The State of Play

A living index of AI adoption across industries — where established practice meets the bleeding edge
UPDATED DAILY
← All domains

🎧 Customer Operations

AI for supporting, retaining, and understanding customers after the sale. The highest concentration of good-practice tiers: chatbots, ticket routing, sentiment analysis, and voice-of-customer are deployed at scale in most industries. Bleeding-edge frontiers include autonomous resolution without human escalation and real-time emotion detection. Momentum is steady but churn prediction and proactive outreach remain stalled.

18 practices: 11 good practice, 4 leading edge, 3 bleeding edge

The Headline

AI in customer service still pays best where a person signs off. New tests show AI phone systems that succeed on one call often fail when the same task is repeated.

18
practices tracked
7
at the frontier
1
moving this fortnight
3,506
evidence items

The Picture

Most companies now use AI behind their staff: drafting replies for employees to approve, summarizing calls and routing tickets are standard features in Zendesk, Salesforce and Amazon Connect. A small group runs customer-facing agents (software that acts on its own without being prompted) at real scale, such as Vodafone's 60 million conversations a month. Far more have tried and retreated: in a Sinch survey of 2,527 decision-makers published in May, 74% had rolled back or shut down a live AI customer agent. Customers have not come round either. In a Gartner survey from early 2026, only 7% had used a company's own chatbot in their latest service interaction, a share statistically unchanged since 2022. If your AI sits behind a human reviewer, you are in the pack. If you run autonomous agents that have stuck, you are ahead of most.

This Fortnight

  • A five-run test of six AI voice platforms found every one far less reliable than a single call suggests. In tests by G2 and the testing firm Bluejay, success across five runs of the same 90 tasks came in 15.6 to 28.2 points below single-call success, and most failures traced to the AI's own decisions rather than speech recognition. A voice pilot judged on one pass through a demo script overstates what your customers will get.

  • Twilio's chief executive put voice AI at about 1% of available voice minutes, even as a health insurer signed Twilio's first eight-figure voice-agent deal. He cited no source, and the insurer has published no figures on how many calls the system handles or how satisfied callers are. Big phone-automation contracts are being signed well before the technology has proved itself at volume, so a peer moving early tells you little about whether it works.

  • Singlife now lets its AI email agent answer routine inquiries without staff review, after accuracy rose from about 30% to roughly 80%. Complex or sensitive emails still get drafts that officers check, and Singlife's Romil Sharma says "You can't just leave AI agents running by themselves." Zendesk also made confidence-gated drafting generally available (out of beta), showing suggestions only above a quality threshold. Unreviewed sending is earned one topic at a time, after months of reviewing failures.

  • Dialpad announced a combination with chatbot maker Ada on 4 November, which BetaKit reports as an all-share deal valuing Ada at about C$600 million. It follows Salesforce's agreement in June to buy Fin for about $3.6 billion and Automation Anywhere's agreement to buy Boost.ai in October. Independent customer-service AI vendors are being absorbed into larger platforms, so check what your contract says if your provider changes owner.

  • A re-analysis of a randomized trial covering 64,000 customers found that targeting the highest-risk customers barely beat picking at random. Targeting the customers most likely to respond to contact did far better, though the data covers marketing email rather than retention offers. Before funding a churn model (software that predicts who will leave), budget for a no-contact control group that proves which interventions actually change behavior.

Coming Up

  • Zendesk will remove or rework three AI features between now and 19 January 2027. Its legacy scripted bot is removed on 10 December, migrations moving AI-detected topic, language and sentiment into standard ticket fields run until 4 December, and suggested first replies retire on 19 January. Have someone audit every workflow, trigger and report that relies on the old tags or bot flows now, before routing quietly breaks.

  • More independent customer-service AI vendors may change hands, with Decagon reported in October to have hired a bank to test acquisition interest. Salesforce, Automation Anywhere and Dialpad have each agreed to buy or combine with a specialist since June. Before signing or renewing, negotiate data-export rights and pricing protections that survive a change of owner.

  • Zendesk plans to release built-in two-way voice translation for support calls in the first quarter of 2027. Analyst firm Futurum cautions that quality on complex, sensitive or urgent calls is unproven, and California canceled a statewide AI translation rollout in its courts after a pilot found serious mistranslations. If language coverage limits your phone support, pilot it on routine queues and keep human interpreters for regulated or high-stakes calls.

