Choose support automation from completed customer work and human effort. Keep assisted productivity, adoption and forecasts separate, and plan a clear route to a person.
What you'll learn
- Check the definition and denominator of resolution rates.
- Count review, corrections and repeat contacts in the trial.
- Use your enquiry mix to set scope and handover rules.
The published Generative AI at Work study reports 15% more issues resolved per hour for staff with access to an AI assistant. It studies human support agents, not an autonomous service desk. Gartner’s 80% of common issues by 2029 is a forecast, not a current measured resolution rate.[1][2]
For your business, the useful question is which enquiry types need less handling time without increasing errors or repeat contacts. Keep productivity, adoption, customer preferences and autonomous resolution separate when comparing a supplier’s numbers.
What do the "resolution rate" numbers really measure?
A resolution rate needs a denominator and a definition of “resolved”. Gartner’s March 2025 forecast concerns autonomous resolution of 80% of common customer-service issues by 2029. It does not say that share of all contacts is already resolved without people.[2]
A supplier may instead count chats that stay inside a bot, tickets not transferred to a person, or customers who stop replying. Those measures can help run a service, but ask how the supplier checks that the customer’s issue was actually settled. A closed chat and a completed refund are different outcomes.
LiveAgent’s January 2026 account illustrates why the denominator matters. It reports 75% resolution across its website chats, including blogs, directories and glossary pages, and 40% on product pages. These are the vendor’s own deployment results, not a market-wide success rate.[12]
Zendesk’s 2026 outlook is different again: 90% of its CX Trendsetter group expect 80% of issues to be resolved without human intervention in the next few years. That describes a group’s expectations, rather than an observed rate across all customer-service teams.[11]
Before comparing resolution percentages, ask: which contacts were eligible, what counted as a completed resolution, when was it measured, and what happened if the customer returned? Keep forecasts and individual vendor cases out of your baseline until you can compare the same task and measure.
For a pilot, keep the incoming question, the action taken and the final outcome together. Separate requests solved without help from those drafted by AI and approved by staff. Include reopened cases and failed transfers so a smaller queue does not conceal unfinished work.
What does AI resolve in customer service today?
Choose the task and its permissions before choosing a resolution target. Routing, drafting, self-service guidance and follow-up are different forms of assistance; none should be treated as completed customer work merely because a message was generated.
Triage and routing. Let the agent suggest a queue from the enquiry and available customer details. Test ambiguous requests as well as obvious ones. A wrong route can add a transfer rather than save work, so keep a route for unrecognised or sensitive enquiries.
First-response drafting. For an order query or appointment change, prepare a reply against the current record and policy. A person can check it before sending. Measure the time spent reviewing and correcting it as part of the task, rather than counting every draft as a saving.
Self-service guidance. Use approved answers and confirm what the customer is entitled to do. A password-reset explanation is different from permission to change an account. Keep identity checks and exception handling attached to the action.
Follow-up. Remind the responsible person when a case remains open. If the agent may send updates, define the permitted wording and stop conditions. A reminder is useful only if the right person can still take over.
| Inquiry type | Proposed AI role | Human responsibility | Frequency to record |
|---|---|---|---|
| Order status and tracking | Retrieve status; draft or send permitted update | Missing records and delivery disputes | Your observed volume |
| Standard refund request | Collect details; draft and route | Check eligibility and approve money movement | Your observed volume |
| Appointment reschedule | Suggest permitted available slots | Conflicts and exceptional changes | Your observed volume |
| Password reset | Guide the approved identity/reset process | Failed checks and account-recovery exceptions | Your observed volume |
| Billing dispute | Summarise records and route | Decide the disputed amount and remedy | Your observed volume |
| Product defect complaint | Acknowledge and collect details | Investigate safety and remedy | Your observed volume |
| Complex account issue | Prepare history and hand over | Own the decision and customer contact | Your observed volume |
This is a proposed division of work, not a claim that every product can perform each action. Build the volume column from your own tickets. A routine-looking category can still contain an exceptional case that needs a person.
See what an AI agent is for the capability distinction. For the first trial, choose a task with an observable outcome and a clear handover rather than assuming the most frequent enquiry is ready for autonomy.
What is the productivity effect on human agents?
