Future of Work & Skills Observatory · Customer Service & Customer Experience
How Will AI Change Customer Support Jobs by 2030?
Which service tasks may be automated, where empathy and accountability remain essential, and how can support teams build the right skills? With a 90-day plan.
hilighter Editorial Team · Published by SkillsBoard SAS, the company behind hilighter
Published · Reviewed
How will AI change customer support in your organization?
Start from your company: the free overview ranks your departments by AI exposure, then shows which tasks of your support roles may change and which skills to build.
Free and indicative. No lead form is required to read this report.
Customer support is among the clearest examples of AI changing parts of a job, not just its tools. A customer may receive an instant answer to a straightforward order question while a human agent handles the complicated case that follows. Behind the scenes, AI can retrieve help-center articles, summarize conversations, translate responses and draft tickets. But fluent language is not the same as a correct solution.
The central question is not “Will AI eliminate support agents?” It is “Which issues should AI resolve, which require human judgment, and how should the skills of the service team evolve?” The right answer differs sharply between retail returns, technical troubleshooting, financial complaints, medical devices and emergency situations.
The strongest service strategy is not “maximum bot deflection.” It is reliable resolution with clear ownership, a fast route to a person, accessible service and outcomes that customers actually consider fair. This report presents task-level analysis, emerging skills and a practical 90-day workforce plan.
Apply these findings to your own organization: Explore AI Workforce Readiness with hilighter. This resource is freely readable, with no gated report or lead capture, and the assessment is currently in English. The AI readiness journey is a planning aid, not an individual employee rating or regulatory certification.
Quick answer: what changes, what matters, and what to do
| Work dimension | Likely direction | Practical response |
|---|---|---|
| Routine questions | AI may answer grounded FAQs and basic order questions rapidly | Provide correct source content, easy escalation and meaningful disclosure |
| Ticket administration | Classification, routing and concise handover notes may become more automatic | Verify records, permissions and follow-through |
| Live-agent assistance | Suggested answers and troubleshooting steps can support agents | Teach verification, empathy and appropriate refusal |
| Complex or vulnerable cases | High stakes and emotional nuance keep humans central | Preserve trained human escalation and accessibility |
| Service quality and learning | AI can surface patterns for knowledge improvements | Audit actual resolution, repeat contact and unequal outcomes |
Important: the classifications in this article are original qualitative editorial judgments. They are not measured replacement probabilities, numerical risk scores or predictions of headcount reductions.
1. What the evidence actually supports
A frequently cited field study by Brynjolfsson, Li and Raymond, published in The Quarterly Journal of Economics in 2025 after the NBER working-paper version, studied AI assistance in customer support and found a productivity increase of about 14% on average in that setting, with larger gains among less experienced workers. This is evidence for a particular assistance system and workplace, not a claim that all service teams gain 14%, or that 14% of jobs disappear. [1]
Salesforce’s seventh State of Service, based on a 2025 survey of 6,500 service professionals across 40 countries, reports increased use of predictive, generative and agentic AI. Respondents expected a larger share of cases to be handled by AI by 2027, but that is a forecast by service teams, not observed future performance. [2]
Zendesk’s CX Trends 2026 identifies rising expectations for fast answers, cross-channel continuity and explainability. It reports, for example, that 74% of consumers in its research expected 24/7 service because of AI. Treat these as survey findings from a customer-service technology provider, not a universal consumer law. [3]
The ILO’s 2025 task-exposure work warns against equating occupational exposure to job loss; it is common for only some tasks to change. WEF projections on skill change are broad employer expectations and should not be copied onto a specific support role. [4][5] Microsoft’s 2026 research emphasizes that work redesign, management practices and learning systems shape outcomes beyond access to AI tools. [9]
The evidence favors testing assisted support and controlled automation with real cases. It does not justify a blanket promise that customers prefer bots, that agents can be eliminated without quality loss, or that an AI “resolved” ticket represents a solved problem.
How to read the evidence
Large industry surveys describe respondents’ adoption, expectations and perceptions. Controlled experiments offer stronger causal evidence but often test one tool in one workflow. Task-exposure studies identify possible technical overlap, not adoption, safety, business value or actual job displacement. hilighter does not convert any of these into a single percentage of a profession “at risk.”