What's Hard About This

  • Vendors' resolution rates do not measure what your customers experience. Helpdesk vendors count a conversation as resolved after time windows of their own choosing, none checks for repeat contact, and Twilio warns that abandoned calls get counted as handled; an earlier Gartner survey found only 14% of issues fully resolved through self-service. Define resolution as no repeat contact within a set period and hold every vendor to your definition.

  • A human reviewer limits the risk of AI replies but does not remove it. In one e-commerce test, staff sent only 12% of AI drafts unchanged, and in a separate test the review queue caught every off-topic reply that full automation would have sent, yet a busy reviewer could still have approved them. Courts and tribunals, as in Air Canada's 2024 chatbot case, hold the company answerable for what its AI tells customers, so budget real reviewer time and spot-check approvals.

  • The quality of your help content, not the AI model, sets the ceiling on results. ServiceNow's own 5 November patch notes told customers to roll back a broken AI article-generation feature, and no major helpdesk updates existing articles without a person. Brazilian payments firm Asaas reports drafting reliability rising from 49% to 92% after restructuring its knowledge base, so name an owner for help content and fund its upkeep before expanding automation.

2012View this domain's timeline →Today

Practices in this Domain (18)

PRACTICETIERTREND
Agent assist — auto-draft with human reviewGOOD PRACTICE— Steady
Agent assist — autonomous sendBLEEDING EDGE— Steady
Agent assist — response suggestionGOOD PRACTICE— Steady
Agent quality monitoring & coachingGOOD PRACTICE— Steady
Call summarisation & dispositionGOOD PRACTICE— Steady
Customer health scoring & churn predictionGOOD PRACTICE— Steady
Customer support chatbots — autonomous resolutionGOOD PRACTICE— Steady
Customer support chatbots — LLM-powered conversationalBLEEDING EDGE— Steady
Customer support chatbots — scriptedGOOD PRACTICE↓ Declining
Knowledge base generation & maintenanceGOOD PRACTICE— Steady
Proactive customer engagement & onboardingGOOD PRACTICE— Steady
Returns, warranty & claims processing automationLEADING EDGE— Steady
Self-service content & community managementBLEEDING EDGE— Steady
Ticket intelligence — intent, sentiment & language detectionLEADING EDGE— Steady
Ticket routing, triage & prioritisationGOOD PRACTICE— Steady
Voice AI — IVR replacement & phone supportLEADING EDGE— Steady
Voice AI — real-time translation in support callsLEADING EDGE— Steady
Voice of Customer & journey analysisGOOD PRACTICE— Steady
Read the full technical briefing (1,798 words) →

Where AI Stands in Customer Operations

Customer operations is where AI has gone furthest in everyday business use, and the pattern of that progress is now clear. AI has advanced fastest where a person still makes the final call. Drafting replies for agents to approve, suggesting answers mid-call, writing call summaries, scoring interactions and routing tickets all ship as standard features in Zendesk, Salesforce, Amazon Connect, Microsoft Dynamics 365, Genesys and Freshdesk. The productivity case rests on field research as well as vendor case studies. An NBER study of 5,179 support agents found that AI-suggested responses raised issues resolved per hour by 13.8%, and by 35% for junior agents. Octopus Energy now sends roughly a third of its customer emails with AI assistance. TELUS Digital reports 15% more issues resolved per hour across more than 5,000 agents. Truist and Barclays US have each summarised millions of calls. The gains are real, but they are uneven: they fall mostly to newer staff and depend on the knowledge base behind the tool.

The customer-facing side looks different. A few organisations run autonomous agents at genuine scale. Vodafone's SuperTOBi handles 60 million conversations a month. Telstra's Agentforce rollout applies refunds. ElevenLabs reports that Klarna made a voice agent the first line of phone support for all 35 million of its US customers. Salesforce paid $3.6bn for Fin. Yet reversal remains the most common outcome after launch. Sinch's survey of 2,527 senior decision-makers found that 74% had rolled back or shut down a live AI customer-communications agent. Among organisations with mature governance the figure was 81%, because better monitoring exposes failures rather than preventing them. Customers have not come round either. Gartner's survey of 3,566 customers in February and March 2026 found that only 7% had used a company-provided chatbot in their most recent service interaction, a share statistically unchanged since 2022, while their use of third-party GenAI tools nearly doubled. Cresta finds only 9% of conversations handled entirely by AI. ContactBabel found that just 2% of 202 contact centres run agents that plan and carry out actions across systems on their own.

