Gartner reported on 4 November 2026 that 4% of customer service leaders have achieved the return they expected from AI, based on a survey of 275 service and support leaders in July and August 2026 (Gartner, 2026). A second release from the same survey found that 21% had reduced human agent headcount because of AI, 7% had already rehired after earlier cuts and 11% planned to rehire within 18 months (Gartner, 2026).
Gartner's release says returns improve when AI spending targets customer value creation, not just cost reduction. Separate from the survey, one reason cost-reduction cases fall short sits in arithmetic that workforce planners already use. Many business cases assume that automating a share of contacts removes the same share of agent hours. Queueing maths says it removes less.
What a staffing model calculates
Contact centres staff to offered load, not to contact counts. Offered load is the arrival rate multiplied by average handle time, measured in Erlangs. One Erlang is one agent's worth of continuous work.
Planners then staff above that load, because callers arrive at random and some will find every agent busy. The margin depends on the service-level target, such as 80% of calls answered within 20 seconds. The textbook calculation, Erlang C, descends from formulas for loss and waiting time that the Danish mathematician A.K. Erlang published in 1917 for telephone exchanges (Plus Magazine).
Why the margin does not grow in proportion
The standard rule of thumb, the square-root staffing rule described in Gans, Koole and Mandelbaum's 2003 review of call centre research, staffs at the offered load plus a multiple of its square root (Gans, Koole and Mandelbaum, 2003). The margin therefore grows with the square root of the load, not in proportion to it. Garnett, Mandelbaum and Reiman showed that the rule still holds when callers hang up (Garnett, Mandelbaum and Reiman, 2002). The result is an economy of scale: larger queues need proportionally fewer agents for the same service level.
A worked Erlang C example published by Lorikeet, an AI customer support vendor, shows the size of the effect. At an 80/20 target and a five-minute handle time, 50 calls per half hour need 12 agents, or 1.44 agents per Erlang. At 400 calls per half hour the requirement is 74 agents, or 1.11 per Erlang (Lorikeet).
Run that in reverse. When automation drains contacts from a queue, the human queue that remains is smaller, and smaller queues need more agents per unit of work. Required headcount falls by a smaller percentage than load does. The gap is widest in small queues and narrower in large ones.
The contacts that leave are not average
A second effect compounds the first. Automation tends to absorb the shortest, most routine contacts. The ones left for people run longer on average, so average handle time on the human queue rises even when no agent works more slowly. Microsoft, which sells contact centre software, showed this in a 2021 post: removing the shorter calls in a sample through a bot raised average handle time (Microsoft, 2021).
Since load is volume multiplied by handle time, longer calls offset part of the drop in volume. A projection built on pre-automation handle time overstates how much load leaves. We found little published, multi-site evidence on the size of this effect; most treatments are worked illustrations.
How planning approaches differ
A linear model divides total workload by the productive hours of one agent. It is quick to build and misses both effects above.
Erlang C captures the queueing effect for a single queue but assumes no caller abandons. Erlang A, analysed by Garnett and colleagues, adds abandonment.
Neither fits a multi-skill centre, where agents serve several queues. For those, researchers including Cezik and L'Ecuyer estimate service levels by simulation inside the staffing optimisation (Cezik and L'Ecuyer, 2004). That matters when bot handoffs go to a separate queue: splitting one pool into two smaller ones generally gives up some of these economies of scale.
What to check in a savings projection
- Whether the projection runs through an interval-level staffing model or applies the automation rate straight to headcount.
- Which handle time it assumes for contacts that stay with people. A pre-automation average understates the remaining load.
- Whether the automation rate is broken out by interval. Centres staff each interval separately, and low-volume intervals sit furthest from proportional savings, so when in the day the bot works changes what it saves.
- Whether escalations join the main queue or a separate one, and whether the model accounts for abandonment.
What commonly goes wrong
The first error is sequencing: cutting headcount on the projected automation rate before the measured load on the human queue has moved. In February 2026, Gartner predicted that half of the companies that attributed headcount cuts to AI would rehire staff for similar work by 2027, under different job titles (Gartner, 2026). The November survey's 7% rehiring figure suggests that process has started.
The second is reading a rise in human handle time after launch as a productivity problem, when it can be the arithmetic result of the bot taking the short contacts. Microsoft's post makes the same point and argues against judging a bot on average handle time alone.
Sources
- Gartner. "Gartner Survey Reveals Just 4% of Customer Service Leaders Have Achieved the Expected ROI of Their AI Initiatives." 2026. https://www.gartner.com/en/newsroom/press-releases/2026-11-04-gartner-survey-reveals-just-4-percent-of-customer-service-leaders-have-achieved-the-expected-roi-of-their-ai-initiatives
- Gartner. "Gartner Survey Finds 21% of Service and Support Leaders Have Reduced Human Agent Headcount Because of AI." 2026. https://www.gartner.com/en/newsroom/press-releases/2026-11-05-gartner-survey-finds-21-percent-of-service-and-support-leaders-have-reduced-human-agent-headcount-because-of-ai
- Gartner. "Gartner Predicts Half of Companies That Cut Customer Service Staff Due to AI Will Rehire by 2027." 2026. https://www.gartner.com/en/newsroom/press-releases/2026-02-03-gartner-predicts-half-of-companies-that-cut-customer-service-staff-due-to-ai-will-rehire-by-2027
- Plus Magazine. "Agner Krarup Erlang (1878-1929)." https://plus.maths.org/content/node/2154
- Gans, Koole and Mandelbaum. "Telephone Call Centers: Tutorial, Review, and Research Prospects." Manufacturing & Service Operations Management, 2003. https://pubsonline.informs.org/doi/pdf/10.1287/msom.5.2.79.16071
- Garnett, Mandelbaum and Reiman. "Designing a Call Center with Impatient Customers." Manufacturing & Service Operations Management, 2002. https://aviman.technion.ac.il/files/References/abandon02.pdf
- Lorikeet (vendor). "Erlang C Formula: How Many Agents You Need 2026." 2026. https://www.lorikeetcx.ai/articles/call-center-staffing-erlang-c-2026
- Microsoft (vendor). "Contact centers: Intelligent bots can increase average handle time." 2021. https://www.microsoft.com/en-us/industry/microsoft-in-business/customer-experience/2021/04/02/contact-centers-intelligent-bots-can-increase-average-handle-time/
- Cezik and L'Ecuyer. "Staffing Multiskill Call Centers via Linear Programming and Simulation." GERAD working paper G-2004-97, 2004. https://www.gerad.ca/papers/G-2004-97.pdf
