The loudest version of the AI story says that large language model agents will replace customer service. The evidence supports a more useful conclusion: adoption is rising quickly, but most organizations are still learning how to combine AI with people, trusted knowledge and workflow controls.
Salesforce’s 2025 State of Service survey found that 69% of service professionals said their organization used at least one form of AI and 39% said it used agentic AI. Respondents expected the share of cases resolved by AI to rise from 30% in 2025 to 50% in 2027. These are survey findings and expectations, not a guarantee that half of every company’s cases will be autonomously resolved.
The safe boundary shifts with ambiguity consequence and authority
|
Workflow |
AI contribution |
Human control |
Default mode |
|
FAQ answer |
High |
Low |
AI resolves within approved knowledge |
|
Thread summary |
High |
Low |
AI prepares for person or workflow |
|
Suggested reply |
High |
Moderate |
Person verifies and sends |
|
Routine action |
Moderate |
Moderate to high |
Narrow authority and audit trail |
|
Exception or dispute |
Low |
High |
Human decides and owns outcome |
A risk based human control framework for service workflows
Four different meanings of AI in customer service
1. Classification. AI detects intent, topic, language, urgency or sentiment and applies a tag or routing suggestion.
2. Assistance. AI summarizes a thread, retrieves relevant knowledge, translates text or drafts a reply for a person to review.
3. Resolution. AI answers the customer directly when the request is within approved knowledge and policy boundaries.
4. Action. An agentic system changes something: resets access, updates an address, issues a credit, reschedules a service or calls another system. This requires the strongest identity, authorization, audit and exception controls.
Calling all four “AI agents” hides important differences in risk. A suggested email and an autonomous refund are not the same deployment.
Workflows being adopted now
|
Workflow |
AI contribution |
Human role |
Good first use? |
|
Topic and intent categorization |
Suggest tags, priority and route |
Correct exceptions and refine taxonomy |
Yes |
|
Thread summarization |
Compress long histories and extract open actions |
Check missing commitments or facts |
Yes |
|
Knowledge retrieval |
Surface FAQ, policy and similar resolved cases |
Choose the applicable source and detect outdated content |
Yes |
|
Reply drafting and translation |
Create a first draft in the right tone and language |
Verify facts, promises and customer-specific nuance |
Yes |
|
Conversation quality review |
Flag tone, completeness, policy and compliance risks |
Coach agents and adjudicate false positives |
Yes, with calibration |
|
Self-service FAQ answers |
Respond from approved, current knowledge |
Own the content, test failure modes and provide escape to a person |
Yes, for bounded topics |
|
Order or account enquiries |
Retrieve authenticated status and explain next steps |
Handle identity failures and exceptions |
After integration controls |
|
Routine system actions |
Execute an approved, reversible action within limits |
Set authority, monitor logs and manage exceptions |
Later stage |
|
Complaints and disputes |
Summarize evidence and propose options |
Listen, negotiate, decide and protect the relationship |
Human led |
|
Financial, legal, safety or regulated decisions |
Retrieve policy and prepare a case file |
Make or approve the consequential decision |
Human controlled |
What the adoption statistics actually say
- Salesforce 2025: 69% reported use of at least one form of AI in service; 39% reported agentic AI use. AI use was associated in the survey with more complex human caseloads, and 82% agreed complex cases are best resolved by humans and AI together.
- McKinsey 2025: 88% of respondents reported regular AI use in at least one business function, 62% said their organization was at least experimenting with AI agents, and 23% said an agentic system was being scaled somewhere in the enterprise. In any single function, no more than 10% reported scaling agents. This is a useful correction to the hype.
- Gartner 2024 survey for 2025 plans: 85% of customer service leaders expected to explore or pilot customer-facing conversational generative AI. For voicebots specifically, 44% were exploring, 11% piloting and 5% deployed. Gartner also found 61% had a backlog of knowledge articles to edit.
- Zendesk 2025 trends: 75% of CX leaders expected 80% of interactions to be resolved without human intervention within a few years. This is an expectation reported by a vendor survey, not current resolution performance.