What sets this domain apart is that the model is rarely the constraint. Where deployments stall, the causes are stale knowledge, systems the agent cannot reach, handoffs that drop context, and metrics that make results look better than they are. The predictive and proactive end of the domain is stuck for a related reason: knowing is not the same as acting. Churn models can flag at-risk accounts, but a Customer.io survey of more than 250 lifecycle marketers found that only 26% of teams act on churn signals. Randomised evidence shows that acting on risk scores alone can backfire: at a wireless carrier, usage-based plan recommendations sent to customers raised three-month churn to 10%, against 6% in the untouched control group. In a Gartner survey of more than 6,000 customers, only 13% had experienced proactive service from brands they had recently dealt with. Across the domain, the hard part has shifted from what the technology can do to the operational discipline around it: owning the knowledge base, defining what counts as resolved, and deciding who acts on what the system flags.

What's New, 2026-11-01 to 2026-11-15

The most telling news this fortnight came from reliability testing, not product launches. G2 and the testing firm Bluejay ran six voice-agent platforms (LiveKit, Pipecat, Retell, ElevenLabs Agents, Vapi and Bland) through 90 tasks five times each. Five-run consistency came in 15.6 to 28.2 points below single-call success on every platform, and Vapi fell to 35.6%. Twilio's chief executive put voice AI at about 1% of total available voice minutes, without citing a source, even as a health insurer signed Twilio's first eight-figure deal for a voice agent. ServiceNow's Brazil Patch 2 notes, released on 5 November, record that multi-knowledge-base article generation was broken and that the Mosaic migration should be rolled back. A competing vendor's audit of 68 US listed companies whose chat opens with a bot found that only four fully answered three basic visitor questions, and that button-menu and keyword bots fully answered none. In churn, an independent re-analysis of Kevin Hillstrom's randomised dataset of 64,000 customers found that targeting the highest-risk decile barely beat random selection (49 against 45 extra visits per 1,000 emailed), while uplift targeting reached 73. The caveat is that the data concerns marketing email, not retention offers. Gartner's survey of 275 service leaders, dating from August, also circulated again. Only 4% had achieved the AI return they expected, and 7% had already rehired after cutting staff.

Meanwhile the human-gated pattern kept spreading. Zendesk made confidence-gated auto assist and in-language draft translation generally available. Tokyo Metro now drafts replies to about 100,000 web-form inquiries a year, and each draft is edited by an operator and checked by a manager before it goes out. Singlife's Agentforce email agent answers more than 20% of incoming email without review, after accuracy rose from about 30% to roughly 80%, yet its own Romil Sharma says "You can't just leave AI agents running by themselves." Talkdesk reports that Michaels cut after-call work by 93%. Shinhan Bank's AI interpretation, now at 61 branches, cut counter task times by 18.8% to 33.3% against its earlier three-party interpreter calls. Consolidation continued: Dialpad announced a combination with Ada on 4 November, which BetaKit reports as an all-share deal valuing Ada at about C$600m. One proactive-outreach result shows how thin the returns can be: an AI voice campaign reached 10,672 of 16,801 abandoned FatakPay sign-ups, and 111 paid, or 0.66% of the leads. None of this changed where any practice stands. It sharpened the existing picture.

Key Tensions

  • Containment is not resolution. Twilio warns that abandoned voice conversations get counted as contained, and the USPS inspector general found that hang-ups before transfer inflated the Postal Service's containment figures in the same way. Helpdesk vendors count a conversation as resolved using windows that range from two hours to 72 hours, and none checks for repeat contact afterwards. Gartner's survey of 5,728 customers found that only 14% of issues were fully resolved through self-service. Until resolution has a shared definition, vendor rates of 70–90% cannot be reconciled with what operators actually see.

  • The review gate limits risk but does not remove it. Draft-and-approve has become the default because autonomy keeps reversing, and rulings such as Air Canada's chatbot case and the Higher Regional Court of Hamm's make companies answerable for what their AI tells customers. But in one eesel test, agents sent only 12% of 284 drafts unchanged. In a 12-prompt off-topic test, shadow mode held every ungrounded reply that autopilot would have sent, and a busy reviewer could still have approved them. An MIT meta-analysis of 106 experiments found that human-AI pairs underperformed on decision tasks, which suggests a draft can anchor the reviewer as much as it helps.

