Only 4% of customer service organisations have achieved the return they expected from AI (Gartner, n=275 service and support leaders, summer 2026). The other 96% did not all buy bad software. Most of them skipped the step that decides AI customer service ROI, and almost none of them did the break-even arithmetic first.
TL;DR: Gartner found that only 4% of service teams got their expected AI return, and that the winners invest nearly 1.8× more in foundations than in AI use cases; break-even depends mostly on volume and fixed cost.
Methodology & sources · checked 17 Nov 2026
- Gartner findings as reported by Customer Experience Magazine (5 Nov 2026), The CX Brief (10 Nov 2026) and CX Dive (17 Aug 2026). Gartner has not published the underlying use-case list.
- Human cost per contact: Gartner's customer service cost benchmark (Feb 2024). AI prices from the Aissist and Intercom pricing pages.
- The break-even table is a model with stated assumptions, not measured customer data. Replace the inputs with yours.
- Disclosure: this is Aissist's blog. Aissist sells an AI operational layer for customer service and sales.

How many customer service teams actually get ROI from AI?
Very few: 4% of service organisations got the return they expected, and only about a quarter of individual AI use cases produce one. The second figure comes from Gartner's July 2026 analysis of 432 customer service AI use cases, which CX Dive reported on 17 August: roughly one quarter return value, one quarter lose money, and 42% have unclear ROI. Only 11% of customer support use cases break even.
The money is not slowing down. Gartner's August 2026 report, cited by Customer Experience Magazine, found service AI spending up 38% while overall service and support budgets grew 2%. CX Dive adds that service teams run nearly five AI use cases on average and commit about 13% of their functional budget to AI.
So spending is rising nineteen times faster than the budget that pays for it, and two in five use cases cannot say whether they worked. That is not a technology problem yet. It is an accounting problem with a technology bill attached.
The useful reading of 4% is narrower than the headline. It does not say AI cannot resolve customer issues. It says that most organisations cannot show the return, which usually means they never defined it, never baselined it, or bought something that could not deliver it.
Why does AI customer service fail to deliver ROI?
AI customer service fails to deliver ROI when teams buy an application without funding the data, knowledge and integration layer it needs. Gartner's Christopher Sladdin put a number on it: AI-forward organisations invest nearly 1.8 times more in data and knowledge foundations than in the AI use cases themselves, according to Customer Experience Magazine's report from Gartner's Customer Service & Support conference.
Read that ratio the other way round. For every dollar the winners spend on the AI agent, they spend almost two on the things underneath it: a knowledge base that is current, order and account systems the AI can act on, and someone who owns quality. Most buyers did the opposite. They bought the agent, pointed it at the help centre, and hoped.
"AI's superpower is that it enables leadership to learn from customer interactions at scale." — Christopher Sladdin, Senior Director Analyst, Gartner, via Customer Experience Magazine
Sladdin also told leaders to deploy fewer, higher-value use cases that connect to each other, and Gartner predicts that by 2028 organisations building AI portfolios will get twice the business value of those buying standalone use cases.
This is the argument for an AI layer on top of the stack you already run, rather than another standalone tool next to it. An agent that can read the order, issue the refund and update the CRM resolves the request. An agent that can only quote the help centre deflects it, and the customer comes back.
Where is the break-even point for AI customer service?
An AI support deployment breaks even when monthly resolutions × (human cost per ticket − AI cost per resolution) covers the fixed cost of running it. As a formula: break-even resolution rate = monthly fixed cost ÷ (monthly volume × (human cost − AI cost per resolution)).
The table applies that formula. Assumptions, stated so you can replace them: $8,000 a month in fixed foundation cost (roughly one person owning knowledge, integrations and QA), $40,000 one-off to clean the knowledge base and connect systems, and payback calculated at 60% resolution, the low end of the range Armatis gives for well-performing bots.
| Monthly ticket volume | Blended cost per human-handled ticket | Cost per AI resolution | Resolution rate needed to break even | Payback in months (at 60%) |
|---|---|---|---|---|
| 1,000 | $6.00 | $0.60 | Not reachable (>100%) | Never |
| 1,000 | $13.50 | $0.60 | 62% | Never at 60% |
| 2,500 | $6.00 | $0.60 | 59% | 400+ (effectively never) |
| 2,500 | $13.50 | $0.99 | 26% | 3.7 |
| 5,000 | $6.00 | $0.60 | 30% | 4.9 |
| 5,000 | $13.50 | $3.00 | 15% | 1.7 |
| 20,000 | $6.00 | $3.00 | 13% | 1.4 |
| 20,000 | $13.50 | $0.60 | 3% | 0.3 |
Where the inputs come from: $13.50 is Gartner's median cost per assisted contact (Feb 2024); $6.00 sits inside Aissist's US text-ticket estimate of $4–$8. On the AI side, $0.60 is the cap on Aissist's pricing page ($0.09 per AI interaction, capped at $0.60 per resolution, read Nov 2026); $0.99 is the entry price per Fin outcome on Intercom's pricing page (read Nov 2026); $3.00 reflects Gartner's forecast, reported by Customer Experience Magazine, that GenAI cost per resolution will exceed $3 by 2030.