The published 2025 Generative AI at Work study finds a 15% average increase in issues resolved per hour among 5,172 customer-support agents with access to assistance. Effects vary with experience: newer and lower-skilled workers gain more, while experienced workers show smaller speed gains and small quality declines.[1]
The measured workflow keeps people answering customers with AI suggestions. Use that result to frame an assisted-work trial, not a promise that software will independently close the same share of your tickets. An average from the study is a useful comparison, but your enquiry mix, information and review process still matter.
An illustration of the arithmetic. Say your team resolves 500 comparable issues in its current working time. Applying the study’s 15% uplift mechanically would give 575, or 75 additional issues. Those are hypothetical inputs and a calculation, not a forecast of your team’s output. Check quality and repeat contact before treating extra throughput as usable capacity.[1]
IBM reports 17% higher customer satisfaction among mature AI adopters, citing its 2025 Institute for Business Value research. Its public summary does not specify the comparison group or enough measurement detail to turn that into a causal promise for your business.[4]
Named cases add operating context. IBM’s Camping World account reports 33% higher agent efficiency and an overall wait time of 33 seconds, with results dated March 2022. The work included a virtual assistant and agent-desktop integration; the outcome belongs to that broader deployment.[5]
HubSpot’s Nutribees case reports a 77% reduction in tickets handled by people, alongside improved conversions and satisfaction. This is the vendor’s account of a customer deployment, not a measured universal autonomous-resolution rate.[6]
Measure the work people finish with assistance, including the time they spend checking it.
Where are customer service teams in adoption?
Salesforce’s May 2026 report says 66% of surveyed customer-service organisations use AI agents, up from 39% in the prior year. Its double-anonymous survey collected 3,075 responses from service professionals across several world regions. This is reported organisational adoption, not the percentage of issues resolved without staff.[3]
Keep that recent survey separate from older benchmarks. The undated 8x8 report’s 52% “invested” figure combines 29% currently using conversational AI with 23% in the process of adopting it; another 44% planned adoption. It does not mean 52% were already live. The report describes 325 contact-centre and IT respondents and should not be relabelled a current-year census.[9]
HubSpot’s August 2023 survey summary supplies a dated view of use cases. Its questions differ: routing is described as a preferred use case, feedback as a use, and chatbots/self-service as a primary use case. They are not additive slices of the whole market.[8]
| Use case | HubSpot’s reported survey finding | How to use the comparison |
|---|---|---|
| Routing requests | 29% of service experts named it as a preferred use case | Test correct and incorrect routing |
| Customer feedback | 28% used AI to collect and analyse feedback | Compare with the original messages |
| Chatbots and self-service | 26% chose it as their primary use case | Check actual completion and handover |
| Response drafting | Practical candidate; no share assigned here | Include review and correction time |
| Escalation assistance | Practical candidate; no share assigned here | Check who receives and owns the case |
The table helps identify trial questions, rather than a prescribed sequence for your business. A team can use AI to prepare a response without allowing it to send one; an adoption label alone does not tell you which permission was granted.
LivePerson’s 2024 report shows why leadership confidence needs its own label: 96% of surveyed business leaders expected generative AI to improve customer interactions, compared with 55% of consumers. Expectations can influence investment, but your acceptance test still needs completed work and customer outcomes.[7]
What do customers actually want from AI support?
LivePerson’s 2024 report finds a preference for reachable human support: 73% of consumers say they are more likely to buy from a company whose AI is monitored by people they can transfer to. In the same research, 77% of business leaders say AI should be monitored by humans.[7]
The study surveyed 1,505 adults in the US, UK and Australia and 504 US executives at companies with more than 500 employees, in November 2023. Keep those populations and timing attached; they are not a survey of every small-business customer.[7]
Monitoring and transfer do not necessarily mean a person approves every outgoing answer. For your service design, decide which actions need advance permission and which can be checked afterwards. Make the human route visible and pass the conversation history along so the customer does not have to start again.
Salesforce’s June 2022 release reports that 69% of banking customers were willing to use AI to improve their experience. That is a sector-specific willingness measure, not a claim that all consumers prefer AI self-service over people.[10]
| Customer or leader finding | Reported share and scope | Practical design question |
|---|---|---|
| Willing to use AI to improve experience | 69% of banking customers; Salesforce 2022[10] | Does the task and sector match your proposed service? |
| More likely to buy with human-monitored, transferable AI | 73% of consumers; LivePerson 2024[7] | Can a customer reach a person with the case history? |
| Say AI should be monitored | 77% of business leaders; LivePerson 2024[7] | Who owns monitoring and intervention? |
| Expect generative AI to improve interactions | 55% of consumers; LivePerson 2024[7] | What do your customers actually experience? |
Treat preference as a design input, then test the service itself. Ask customers whether they finished the task and could get help when needed. A fast answer can still leave a billing dispute unresolved.