2. Define the profession before assessing it
A job title is an imperfect proxy for what employees do. The relative importance of each task changes with company size, business model, geography, customer segment, technology stack and seniority. Analyze your own tasks before choosing training or software.
- Tier 1 / frontline support agent: Handles frequent general inquiries where accurate FAQ retrieval, account validation and good escalation rules have the greatest potential.
- Technical support specialist: Troubleshoots configurations, integrations and incidents; AI can retrieve steps but wrong instructions may create operational or security harm.
- Customer care / complaint handler: Deals with distress, complaints, disputes and complex service recovery. Fairness and emotional intelligence remain central.
- Multilingual contact center agent: Serves across languages and channels; translation helps but tone, cultural context and regulated messages need quality checks.
- B2B support or customer success support: Works with contractual SLAs, business continuity and account stakeholders. Context, commitments and precise handovers matter.
- Support team lead / quality analyst: Coaches staff, designs escalation policies, monitors root causes and evaluates whether AI improvements are genuine.
3. Task-by-task AI impact map
How to read the map: each task has two separate ratings. Change potential: High workflow change means relatively repeatable activities that may be assisted or partially automated with good data and review; Medium / conditional means useful AI help depends strongly on context and systems; Human-led / consequential means tools may prepare inputs, but accountable human control remains central. Risk and trust is a separate dimension (Low, Medium, High, Critical): it addresses privacy, accuracy, rights, fairness and the cost of an error. Neither rating is a probability.
| Customer Support Specialist task | Change potential | Risk and trust | What AI can contribute | What people must check |
|---|---|---|---|---|
| Answer repetitive FAQ questions | High workflow change | Medium | Suggest grounded answers from versioned and authorized knowledge articles. | Display supporting sources; escalate unknown, conflicting or sensitive requests. |
| Tag, triage and route incoming tickets | High workflow change | High | Classify intent, urgency and product with confidence indicators. | Audit false priority, accessibility needs and potentially harmful misrouting. |
| Summarize conversations and handover context | High workflow change | High | Draft brief, structured case histories across channels. | Check omitted facts, privacy, consent and human handover quality. |
| Translate standard customer messages | High workflow change | Medium | Draft multilingual responses and terminology-consistent templates. | Check local legal terminology, cultural tone and safety-critical instructions. |
| Resolve low-risk order-status and policy questions | High workflow change | High | Retrieve account-specific facts through scoped, verified integration. | Confirm authentication, actual status and limits on account action. |
| Suggest replies to live agents | Medium / conditional | Medium | Offer ranked knowledge passages and draft responses. | Agent checks truth, empathy, completeness and any commitments. |
| Guide routine product troubleshooting | Medium / conditional | High | Propose diagnostic steps from validated manuals and logs. | Stop when evidence is ambiguous or action could cause damage or injury. |
| Support return, refund and warranty cases | Medium / conditional | High | Present policy options and prepare transaction drafts. | Human handles disputes, exceptions, high value and legally significant rights. |
| Analyze sentiment and recurring feedback themes | Medium / conditional | High | Cluster cases and suggest knowledge-base updates. | Check bias, sarcasm, multilingual errors and low sample counts. |
| Maintain knowledge-base drafts | Medium / conditional | Medium | Suggest missing articles and contradictions from case patterns. | Qualified owner approves, versions and retires outdated instructions. |
| Handle safety, fraud, vulnerability or crisis reports | Human-led / consequential | Critical | Surface relevant guidance and priority routing for trained responders. | Empowered person investigates, decides and follows safety protocols. |
| Resolve complex complaints and rebuild trust | Human-led / consequential | Critical | Compile chronology, contractual facts and possible options. | Person listens, uses discretion and owns resolution and remedies. |
Five workflows worth examining first
FAQ automation depends on knowledge quality. If the refund policy is outdated, a confident bot can scale the wrong answer to thousands of customers. Ground responses in approved material, identify content owners and test exceptions rather than judging success by low response time alone.