Two things fall out. First, volume is the hidden variable: below about 2,500 text tickets a month, fixed costs eat the savings at any realistic resolution rate. Second, price per resolution matters a lot when human tickets are cheap and very little when they are expensive. Put your own numbers into the AI support price comparison tool before you compare vendors' price lists.
What are the five reasons the return did not show up?
Most failed deployments share five causes, and only one of them is the AI itself.
- They bought the application and skipped the foundations. Gartner's 1.8× ratio is the evidence. A stale knowledge base produces confident wrong answers at scale.
- They measured deflection, not resolution. Deflection counts conversations that never reached a human; resolution counts problems solved. A deflected customer who calls back costs you twice. We unpack the difference in deflection vs resolution rate.
- They never set a baseline or an owner. In a separate study from Gartner's, Sprinklr research cited by Customer Experience Magazine found that 4% of organisations with AI in live operation had defined success metrics, recorded a baseline and named an outcome owner. That is how 42% of use cases end up "unclear."
- They ran too many disconnected use cases. Five pilots that each automate a fragment rarely beat one deployment that finishes the whole request.
- The savings never left the spreadsheet. Below break-even volume, or with freed hours that were never redeployed or removed, the model shows a saving the P&L never does.
"Containment is also too often mistaken for success." — Julie Geller, Principal Research Director, Info-Tech Research Group, via CX Dive
Reasons two and three compound. If you count deflections as wins and never recorded the before-picture, the dashboard will look healthy right up until the budget review. For the full metric set, see how to measure the true ROI of agentic AI; for project-level causes outside customer service, see why AI projects fail.
What is a realistic payback period for AI customer service?
At 5,000 or more tickets a month with funded foundations, the model above pays back in under six months; below about 2,500 text tickets, it may never pay back. The payback clock starts when resolution is real, not when the agent goes live.
Three conditions shorten it. Define a resolution in the contract as a solved request with no human and no repeat contact, so you pay for outcomes rather than attempts. Record the before-picture: cost per ticket, repeat-contact rate and CSAT. And give the AI actions, not just answers, so it can finish the request inside your helpdesk and back-office systems.
"The question isn't 'does it cost less?' but 'is the ROI good?'" — David Gillaux, Chairman, Armatis
Foundations also pay back in ways the cost table misses. In a published Aissist customer story, a global telecom provider's AI agents absorbed a 20× support surge within an hour of a carrier outage, the root cause was found in about 25 minutes, and CSAT held at 4.8/5.0. That is Sladdin's "learn from interactions at scale" in practice.
For reference, Aissist reports an 83% average resolution rate (up to 98% in mature deployments) and 4.8/5.0 CSAT across deployments. The table above uses 60% deliberately, so the arithmetic does not depend on any vendor's figures, ours included.
The 4% funded the layer underneath, not just the agent
The Gartner numbers do not show that AI customer service fails. They show that most buyers paid for an application and skipped the foundations, then measured deflection and called it success. The break-even arithmetic is simple, and it is mostly about volume and fixed cost, not about a few cents per resolution. Before you sign anything, run your volume, your human cost and a conservative resolution rate through the formula. If the line does not cross, fix the foundations or the scope first. If it does, insist on a contractual definition of resolution and measure it from day one.
Break-even starts with knowing what was actually resolved. Pulse™ reads every ticket and scores what happened to the customer, so your resolution rate is measured, not assumed. Book a consultation →
Frequently asked questions
What percentage of companies get ROI from AI customer service?
Gartner found that 4% of customer service organisations achieved the return they expected from AI, based on a 2026 survey of 275 service and support leaders. Separately, about a quarter of 432 customer service AI use cases Gartner analysed produce a return.
How do I calculate AI customer service break-even?
Divide your monthly fixed AI cost by monthly ticket volume multiplied by the difference between your human cost per ticket and the AI cost per resolution. The result is the resolution rate you need to break even.
What is the cost of a human-handled support contact?
Gartner's benchmark puts the median cost per contact at $13.50 for assisted channels and $1.84 for self-service. Text-only queues often cost less, so use your own fully loaded figure.
Does a cheaper price per resolution guarantee ROI?
No. When human tickets cost $13.50, moving from $0.60 to $3.00 per AI resolution shifts break-even only slightly. Volume, fixed costs and real resolution rates matter more.
How much does Aissist cost per resolution?
Aissist's pricing page lists $0.09 per AI interaction, capped at $0.60 per resolution, with escalated cases not charged. A free tier covers up to 1,000 tickets a month.
Why is service AI spending rising if ROI is rare?
Gartner reported service AI spending up 38% against 2% overall service budget growth. Leaders expect returns, and more than three-quarters plan to increase AI investment in 2026, according to CX Dive.
Changelog
- 17 Nov 2026: Rebuilt around Gartner's 4% finding (n=275) and the 1.8× foundations ratio; added the break-even and payback table built from published prices; removed the previous unit-economics model and Aissist research figures.
- 17 Sep 2026: First published.




