servicedeskagents.com is an independent enterprise-IT reference. Not affiliated with ServiceNow, Moveworks, Aisera, Freshworks, Atlassian, Zendesk, or any AI ITSM vendor. Pricing reflects published rates where available and directional estimates where vendors do not publish; validate with vendor before procurement. // Last verified April 2026
[DEFL-BENCH-2026]P1 / BENCHMARK

AI Service Desk Deflection Rates

There Is No Published Gartner Number. Here Is What There Is.

Sources re-checked 7 September 2026
Direct answer

There is no published Gartner deflection rate for AI IT service desks. We went looking on 7 September 2026 and found no Gartner press release, newsroom item or public research page that states one. Gartner lists deflection rate as a service desk KPI in its public Peer Community material, with no number attached. The 20-30% average and 40-60% best-in-class pair that circulates under Gartner's name appears only in vendor and agency blog posts citing one another.

Every deflection figure you can actually read today is a vendor claim on a vendor site, and the vendors do not share a definition. Some of those claims do not survive a visit to the page they are attributed to. Measure your own baseline and treat any unsourced percentage, including one you read here in the past, as marketing until someone shows you the document.

None
Public independent benchmarks found
3
Incompatible definitions in use
72h
Re-open window in the strict definition
SECTION 01

Why "Deflection" Is Inconsistently Defined

Definitional drift is the reason cross-vendor deflection comparison does not work, and it is the reason a single industry benchmark would be close to meaningless even if one existed. The word "deflection" is used across the industry to mean three materially different things:

  • 1.Auto-reply deflection: the AI sent any response, regardless of whether the user was satisfied
  • 2.Self-service containment: the user interacted with the AI and did not open a new ticket within 24 hours
  • 3.Full AI resolution: the AI completely resolved the request without any human agent involvement and the user did not re-open within 72 hours

This reference uses definition 3 as the strict standard. Vendor-published numbers are typically definition 1 or 2, and vendors rarely say which. A vendor quoting definition 1 and a vendor quoting definition 3 can differ by tens of percentage points while describing identical performance, so the first question to ask about any deflection number is not how high it is but what it counts.

We previously published a "strict equivalent" column here that converted vendor figures by applying a flat 10-percentage-point compression. That conversion was our own invention with nothing behind it, so it has been removed. Deriving a number from another number does not make it a measurement.

SECTION 02

What Is and Is Not Publicly Published

This page used to carry three tiles here presenting 20-30%, 40-60% and 20-35% as independently verified baselines from Gartner and HDI. They are gone. When we went back to check them on 7 September 2026 we could not put a document behind any of them, and a benchmark with no document is not a benchmark.

What we searched: Gartner's public site, including its newsroom and its public research and Peer Community pages. Gartner names deflection rate as a service desk KPI and sells IT benchmarking as a service, but publishes no AI service desk deflection percentage that a member of the public can read. The HDI and MetricNet figures sit behind a paid survey, so they cannot be independently checked either. A number nobody outside the paywall can verify is not something we will assert as fact.

What we did find, and could verify at source: Atlassian publishes on its own site that Jira Service Management includes 1,000 Virtual Service Agent assisted conversations per month at no charge, and that conversations above that threshold start at $0.30 each with volume discounts, effective 16 October 2024. That is a commercial fact rather than a performance benchmark, but it is a real, dated, primary number, and it tells you something useful: the vendor built its packaging around AI conversation volume, not around a promised deflection percentage.

Sources checked 7 September 2026: gartner.com public pages; aisera.com/customers/cisco; moveworks.com; atlassian.com/blog/announcements/jsm-cloud-pricing-packaging-update.

SECTION 03

What Happened When We Checked Each Circulating Figure

Every widely repeated deflection number, the source it is attributed to, and what was actually on that source when we opened it on 7 September 2026. This is not a benchmark table. It is an audit of the claims.

Circulating claimAttributed toStated sourceWhat we foundVerdict
65% deflection at CiscoAiseraaisera.com/customers/ciscoChecked 7 Sep 2026: that page publishes no percentage at all.Not at stated source
75% average deflectionMoveworksmoveworks.com case studiesChecked 7 Sep 2026: moveworks.com publishes a customer quote of a 20-30% ticket volume reduction at BambooHR, not a 75% deflection rate.Not at stated source
20-30% industry averageGartnerNo document namedChecked 7 Sep 2026: no public Gartner press release or research page states this. Gartner lists deflection rate as a KPI without a number.No source found
40-60% best-in-classGartnerNo document namedChecked 7 Sep 2026: same result. The figure appears only in vendor and agency blog posts citing each other.No source found
20-35% first-year L1HDI / MetricNetPaywalled benchmarking surveyNot publicly readable. A figure behind a paywall cannot be independently checked, so we do not repeat it as fact.Not publicly checkable
1,000 assisted conversations included, then from $0.30 eachAtlassianatlassian.com/blog/announcements/jsm-cloud-pricing-packaging-updateVerified 7 Sep 2026 on Atlassian's own site. Effective 16 October 2024, with volume discounts above the threshold.Verified at source
note:A "not at stated source" verdict means the figure may still be true and may exist in a deck, a webinar or a page that has since changed. It means we could not read it on the page it is credited to, so we will not repeat it as fact.
SECTION 04

Measure Your Own Rate Instead

Since there is no benchmark worth budgeting against, the only deflection number that means anything about your organisation is the one you measure. Fix the definition first, in writing, before anyone builds a dashboard. Then count tickets the AI resolved end to end, divide by total inbound requests that reached the AI channel in the same window, and exclude every ticket a human agent touched or the user re-opened inside 72 hours.