Escalation is a product feature, not a failure. Good automation knows when to stop. Customers must reach a competent person if an answer is uncertain, an issue repeats, or emotional, safety or accessibility needs arise. The handover must carry relevant context without unnecessarily exposing personal data.
AI assistance changes what agents practice. New agents may learn procedural shortcuts from suggestions but risk becoming less capable at independent troubleshooting. Train agents to challenge a proposed answer and explain their reasoning. Regular no-assist exercises can preserve foundational competence.
Agent actions require stronger controls than drafting. A bot that drafts refund advice is not equivalent to one that actually issues refunds or disables an account. Differentiate read access from external action, define approval thresholds and create a rollback path.
Quality metrics must follow customer outcomes. An automated response that causes a second contact is not necessarily a successful resolution. Measure repeat contacts, verified first-contact resolution, complaint escalation and customer effort alongside speed and operating cost.
4. Skills and work habits that may lose relative value
Important distinction: a declining need for a routine activity does not make an employee or the underlying profession obsolete. Many foundational skills remain necessary to catch AI errors. These are activities whose standalone market value could decrease, not skills people should stop learning entirely.
- Memorizing standard FAQ wording: Retrieval can supply consistent approved copy; contextual interpretation and detecting exceptions remain valuable.
- Copy-pasting ticket templates: Standardized replies are easier to draft; customer understanding and accuracy checking still differentiate service.
- Manual ticket tagging: Classification can be assisted, though critical misrouting requires oversight and correction.
- Reconstructing prior conversations manually: Summarization can help if records are permissioned and key commitments are verified.
- Simple status lookups: Secure integrations can retrieve basic account facts; authentication and fraud checks remain essential.
- Speed-only service targets: When AI increases throughput, resolution quality, fairness and trust become more important than a short handling-time number.
5. Eight priority skills to develop
The useful standard is not whether someone has attended an AI webinar, but whether they can demonstrate a safe, repeatable outcome on relevant work. The suggested evidence below is illustrative and should be adapted by employees and managers.
1. Source-grounded answer verification
Distinguish an approved policy or knowledge article from a plausible but unsupported response. Evidence of competence: A set of corrected responses with citation and freshness checks. First practice: Audit 20 AI-generated support replies against the official help center.
2. Escalation and exception judgment
Identify when automation should stop because of harm, legal rights, vulnerability or uncertainty. Evidence of competence: An escalation decision tree tested against ambiguous and urgent cases. First practice: Classify ten challenging anonymized ticket scenarios and explain each handoff.
3. Empathy and complex service recovery
Listen to emotional and commercial context when a case cannot be solved by a standard answer. Evidence of competence: Observed role-play with fair explanation, de-escalation and ownership. First practice: Practice an upset customer dispute with an independent coach.
4. Service workflow and AI-agent supervision
Understand intent classification, retrieval, permission boundaries, external actions and rollback. Evidence of competence: An approved flow diagram with human takeovers and audit events. First practice: Design a controlled order-status bot that cannot issue refunds.
5. Privacy and secure customer-data handling
Recognize sensitive identifiers and avoid disclosing information across customers or channels. Evidence of competence: A least-privilege permission review and redacted case examples. First practice: Run a privacy exercise with misdirected customer-account information.
6. Service quality analytics and experiment design
Interpret deflection, re-contact and resolution metrics without confusing cost reduction with customer benefit. Evidence of competence: A pilot dashboard with baseline, sampling rules and quality guardrails. First practice: Recalculate first-contact resolution when repeat contacts are included.
7. Knowledge-base curation and content governance
Improve article coverage, versioning, source attribution and ownership as a core service capability. Evidence of competence: An audited set of revised articles with approval and expiry owners. First practice: Fix five obsolete troubleshooting articles and regression-test AI answers.
8. Cross-channel communication and accessibility
Translate clearly and preserve customer context across text, voice and image while respecting accessibility needs. Evidence of competence: An accessible resolved case across two channels with complete handover. First practice: Test a multilingual ticket with accessibility and language-quality review.
