational AI patterns for CX covers the design choices that move a programme from containment to genuine resolution.

What is genuinely measured: adoption and pressure

Adoption data is the strongest evidence base in this market, because two large vendors run disclosed-methodology surveys.

Salesforce's State of Service: AI Agents Edition, published 20 May 2026, reports that 66% of customer service organisations now use agentic AI, up from 39% in 2025, a 1.7 times increase, with 85% using at least one form of AI. The methodology is stated: 3,075 responses collected between 9 March and 4 April 2026. Note the precision of the claim. The 66% is agentic AI specifically, and conflating it with the broader 85% overstates the case.

Gartner's survey of service and support leaders, released 18 February 2026, found that 91% reported pressure from executive leadership to implement AI. The sample is 321 leaders surveyed in October 2025.

The most quoted market-level number, Gartner's projection that conversational AI would reduce contact-centre agent labour costs by $80 billion in 2026, is real but old. It comes from a press release dated 31 August 2022, which also forecast that one in ten agent interactions would be automated by 2026, against an estimated 1.6% at the time. It is a four-year-old forecast whose target year is now, and no Gartner release revisiting it could be found. Cite it as a 2022 prediction or not at all.

Building a business case you can defend

Start from your own numbers, because the borrowed ones do not survive scrutiny. Take your monthly contact volume and your fully loaded cost per human resolution, which you can compute from payroll, benefits, tooling, management overhead and the cost of staffing for peak. That figure is knowable inside your business and unknowable outside it, which is precisely why no published benchmark is reliable.

Then apply a vendor's published rate to the share of contacts you expect it to resolve, using the vendor's own definition of a billable event. Model the failure case explicitly: on a per-conversation model, unresolved interactions still bill. Add platform and seat costs, which for a 20-agent team can run into thousands of dollars a month before a single AI resolution is counted, and add implementation. The real cost is usually the integration and the knowledge-base work, not the inference.

Finally, commit to reporting resolution rather than containment, and name the downside scenarios: an AI share pushed past what resolution quality supports, or a knowledge base left to rot. A case built this way is defensible in a board meeting a year later. A case built on a $0.62 figure is not, because the figure cannot be produced on request. If you are weighing an in-house build against a platform, our analysis of build versus buy for an AI support agent works through the same arithmetic, and the wider enterprise AI agents in production pillar covers the governance that keeps a deployment safe.

Choosing a delivery partner

Buying a platform and running one are different problems. The published rate is the smallest line in a first-year budget once integration with your helpdesk, knowledge-base restructuring, escalation design and evaluation harnesses are included. A chatbot development company earns its fee on those, not on the model choice, and the work that separates a 30% resolution rate from a 70% one is content and routing rather than prompt wording.

Scope the permissions question early. An agent that can issue a refund or change account details needs bounded permissions and a human checkpoint on irreversible actions. The engineering behind that is covered in our notes on shipping AI agents to production, and the per-task arithmetic in our work on AI agent unit economics.

India-specific considerations

For Indian businesses, and for the global support operations run from India, two constraints sit alongside the cost case. Under the Digital Personal Data Protection Act, 2023, an agent handling customer personal data such as names, contact details and order history needs a lawful basis, purpose limitation and an audit trail. Build consent capture and logging into the first channel rather than retrofitting them across five.

The second is language. Resolution quality, not containment, depends on the agent handling the languages your customers actually use. An English-only bot that closes a Hindi conversation without solving anything scores well on containment and badly on everything that matters. Where agent salaries are lower in rupee terms, the arithmetic that makes AI support attractive in a US cost base compresses, so the India business case has to be run on Indian loaded costs rather than imported per-ticket figures.

FAQ

How eCorpIT can help

eCorpIT is a Gurugram-based, CMMI Level 5, MSME-certified and ISO 27001:2022 certified technology organisation whose senior engineering teams build AI customer-service systems measured on resolution rather than containment. We design hybrid workflows with clean escalation, maintainable knowledge bases, scoped-permission agents with human checkpoints on irreversible actions, and reporting that separates confirmed resolutions from contained conversations. We design applications aligned with Digital Personal Data Protection Act requirements. If you want a cost model built on your contact volume rather than a borrowed benchmark, talk to us through our contact page.

References

1. AI customer service agent pricing comparison, 2026 — Intercom, July 8, 2026.

1. Zendesk introduces outcome-based pricing — Zendesk newsroom, August 28, 2024.

1. Zendesk plans and pricing — Zendesk.

1. Gartner survey finds 91% of customer service leaders under pressure to implement AI in 2026 — Gartner, February 18, 2026.

1. Gartner predicts conversational AI will reduce contact center agent labor costs by $80 billion in 2026 — Gartner, August 31, 2022.

1. AI service agents improve customer satisfaction: State of Service, AI Agents Edition — Salesforce, May 20, 2026.

1. Building trust: How customer care leaders pull ahead with AI — McKinsey and Company, February 2026.

1. The next frontier of customer engagement: AI-enabled customer service — McKinsey and Company, March 2023.

1. The contact center crossroads: Finding the right mix of humans and AI — McKinsey and Company, March 19, 2025.

1. ROI of AI customer service agents: benchmarks and data — Fin.ai, March 19, 2026.

1. Customer service AI agent statistics 2026 — Digital Applied, April 22, 2026.

1. AI customer service cost savings statistics — theStacc.

1. AI customer service statistics with sources — Lorikeet.

_Last updated: August 21, 2026._