Take the baseline before go-live, over the same weekday pattern and the same length of window you will use afterwards. This matters more than it sounds: inbound volume moves for reasons that have nothing to do with the AI, and a ratio with a moving denominator will show improvement that did not happen. Hold the window fixed and you can attribute the change.

Deflection also is not static once it starts. It climbs as the intent library fills and knowledge-base gaps close, which means an early reading understates the steady state and a vendor pilot timed for month twelve overstates what month one looked like. Report the curve, not a single figure, and say which month each reading came from.

The trajectory is driven by four factors. Knowledge-base hygiene is the largest single variable: RAG accuracy improves as stale articles are updated and gaps are filled. Intent library maturity is the second: the more ticket patterns the system has seen and correctly classified, the better it routes new inputs. Action framework depth is the third: platforms that can actually execute actions (not just answer) achieve higher deflection because they resolve multi-step tickets autonomously. Change management investment is the fourth: end-user adoption of the AI channel determines whether the AI has the ticket volume to learn from.

One caution on forecasts. Long-range autonomous-resolution predictions circulate widely and are usually attributed to analyst houses without a linkable document. We do not repeat them here for the same reason we removed the baselines above. If a vendor quotes you a future resolution percentage, ask for the research note ID and the publication date, and treat a proposal that cannot supply both as a proposal without evidence.

SECTION 05

What Drives Deflection Variance

Knowledge-base hygieneImpact: Highest

RAG accuracy depends directly on source quality. Fragmented or outdated KB is the most common cause of underperformance.

Intent library maturityImpact: High

The more intent patterns configured, the broader the AI's coverage. Start with 5-10 high-volume ticket types; expand over 12 months.

Action framework depthImpact: High

Platforms that execute actions (password reset, provisioning) deflect more than platforms that only answer. Integration investment is the constraint.

Channel coverageImpact: Medium

End-users must find and use the AI channel. Slack/Teams integration drives higher adoption than portal-only access.

Change managementImpact: Medium

End-users have to choose the AI channel over the old one. Under-investment here is a common cause of slow deflection, because an AI nobody routes tickets to cannot deflect them.

Time on platformImpact: Medium

Deflection is almost always higher at month 18 than month 6. Plan for a 12-18 month ramp to steady-state.

SECTION 06

FAQ

What is Gartner's IT service desk AI deflection rate?
There is no published one. We searched Gartner's public site on 7 September 2026 and found no press release, newsroom item or public research page stating an AI service desk deflection benchmark. Gartner names deflection rate as a service desk KPI with no number attached. The 20-30% and 40-60% pair circulating under Gartner's name appears only in vendor and agency blog posts citing each other.
What deflection figures can actually be verified?
Only vendor claims on vendor sites, and fewer than the industry repeats. On 7 September 2026 aisera.com/customers/cisco published no percentage at all, despite being the stated source for a widely quoted 65% Cisco figure. Moveworks' own site publishes a customer quote of a 20-30% ticket volume reduction at BambooHR, not a 75% deflection rate. The one number we verified at source is commercial: Atlassian includes 1,000 Virtual Service Agent assisted conversations per month and charges from $0.30 each above that, effective 16 October 2024.
What does 'deflection' mean in AI service desk?
Three incompatible things. Auto-reply deflection counts any AI response. Containment counts a user who did not open a new ticket within 24 hours. Full AI resolution counts only tickets the AI resolved end to end with no human agent and no re-open within 72 hours. Two vendors with identical performance can quote numbers tens of points apart, so ask which definition a figure uses before comparing it.
How do I measure my own deflection rate?
Fix the definition in writing, then baseline before go-live. Count tickets the AI resolved end to end, divide by total inbound requests reaching the AI channel in the same window, and exclude anything a human touched or the user re-opened inside 72 hours. Keep the window length and weekday pattern identical before and after, because inbound volume moves on its own and a moving denominator will manufacture an improvement.
Why does my AI service desk have low deflection?
The most common causes: (1) fragmented or out-of-date knowledge base undermining RAG accuracy; (2) sparse intent library missing high-volume ticket patterns; (3) lack of action framework integration meaning the AI can answer but not execute; (4) low end-user adoption of the AI channel due to inadequate change management investment.