6. Emerging responsibilities already becoming plausible
These are work responsibilities, not necessarily universally established job titles. In a small business, one employee may combine several of them.
- AI service-response quality auditor. Tests answer grounding, critical failures, policy exceptions, multilingual performance and repeated customer contacts. Maturity: emerging responsibility, not a standardized occupation.
- Human–AI escalation designer. Builds reliable transition rules for bot-to-human cases with context and clear ownership. Maturity: emerging responsibility, not a standardized occupation.
- Customer knowledge provenance steward. Connects support content to source owners, revision dates, authorized claims and retrieval quality. Maturity: emerging responsibility, not a standardized occupation.
- Service action-automation guardian. Sets permissions and monetary thresholds before bots can process returns, cancel accounts or trigger workflows. Maturity: emerging responsibility, not a standardized occupation.
- AI-era customer experience recovery analyst. Studies failed automations and customer frustration to improve service design rather than conceal escalations. Maturity: emerging responsibility, not a standardized occupation.
7. Future skills that may appear by 2028–2030 (hypotheses, not facts)
The following are scenario hypotheses. Their labels, timing and qualitative confidence are hilighter editorial assessments. They have not been independently observed at scale and should not be used to promise a career outcome.
- Multimodal service reliability assurance. Medium confidence (conditional, 2028–2030). Test whether AI handles images, voices and text consistently, especially for vulnerable users. What would validate it: Formal quality tests across languages and modalities become routine.
- Customer-side AI agent negotiation. Low–medium confidence (conditional, 2028–2030). Handle requests originating from customer-owned shopping or service agents and verify their authority. What would validate it: Recognized identity and delegation mechanisms appear in customer interactions.
- Service autonomy-boundary design. Medium confidence (conditional, 2028–2030). Define which transactions can be executed independently and which need manual authorization. What would validate it: Widely adopted transaction policies and auditable automated-action controls.
- Customer answer provenance and appeal design. Medium confidence (conditional, 2028–2030). Produce understandable evidence trails for AI-issued service explanations and complaints. What would validate it: Customer-service tools expose usable evidence and appeal routes by default.
- Dynamic complexity and vulnerability routing. Low–medium confidence (conditional, 2028–2030). Adapt handovers to customer vulnerability and uncertainty without harmful profiling. What would validate it: Privacy-preserving and independently validated escalation research.
8. A real workflow: before and after thoughtful AI adoption
This example is a process-design illustration, not a live hilighter customer result or a promise of cost savings. Teams should compare it with their baseline and protect quality, not simply reduce handling time.
An order problem that becomes a complex complaint
| Stage | Conventional approach | AI-assisted approach | Control that must remain |
|---|---|---|---|
| 1. Customer contacts support | Agent opens a ticket and asks for the order number. | Secure intake checks intent and offers a human option. | Avoid premature authentication assumptions. |
| 2. Retrieve order data | Agent searches systems and policy pages. | Scoped retrieval returns status and relevant policy. | Check identity, accuracy and access rights. |
| 3. Propose initial answer | Agent uses standard wording. | Grounded assistant drafts response with source links. | Agent or bounded bot checks whether exception applies. |
| 4. Identify exception | Customer repeats explanation to another agent. | System detects mismatch and proposes escalation. | Human decides severity and routing. |
| 5. Hand over context | Agent reviews long chat history. | AI produces a short factual handover. | Confirm missed promises and redact extra identifiers. |
| 6. Resolve dispute | Human decides remedial action. | AI supplies verified policy options. | Empowered person owns final remedy. |
| 7. Improve knowledge | Issue might recur unnoticed. | Anonymized failure theme goes into article backlog. | Knowledge owner validates fix and tests outcomes. |
In this scenario, the goal is correct resolution with preserved context. Ticket deflection is a secondary metric. A customer transferred to a human for the right reason is not a failure; a hidden unresolved problem is.
