CS AI Copilots: The Competitive Landscape by the Numbers, Spring 2026
Customer support AI copilots have moved completely past the "FAQ bot" phase of pre-2024. By the second half of 2025, a deflection rate above 50% became the baseline that leading vendors were expected to clear. Deflection rate — the share of inquiries resolved without human agent intervention — is the common industry KPI, and four brands — Intercom Fin 3, Zendesk Copilot, Ada, and Dialpad Ai — are splitting the market along very different strengths.
This post focuses on the results of a comparative evaluation KGA conducted in Q1 2026 across domestic B2B SaaS, e-commerce, and financial services clients, covering deflection rate, human-escalation detection accuracy, CSAT impact, Japanese formal keigo handling, and compliance recording capabilities.
Intercom Fin 3: Top Deflection Rate in the Industry
Intercom has pushed deflection rate sharply upward with each version — Fin 1 in 2023, Fin 2 in 2024, and Fin 3 in Q4 2025. Fin 3 uses a hybrid of Claude 4.5 Sonnet and GPT-4.5, dynamically switching models based on query complexity.
Intercom's published deflection rate for Fin 3 is 67% average across SaaS customers — a leading figure industry-wide. KGA's Q1 2026 measurements across three domestic SaaS companies confirmed an average of 58–64%, close to the published number. Performance scaled with the completeness of help center articles, with environments having 300+ articles regularly exceeding 70%.
Human-escalation detection — the AI's ability to correctly identify when a query should be handed off to a human — is a major evolution in Fin 3. The "Handoff Intelligence" feature integrates user sentiment analysis (frustration detection), query specificity, and historical escalation patterns. In KGA's evaluation, the recall for cases that genuinely needed escalation was 91%, and the false positive rate (unnecessary escalations) was 8%, a substantial improvement from Fin 2's 78% recall.
Zendesk Copilot: Enterprise Fit
Zendesk acquired Ultimate.ai in 2024, integrated it as "Zendesk Copilot" in 2025, and established a clear counter-position to Intercom. For enterprises, Zendesk has an edge — particularly in large contact center operational track records, SLA management, multi-brand support, and granular permission controls, all of which exceed Intercom's capabilities.
Deflection rate came in at 52–58% in KGA's measurements — somewhat lower than Intercom Fin 3 — but Zendesk's differentiation is in the AI Agent Workspace, specifically the quality of AI assistance for human agents. While a human agent is composing a reply, the system displays a real-time next-response draft, relevant knowledge articles, similar past tickets, and a sentiment score in a side panel. Average Handle Time (AHT) reduction of 25–35% exceeds industry benchmarks.
Compliance recording and audit logging is where Zendesk stands alone. In regulated industries — finance, healthcare, government — full call and chat auto-transcription, automatic sensitive data masking (credit card numbers, My Number IDs, phone numbers), and GDPR/HIPAA/PCI DSS-compliant export are all standard. Intercom offers equivalent features as add-on modules; Zendesk has the advantage here.
Ada: Fully Committed to Deep Automation
Since 2024, Ada has positioned itself as an "AI Agent" — moving from simple FAQ answering to automating multi-step transactional processes: returns, subscription changes, reservation modifications. Where Intercom and Zendesk emphasize the copilot model (AI and human agents collaborating), Ada's philosophy is clearly "let the AI Agent complete as much as possible autonomously."
Ada's published deflection rate is 72% industry average — the top figure in the comparison — but this reflects deflection including complex automated business processes, so direct comparisons require care. KGA's measurements at two e-commerce companies came in at 62–68%, on par with or slightly above Intercom Fin 3.
Ada's distinguishing feature is its Reasoning Engine, which lets AI autonomously plan multi-step procedures and call internal APIs in sequence. For a request like "I want to exchange this item for a different size," the agent executes the full sequence: check order history, verify inventory, issue return label, arrange new shipment. Intercom and Zendesk's Fin and Copilot cannot handle this level of process automation at present.
The tradeoff is implementation cost. Ada requires API integration and business workflow design, with a total implementation timeline of 3–6 months and an upfront cost in the range of ¥8–20 million. Compared to Intercom's standard deployment (roughly 1 month, ¥3 million+), Ada is a commitment.
Dialpad Ai: Voice-Specialized Independent Evolution
Dialpad Ai is the only product among the four designed specifically for voice contact centers. For customer touchpoints that are primarily phone-based — insurance, finance, real estate, medical appointment scheduling — it is stronger than Intercom, Zendesk, or Ada.
The technical core is Real-time Transcription + Coaching: during a call, the agent screen shows live transcription, sentiment score, recommended next script, and relevant knowledge articles. Post-call summary, automatic CRM recording, and follow-up task creation are all automated. The Q1 2026 feature "Ai Voice Agent" has reached practical-use level for fully automated responses to simple phone inquiries — business hours, store locations, reservation confirmation.
Compliance recording functionality tailored for financial institutions includes full call recording, automatic masking of sensitive information, and PCI DSS/FINRA/MiFID II-compliant export as standard features. In Japan, the product also complies with recording obligations under the Insurance Business Act, Payment Services Act, and Money Lending Business Act.
Japanese Formal Keigo Handling
Evaluation of all four products on real Japanese-language data revealed clear differences in keigo handling. The most natural Japanese output came from Intercom Fin 3, with kenjogo/sonkeigo usage at a practical level for B2B SaaS. Zendesk Copilot ranked second, followed by Dialpad Ai. Ada, as a US-origin product, retained a translated feel in its Japanese output.
The most critical criterion in practice is preventing inappropriate casual expressions from appearing — controlling output to exclude casual sentence-ending forms like 「〜ですよね」「〜かも」「〜なんです」. Intercom Fin 3's December 2025 Japanese model refresh introduced three tone presets — business formal, customer-friendly, and technical — and this flexibility has been the deciding factor in domestic adoption decisions.
Measured CSAT Impact
In KGA's Q1 2026 comparative evaluation across three domestic SaaS companies, two e-commerce companies, and two financial services companies (seven in total), CSAT changes after deployment showed the following patterns. Intercom Fin 3: +4–8 points. Zendesk Copilot: +2–5 points. Ada: +3–10 points (high variance depending on business process fit). Dialpad Ai: +5–8 points on the voice channel.
The primary driver of CSAT gains is "immediate response" and "24-hour availability" — quantifiable elimination of missed-contact opportunity cost, especially for late-night and early-morning inquiries. However, cases where incorrect escalation judgment delays human response actually lower CSAT — tuning Handoff thresholds is the most critical task in the early deployment period.
Practical Guidance for Deployment Decisions
Maximize deflection rate for general B2B SaaS: Intercom Fin 3. Enterprise and regulated industries requiring SLA and compliance: Zendesk Copilot. Automate end-to-end complex business processes: Ada. Phone-primary contact centers: Dialpad Ai. These four axes will keep you from going seriously wrong.
For CX digital transformation at Japanese enterprises, the approach that maximizes deflection rate is: spend the first three months intensively building out help center articles, then deploy the CS AI. Installing AI without article coverage produces no results. For mid-sized companies, an annual budget in the range of ¥10–30 million is realistic.