Industries where agent use is showing up first
McKinsey’s 2025 global survey reported the widest use of AI agents in technology, media and telecommunications, and healthcare. Across AI generally, media and telecommunications, insurance and technology were the leading industries in reported use. The most common service-related uses included information capture and delivery through conversational interfaces and contact-centre or customer service automation.
The operational pattern explains the lead: technology and telecom have digital service volumes and structured product data; insurance has document-heavy, repeatable workflows but requires controls; healthcare has high administrative demand but also high consequence, making supervised workflows more credible than unrestricted autonomy. Retail and ecommerce are also prominent vendor use cases, especially for order status, returns and product recommendations.
Where geography appears to matter
There is no single public 2025 dataset that compares customer-service-agent deployment by both industry and geography using one definition. Geography claims should therefore remain cautious. Broader enterprise evidence nevertheless shows strong APAC momentum. Forrester reported that APAC companies were more likely than North American or European peers to invest US$400,000–500,000 in AI, and that Singapore, Australia, New Zealand and South Korea ranked among the leading countries in an external AI usage index. Adobe’s Asia study found practitioners reporting returns from chat and customer support as one of several generative-AI use cases, while IBM’s APAC outlook placed customer experience first among 2025 AI investment priorities at 21%.
These studies do not prove that APAC has the largest installed base of customer service agents. They show investment appetite and relevant use cases. North America remains a major centre of product supply and operational experimentation; Europe’s tighter regulatory environment increases emphasis on governance, transparency and human oversight. Any headline claiming a geographic winner should state the metric used.
The human intervention test
Require a person when one or more of the following is true: the relevant facts are incomplete; sources conflict; the requested action changes money, access, rights or safety; policy permits discretion; identity cannot be established; the customer disputes the organization’s position; the issue could create legal or reputational exposure; or the organization has not defined who is accountable for the outcome.
Human intervention should not mean starting again. The AI should pass the complete thread, customer identity state, sources used, actions already attempted, confidence or uncertainty, and a recommended next step. A handoff that loses context is an automation failure even if the routing itself succeeds.
The implementation sequence for a lean team
1. Build the knowledge foundation. Clean FAQs, define authoritative sources and capture useful prior resolutions. Gartner’s knowledge backlog finding is a warning: an LLM cannot compensate for unmanaged knowledge.
2. Assist before acting. Begin with classification, summaries, retrieval and drafts. Measure corrections and failure patterns.
3. Automate bounded resolutions. Select high-volume, low-consequence intents with clear policies and an obvious human escape route.
4. Authorize actions narrowly. Use identity checks, transaction limits, reversible steps, logs and explicit exception ownership.
5. Measure outcomes, not deflection alone. Track recontacts, reopened issues, complaints, correction rates, escalations, customer effort and successful resolution—not just how many conversations avoided a person.
What this means for Querylah
Querylah’s opportunity is to make the human–AI service loop practical for a small team. The valuable foundation is already the operating model: communication, searchable FAQs, prior resolutions, ownership and specialist collaboration in one place. AI can then categorize issues, surface similar cases, summarize histories and prepare replies. Autonomous answers or actions should be introduced by topic and risk, with visible sources and controlled handoff—not as a single “turn on the agent” feature.
Sources
Salesforce State of Service Seventh Edition - AI use, agentic AI, case-resolution expectations, top use cases and human–AI collaboration. Source
McKinsey State of AI 2025 - Global enterprise adoption, agent experimentation and scaling by function and industry. Source
Gartner Customer Service GenAI Survey - Plans for customer-facing conversational GenAI and knowledge-management barriers. Source
Zendesk 2025 CX Trends - Consumer and CX leader expectations for copilots and autonomous service. Source
Forrester APAC AI Adoption - Regional enterprise AI investment and adoption patterns; not customer-service-specific. Source
Adobe 2025 AI and Digital Trends Asia - Asia use cases and reported ROI, including chat and customer support. Source
IBM APAC AI Outlook 2025 - APAC AI investment priorities and human-centred agentic workflow outlook. Source
