The AI Examiner — tier-1 customer support, 2026-08-07

Headline

Fin by Intercom markets a 76% resolution rate; an independently tested deployment measured 38%.

Summary

Fin's marketed rate is drawn from easier, high-volume customer segments than the ones production deployments actually face, and the metric underneath it — "resolution" — is defined differently across the industry, with some vendors counting a customer who goes quiet as resolved. The same shortfall shows up at the company level — Gartner is forecasting that half of the companies that cut customer-service staff for AI will rehire by 2027, and Klarna, the most-cited early adopter, already reversed course after resolution quality dropped.

The evidence

The billing mechanism

Tier-1 customer-support AI is sold on three pricing structures: flat per-seat licensing, per-ticket/per-session fees, and per-resolution billing, where the vendor charges only when the AI resolves a conversation end-to-end (fetched, observed — consistent across multiple vendor-comparison sources). Per-resolution is the structure vendors lead with, and published rates cluster tightly: Quickchat $0.50, Fini $0.69, Gorgias $0.60–$1.27, Lorikeet $0.80–$0.95 (chat/email/SMS) and $1.20–$1.50 (voice), Fin by Intercom $0.99, Ada $1–$3.50, Zendesk AI Agents ~$1.50, Salesforce Agentforce $2.00 (fetched, observed — vendor-published rates as compiled by third-party pricing-comparison sites). Per-ticket/session rates run lower: Freshdesk Freddy $0.10–$0.49/session, eesel $0.40/ticket flat (fetched, observed).

The rate a customer is billed depends entirely on how "resolution" is defined, and that definition is not standardized — some vendors count only a conversation an LLM-verification step confirms was addressed; others count any conversation where the customer didn't ask for a human, including one where the customer simply stopped replying (fetched, observed — stated directly in vendor-pricing-comparison coverage).

What vendors claim vs. what gets measured in production

Fin by Intercom publishes the most specific claim found: a 76% average resolution rate across more than 12,000 customers, improving roughly 1 percentage point a month, plus a separately-claimed 93% "accuracy" figure from vendor-selected testing (fetched, self-reported — Fin's own site). Resolution rate and accuracy are different metrics — the first is the fraction of tickets an AI closes without escalation, the metric the per-resolution price is billed against; the second is whether a given AI response was correct. The independent evidence below speaks to resolution rate, the billed metric, not accuracy.

One reviewer ran Fin across 4 small-business clients with a combined volume of 500 tickets/month, tracked over 60 days, and measured a 38% average resolution rate — clients with strong help-center documentation reached 47–52%, clients with sparse documentation reached only 28–31% (fetched, observed — the reviewer's own disclosed test methodology, n=4 accounts / 500 tickets monthly). A separate source puts Fin's production resolution rate at 45–53%, attributing the shortfall to vendor benchmarks being drawn from high-volume, low-variance B2C customer cohorts rather than the messier B2B ticket mixes production deployments actually see (fetched, inferred — this figure could not be independently re-fetched to confirm exact wording; treat as lower-confidence than the 38% figure above).

Either number is well below the marketed 76%. The reviewer who ran the disclosed test attributed the gap primarily to knowledge-base quality rather than to the AI itself: accounts that spent 2–4 weeks cleaning up documentation before launch averaged 12 points higher resolution than accounts that didn't (fetched, observed, same test).

The definition problem is structural, not just Fin's

No vendor in this space publishes a standardized, third-party-audited resolution-rate figure, and "resolution" itself is defined differently vendor to vendor, including definitions that count a customer going silent as success (fetched, observed). That is a market-structure fact, not a limitation of this run. A buyer comparing Fin's 76% against Zendesk's or Salesforce's rate is comparing numbers built on different definitions of the word "resolution," and no independent body checks any of them against production outcomes.

The macro correction

Gartner is forecasting that 50% of companies that cut customer-service staff because of AI will rehire for similar roles — under different titles such as "Solution Consultant" — by 2027, citing AI's difficulty with complex, judgment-requiring cases, declining customer-satisfaction scores, and deployments that prioritized headcount reduction over actually solving customer problems (fetched, observed, dated 2026-02-03). Klarna is the most-cited case: in 2024 the company said an AI agent had replaced roughly 700 customer-service roles and was handling two-thirds of chats with under 2-minute average resolution times; by mid-2025 it was rehiring after customer satisfaction dropped, with its CEO stating "we focused too much on efficiency and cost — the result was lower quality, and that's not sustainable," and has since moved to a hybrid AI-plus-human model (fetched, observed — corroborated across multiple independent outlets). Uber cut 10% of its customer-support organization in 2026, its second such reduction in under two months, citing AI as part of the rationale (fetched, observed).

Adoption figures show the same claimed/observed split at the industry level. Seventy-eight percent of firms surveyed expect agentic AI to run customer support within 18 months; only 16% have deployed it organization-wide, and only 12% of the 98% of contact centers that use some AI describe their strategy as fully optimized (fetched, observed, Adobe 2026 survey as cited in secondary coverage). Separately, 23% of organizations report scaling agents in at least one function and 39% are experimenting, with no single function above roughly 10% of organizations at scale (fetched, inferred — McKinsey figures as cited in secondary coverage, not independently re-fetched).

What people are asking

In a practitioner IT-support community, the live question isn't whether AI can fully replace tier-1 support — it's phrased as whether the role gets hollowed out or evolves into something more strategic as AI gets better at summarizing logs and escalating with context (fetched, observed — direct quote from a community post). Separately, outsourcing firms that supply tier-1 support as a service are reported to be sounding alarm over direct revenue exposure, as client companies bring AI in-house rather than continuing to contract the work out. India and the Philippines are named as the most exposed BPO markets (fetched, observed/inferred — industry-coverage synthesis, not a single practitioner survey).

What this means for a support leader evaluating per-resolution pricing

The number on a vendor's rate card is the vendor's own definition of "resolution," not an audited one. Definitions vary: one vendor requires an LLM-verification step before counting a resolution, another counts it the moment a customer stops replying. For a support leader comparing quotes, that difference changes what the price actually buys — the definition belongs in the contract, not the sales deck.

Methodology & sample

This issue rests on roughly a dozen WebSearch queries and half a dozen WebFetch pulls, all fetched-tier; examiner/sources/ held no files, so there is no operator-supplied data here. Confidence is strongest on the vendor-published per-resolution rates (cross-corroborated across independent pricing-comparison sources) and on the Klarna and Gartner material (each corroborated across multiple independent outlets). It is weaker on the single disclosed production test (n=4 accounts). It is weakest on the 45–53% Fin figure and the McKinsey scaling figures — both came through search-tool summaries, not a direct primary fetch.

Gaps

• The claimed-vs-observed resolution-rate gap is directly evidenced for one vendor (Fin by Intercom) with both a marketed figure and independently-tested production data. Extending that pattern to the wider per-resolution-pricing market is inferred, not confirmed — none of the other eight vendors priced in this issue has a matching independent production test.
• The 45–53% production-rate figure attributed to a second source could not be independently re-fetched (the page returned a 403) to confirm exact wording or methodology; it is carried at lower confidence than the 38% figure, which came from a test with disclosed methodology.
• Gartner's original press release also returned a 403; the 50%-by-2027 figure is confirmed through one secondary outlet that quotes it directly, not through Gartner's own page.
• This issue did not gather vendor count, market-size, or "how they find clients" data — those didn't bear on this issue's question and were skipped rather than filled in for completeness.
• The demand-survey section rests on one community post and secondary industry coverage of BPO exposure, not a broader sample of practitioner forums.