What does the ROI picture look like in customer service AI?
Build the return calculation from your own completed cases and handling effort. Include software, setup, review, corrections and ongoing ownership. Time released by AI becomes a financial benefit only when it changes paid work, usable capacity or an outcome you can measure.
Scope determines the comparison. Record which enquiry types enter the trial and which go directly to staff. Compare the same mix before and after. An agent may appear faster simply because difficult cases have moved into a separate queue; include the receiving team’s work too.
Count review and recovery. A draft can save composition time while still needing a policy check. An autonomous action can create rework if a record is wrong. Keep those costs beside the completed work rather than reporting only the number of automated replies.
Track customer outcomes. Compare repeat contacts, reopened tickets, complaints and customer feedback. Faster handling is valuable when the customer’s task gets done. Treat an unverified claim of lower churn or higher revenue as a question to measure, rather than adding it to the savings column.
The assistance study’s 15% average is a research result, not a minimum, a first-month schedule or a guarantee of cash savings. Its differences across workers are a reason to compare outcomes by task and experience, not just the team average.[1]
For example, a team could start with status enquiries that use a clear record and retain staff approval for disputed refunds. That is a proposed trial design. The useful evidence is what happens to those actual cases, including exceptions and the customer’s next contact.
Use which workflows to automate first to choose the initial scope. Agree the completion test and handover owner before assigning live actions, then decide from the measured result whether to expand.
Frequently asked questions
What percentage of customer service interactions can AI handle? There is no single transferable rate for your support queue. Gartner forecasts 80% of common issues by 2029; Zendesk reports an expectation among CX Trendsetters, not a current market rate. Named deployments use different denominators. Check eligible cases, completed outcomes and repeat contacts.[2][11][12]
Does AI in customer service increase agent productivity? The published 2025 assistance study reports 15% more issues resolved per hour on average, with substantial differences by worker experience. IBM reports 17% higher satisfaction among mature adopters; Camping World’s named deployment reports 33% higher efficiency and 33-second waits. These are different measures and study designs.[1][4][5]
What do customers think about AI in customer service? LivePerson’s 2024 research reports 73% of consumers more likely to buy where AI is human-monitored and transferable, and 77% of leaders favouring human monitoring. Salesforce’s 69% figure concerns banking customers willing to use AI, not all consumers preferring automated self-service.[7][10]
What is the current AI adoption rate in customer service? Salesforce’s May 2026 research reports AI-agent use in 66% of surveyed service organisations, compared with 39% in the prior year. Older 8x8 figures combine current use and ongoing adoption. Keep the date, population and technology definition attached.[3][9]
Notes
- Generative AI at Work. February 4, 2025 online; May 2025 issue Accessed October 15, 2026.
- Gartner Predicts Agentic AI Will Autonomously Resolve 80% of Common Customer Service Issues Without Human Intervention by 2029. March 5, 2025 Accessed October 15, 2026.
- New Research: AI Service Agents Are Scaling and Delivering CSAT. May 20, 2026; survey March 9–April 4, 2026 Accessed October 15, 2026.
- AI in Customer Service. Living page; cites IBV March 23, 2025 Accessed October 15, 2026.
- Camping World: Driving a reimagined customer experience with an AI-powered virtual assistant. Case results dated March 2022 Accessed October 15, 2026.
- Customer Agent customer success stories. Living customer library; publication day not stated Accessed October 15, 2026.
- State of Customer Conversations 2024: Bridging the AI Gap. 2024 report; survey November 2023 Accessed October 15, 2026.
- Pros and Cons of AI in Customer Service [New Data + Expert Insights]. August 14, 2023 update Accessed October 15, 2026.
- State of Conversational AI in the Contact Center. Undated report; primary PDF has no survey/publication year Accessed October 15, 2026.
- Chatbots and Self-Service Deliver Automation-Enabled Efficiency for Prudential and PenFed Credit Union. June 23, 2022 Accessed October 15, 2026.
- Top ITSM trends to watch now. January 15, 2026 Accessed October 15, 2026.
- How LiveAgent resolved 75% of chat interactions via the AI Chatbot. January 20, 2026 Accessed October 15, 2026.
