Freshworks’ Customer Service Benchmark Says 83% Deflection. Its Earnings Calls Say 50%
The customer service benchmark report’s dataset is genuinely large and its operational benchmarks are useful. Two of its AI figures do not reconcile with what the company tells investors.
Freshworks published its 2026 Customer Service Benchmark Report, built from anonymised usage data across 33,889 accounts and 1.2 billion tickets, and the operational benchmarks in it are genuinely useful. The headline finding is that the top 20 per cent of customer service teams in India resolve issues 95 per cent faster than the rest. The more interesting numbers sit further down: an average AI agent deflection rate of 83.1 per cent, and year-over-year jumps that roughly double deflection across every company size.
Two of those figures need checking against what Freshworks tells its investors, and the comparison is available because the company is public and reports to a market that holds it to a different standard of precision.
What the report actually measures
The methodology is better than most vendor customer service research. Freshworks combined anonymised, aggregated product usage data from June 2025 to May 2026 with its annual customer survey and cohort research, covering 33,889 accounts across 63 industries, and excluded outliers. It reports two reference points for each metric: the regional or industry median, and the top 20th percentile, which it calls trendsetters. It also adds year-over-year comparisons for the first time, which is what makes the deflection figures checkable.
The APAC operating benchmarks are the most useful part. Trendsetters there resolve in 45 minutes 18 seconds against a regional median of 9 hours 6 minutes 43 seconds, achieve first contact resolution on 43.18 per cent of issues against 18.63 per cent, and hit resolution service levels 99.32 per cent of the time against 95.35 per cent. A support leader can hold those against their own queue and learn something true about the operation. The report’s industry sections carry the same shape for retail, software and the other verticals it covers, which makes them usable as a starting reference rather than a single headline.
The distinction between the two data sources matters and the report blurs it. The operational benchmarks come from product usage: what tickets arrived, how fast teams closed them, whether they met the service level. The AI adoption percentages come from a survey of customers who chose to answer. One measures behaviour, the other measures stated intent, and mixing them in a single release invites a reader to treat 56.4 per cent adoption and 45 minutes 18 seconds resolution as though they came from the same instrument. They did not.
The 95 per cent claim is a definition, not a finding
The headline deserves a caveat. Freshworks defines trendsetters as the top 20th percentile of organisations in its own dataset by performance, then reports that this group outperforms the median. That is what the definition guarantees. The useful information is the size of the gap, not the fact of it, and the gap is large: a twelvefold difference in resolution time across the APAC sample is a real operational spread between well-run and poorly-run teams. Presenting it as a discovery rather than as a definition invites a reader to draw a causal conclusion the data cannot support.
The deflection gap
The 83.1 per cent average AI agent deflection rate is where the report and the earnings calls part company. In its second-quarter 2026 call, Freshworks told investors that Freddy AI agent deflection rates averaged 50 per cent and reached as high as 80 per cent for mature deployments. For the fourth quarter of calendar 2025 it reported deflecting more than 50 per cent of tickets. The benchmark report’s 83.1 per cent average is therefore well above the 50 per cent average the company reports to the market, and closer to the top of the range it describes to investors as a mature-deployment ceiling.
A definitional difference could explain it. Deflection can mean the share of conversations an agent closes without human involvement, or the share of contacts it resolves, or the share of eligible intents it handles, and the denominators differ. Freshworks does not publish which one the benchmark uses. But a company reporting 50 per cent to investors and 83 per cent to prospects in the same year owes both audiences the same definition, and until it provides one a buyer should treat the higher number as the ceiling of a favourable subset rather than as a typical result.
The year-over-year jumps
The size breakdown sharpens the question. Freshworks reports AI agent deflection at 81.4 per cent for small businesses, 86.2 per cent for mid-market and 91.5 per cent for enterprise in 2026, against 48, 55 and 56 per cent respectively in 2025. Every segment roughly doubled in a year, and enterprise now sits above 91 per cent. That would mean the largest support organisations hand almost every conversation to an agent, which sits awkwardly beside the report’s own second trend: that AI plus humans is the winning model and that the best teams use AI for speed and volume while people handle judgement. Both can hold only if enterprise deflection means something narrower than a reader would assume.
A CX report from an EX company
The commercial context is worth stating because it explains why the report exists in this form. Freshworks is now an employee-experience company by revenue. In its second-quarter 2026 results, EX annual recurring revenue reached 567 million dollars growing 24 per cent, while CX ARR reached 400 million dollars growing 4 per cent, with management guiding to low single digits for the full year. At its IPO, employee experience was 35 per cent of total recurring revenue; the company expects it to exceed 60 per cent by the end of 2026. CX has been reorganised for profitability and steady-state growth, with operations consolidated in India.
So the benchmark report is a marketing asset for the smaller, slower-growing half of the business. That does not make its data wrong, and the operational benchmarks are worth having. It does mean the AI figures carry a promotional load they would not carry if CX were growing at 24 per cent, and a reader should weigh them accordingly.
The industry-wide measurement problem
The deeper issue is that nobody in this market counts resolution the same way, and the spread of published claims is wide. Intercom says its Fin agent averages a 76 per cent resolution rate across more than 12,000 customers. HubSpot has reported an average of 55 per cent. Freshworks tells investors 50 per cent and publishes 83 per cent. Pricing is moving in the same direction as the metric, with Intercom charging per outcome and Zendesk billing AI agents per automated resolution, which puts commercial weight on a definition that has no standard.
Gartner expects AI agents to resolve 80 per cent of common service issues without a person by 2029, and that forecast will be met or missed depending entirely on which denominator the industry settles on. It is worth noting that the vendors’ published claims already cluster near that figure. A forecast that becomes a target, and then a claim, is a familiar sequence in enterprise software, and the definitional problem is what allows the sequence to run.
For a support leader the useful exercise is narrow and worth doing. Take the three operating metrics that matter to any queue, resolution time, first contact resolution and service-level attainment, and place them against the regional median and the trendsetter figure. If you sit at the median on all three, the gap to the top quintile is the size of the improvement available, and the report’s industry sections suggest where the constraint usually sits: routing, knowledge quality or channel mix. That is more valuable than any single percentage in the release, and it does not depend on the AI figures being right.
What to watch
Three markers will show whether the benchmark holds. The first is whether Freshworks publishes the deflection definition and reconciles 83.1 per cent with the 50 per cent it reports to investors, because the gap is answerable in a footnote. The second is the 2027 edition, since a second year of year-over-year data will show whether the doubling was a step change or a change in how the metric is counted. The third is buyer behaviour, because per-outcome pricing turns resolution rate into a contract term and customers will start auditing it.
What it signals
The India framing deserves a note before the conclusion. The release leads on Indian teams resolving faster, but the benchmark table behind that claim is Asia-Pacific, and the median it compares against is an APAC median rather than an India median. Indian support operations form a meaningful subset of that sample rather than the whole of it, so a reader comparing an Indian team against the published figures is measuring against a regional reference, not a national one. The announcement marks the point where AI deflection has become the number a customer service vendor leads with, and where that number has outrun the definitions behind it.
The report’s real contribution is the operational benchmark: resolution time, first contact resolution and service-level attainment, measured against a regional median, tell a support leader something true about their own operation. The AI figures tell them something about the market’s marketing. Both are worth reading, and only one of them is worth budgeting against until the industry agrees what a resolved ticket is. That agreement will arrive through procurement rather than through research.