9. How the profession could evolve in 2026, 2028 and 2030
| Time horizon | Reasonable scenario | What remains uncertain |
|---|---|---|
| 2026: assisted agents | Grounded knowledge retrieval, summaries and limited bot answers become common in some service environments. | Data silos and difficult exceptions limit scale. |
| 2028: bounded action automation | Some teams enable carefully authorized refund or order actions and multimodal support. | Fraud, privacy, costly mistakes and governance can slow expansion. |
| 2030: complex-case specialization | Human agents may spend more time on exceptions, retention, vulnerabilities and service design. | Customer preferences, channel access and real costs remain uncertain. |
A 2030 scenario is not a forecast. Some sectors will adopt slowly because of legacy systems, customer expectations, regulation, data access or economics; others will redesign workflows quickly. A cautious employer should plan for multiple paths rather than a single predicted future.
10. Governance, trust and employee safeguards
- Tell customers about qualifying direct AI interactions: apply relevant AI Act Article 50 disclosure rules, including context-specific exceptions, rather than pretending a bot is a person. [6]
- Protect customer data: apply least-privilege access, purpose limitation, security controls and appropriate retention to transcripts, accounts and attachments. [7]
- Make escalation genuinely available: implement handoff for uncertainty, repeat contact, accessibility or high-consequence problems; preserve context.
- Separate answers from actions: reading order status, changing an account and issuing a refund require different authorization levels.
- Test failure patterns: include hallucinations, contradictory policies, wrong customer matches, abusive inputs and unsupported languages.
- Respect vulnerable customers: avoid reducing sensitive cases to sentiment scores or opaque profiling; use trained human responders.
- Retain professional competence: train employees to solve cases without AI suggestions and verify tool output independently.
- Measure real resolutions: monitor repeat contacts, reopens, fairness across cohorts and customer complaints; publish clear ownership for errors.
Legal note (EU context): AI Act Article 50 includes transparency rules for certain direct interactions with AI and specified synthetic-content scenarios, with exceptions and context-specific obligations [6]. The GDPR and national rules continue to apply to personal-data processing; in sales and marketing, ePrivacy and local direct-marketing laws may independently restrict outreach [7]. A website article is not legal advice. Verify the rules relevant to your specific workflow and jurisdiction.
11. A practical 90-day action plan for teams
| Phase | Team actions | Individual learning activity | Evidence / go–no-go test |
|---|---|---|---|
| Days 1–30: discover failures | Map case types, top knowledge gaps, handoffs, permissions and privacy constraints. Baseline first-contact resolution, repeats and complaints. | Train agents in verifying AI answers against policy sources. | Approved no-automation list, baseline quality audit and escalation owner. |
| Days 31–60: supervised pilot | Pilot article retrieval and response suggestions on low-risk, reversible cases. Use shadow mode for action workflows. | Test 20 difficult cases including multilingual exceptions and safety escalations. | Correct-source rates, independent quality checks, no loss of human access. |
| Days 61–90: scale or stop | Review repeat contact, customer effort, staff capability and failure modes. Improve outdated knowledge content. | Coach agents on complex-case empathy and escalation decisions. | Expand only with customer-quality evidence and audited access controls. |
Simple organizational scorecard
| Metric | Why it matters | Common measurement error |
|---|---|---|
| Verified first-contact resolution | Tracks resolved problems, not simply a bot answer. | Counting abandoned or prematurely closed chats as resolved. |
| Repeat contact within 7 days | Catches unsolved issues returning through other channels. | Not linking conversations across channels appropriately. |
| Escalation appropriateness | Tests whether human handovers occur for the right reasons. | Labeling every handoff as bad deflection. |
| Knowledge-grounded answer accuracy | Tests whether guidance matches approved sources. | Using model self-confidence as proof of truth. |
| Customer effort and satisfaction | Tracks the real experience of obtaining a solution. | Sampling only people whose tickets were closed. |
| Agent learning and independent competence | Ensures assistance does not erode foundational judgment. | Counting completed courses without observed ability. |
Review results across case types, languages, customer groups and seniority levels rather than treating a single average as proof of improvement. Include a baseline, independent spot checks, escalation quality and an explicit rollback plan.
12. Frequently asked questions
Will customer support agents disappear because of AI?
No reliable universal estimate supports that conclusion. Routine answers and administrative work are exposed, while complex cases, trust and service recovery can expand in importance. [4]
Does the famous 14% productivity finding apply to my team?
Not automatically. The study evaluated a particular AI-assistance deployment; benefits varied across workers and settings. Pilot with your own issue mix and quality checks. [1]
Can AI answer customer questions without a person?
Low-risk, well-grounded questions may be suitable for carefully tested automation. High-consequence or ambiguous issues need escalation and other safeguards.
What new skill matters most for frontline support?
Grounded answer verification combined with knowing when to hand a customer to a trained person.
Should we optimize for ticket deflection?
No. Deflection is meaningful only when the underlying problem is solved. Track verified resolutions, repeated contacts, complaints and customer effort.
Can a bot issue refunds?
Only within clearly authorized, verified and auditable boundaries, including local consumer-law and data-security requirements. Human review should cover exceptions and high-value cases.
How can support managers retrain employees?
Map case types, practice exceptions and empathy, train on knowledge verification and run controlled pilots with regular human quality reviews.
How can hilighter help with service skills?
Use the organization-level AI Workforce Readiness experience to explore your actual service tasks, identify likely capability gaps and plan training with employee and manager validation.
13. What organizations should do next
Customer support is moving from answer production toward resolution ownership. The organizations that do well will improve the knowledge base, build safe handoffs and develop staff capable of handling ambiguous and sensitive cases. They will resist the temptation to measure an automated conversation as successful merely because it ended.
For managers, the immediate priority is a task map and a small, quality-controlled pilot. For employees, it is learning to verify sources, detect limits, escalate appropriately and build the empathy and judgment required where scripts are insufficient.
Ready to examine your actual jobs rather than a generic profile? Explore your organization’s AI Workforce Readiness with hilighter. Identify tasks, document current skills, validate capability gaps with managers and employees, and turn the results into a practical development plan. The assessment is currently in English. It does not automatically preselect a role and does not provide an official legal-risk rating.
Methodology, limitations and editorial disclosure
This is a research-informed, scenario-based editorial resource, created for the hilighter Future of Work & Skills Observatory (SkillsBoard SAS). Task categories and future skill confidence labels are editorial classifications. The report is not a representative employer survey, an assessment of any particular employee, a labor-market forecast, a technical benchmark of AI products, legal advice or a guarantee that a job will or will not disappear. The sources include independent research, vendor-sponsored surveys and regulatory guidance. Those evidence types are not interchangeable. Any vendor-reported percentage is attributed to the respective survey, with its denominator where available. Publication date: 6 November 2026; verify material new developments before future updates.
Editorial maintenance: refresh source links, legal interpretations and examples at least annually, and whenever a material regulatory or technological change occurs. Identify substantive edits and their dates rather than silently changing the reported evidence.
Sources and further reading
- Brynjolfsson, Li & Raymond / NBER and The Quarterly Journal of Economics, Generative AI at Work: evidence from customer-support assistants (2023 working paper / 2025 published study). https://www.nber.org/papers/w31161
- Salesforce, State of Service, Seventh Edition (13 November 2025), survey of 6,500 service professionals. https://www.salesforce.com/service/resources/state-of-service-report/
- Zendesk, Customer Experience Trends 2026 (2026). https://cxtrends.zendesk.com/gb
- International Labour Organization, Generative AI and Jobs: A Refined Global Index of Occupational Exposure (20 May 2025). https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure
- World Economic Forum, The Future of Jobs Report 2025 (2025). https://www.weforum.org/publications/the-future-of-jobs-report-2025
- European Commission, Guidelines on transparency obligations for providers and deployers of AI systems (Article 50) (2026). https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations
- European Commission, Information for individuals: automated decisions and profiling (accessed November 2026). https://commission.europa.eu/law/law-topic/data-protection/information-individuals_en
- NIST, AI Risk Management Framework: Generative AI Profile (26 July 2024). https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence
- Microsoft, 2026 Work Trend Index: Agents, human agency, and opportunity (5 May 2026). https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization
- NBER, Shifting Work Patterns with Generative AI (revised November 2025). https://www.nber.org/papers/w33795