  • Knowledge quality, not model capability, sets the ceiling. Pylon traced more than half of roughly 200 AI-agent escalations in one September week to knowledge gaps. G2's analysis of reviews finds inaccurate or outdated documentation is the dependency complaints mention most. When teams fix their content, results move: Asaas reports Zendesk Copilot reliability rising from 49% to 92% after it restructured its knowledge base, and Thames Water rewrote about 2,000 articles before its assistant went live. Yet a review of six helpdesks found none that updates existing articles without a person, and ServiceNow's own generation feature needed a rollback this month.

  • Detection outruns action. AI quality tools now score every interaction, yet ScorebuddyCX found that only 18% of coaching sessions are triggered by AI analysis. SQM Group cites a study of 5,172 agents in which AI assistance raised productivity by about 15% with no economically significant gain in NPS. For churn, Vidio's renewal model reliably flagged users who would leave, but nothing the team tried moved the number. Kustomer found that 88.5% of CX leaders are confident they can spot churn signals, but only 56% act on them in time.

  • Capital is consolidating faster than returns are proven. Salesforce closed its $3.6bn acquisition of Fin in September, Automation Anywhere agreed to buy Boost.ai in October, and Dialpad announced its combination with Ada in November. Sacra estimates Sierra's annual recurring revenue at $400m. Against that, Gartner found that only 24% of 275 service leaders saw any positive AI return, and Gartner's separate analysis of 432 customer-service AI use cases found a quarter delivering positive ROI, a quarter negative and 42% unclear. Revenue is growing. Evidence that buyers are getting value is not keeping pace.

Top 10 Evidence Items

  1. We ran the same 90 Voice AI tasks five times. Here’s what we learned about reliability. (industry-report) — This is the fortnight's headline reliability finding, showing single-call demos overstate production consistency across every major voice platform. https://ai.g2.com/build-notes/we-ran-the-same-90-voice-ai-tasks-five-times-here-s-what-we-learned-about-reliability
  2. How Singlife keeps its AI agents accurate and costs in check (news-coverage) — Singlife is the clearest case of autonomy earned through iteration rather than granted outright, and even its own lead stresses it still can't run unsupervised. https://www.theedgesingapore.com/digitaledge/digital-economy/how-singlife-keeps-its-ai-agents-accurate-and-costs-check
  3. We Ran the Off-Topic Test (opinion) — Shows the review gate's limit directly: shadow mode caught every bad reply, but a busy human reviewer is the only thing standing between that and a live send. https://shyanming.substack.com/p/we-ran-the-off-topic-test
  4. Brazil Patch 2 • Brazil Release Notes • Docs | ServiceNow (product-ga) — A vendor's own patch notes admitting a generation feature broke and a migration needs reverting undercuts the idea that knowledge maintenance can be automated away. https://www.servicenow.com/docs/r/release-notes/brazil-patch-2.html
  5. Why Churn Prediction Models Fail at Customer Retention (opinion) — Reframes the churn-prediction story: ranking risk barely beats random, and the real gain comes from uplift modelling, not better scores. https://hackernoon.com/why-churn-prediction-models-fail-at-customer-retention
  6. Contact Center AI Agent Buyer's Handbook: Choose the Right Platform (opinion) — The 2% figure is the hard ceiling for 'agents that act across systems', against which every vendor autonomy claim should be read. https://cxtoday.com/ai-agents-in-contact-centers-the-definitive-buyers-guide
  7. Just 4% of Customer Service Leaders Hit Their Expected Automation Returns (news-coverage) — Puts a number on the gap between consolidation-driven valuations and leaders who actually see the AI return they expected. https://www.cmswire.com/customer-experience/just-4-of-customer-service-leaders-hit-their-expected-automation-returns/
  8. Twilio Sees CX Demand Rise for AI Voice and Agent Identity (news-coverage) — Twilio's own CEO estimating voice AI at 1% of minutes, right as the company signs an eight-figure voice deal, captures the hype-versus-scale tension exactly. https://cxtoday.com/twilio-sees-cx-demand-rise-for-ai-voice-and-agent-identity
  9. Dialpad–Ada deal targets customer-service continuity, not just automated answers (news-coverage) — Marks the acquisition wave extending from support bots into voice/CX continuity, with no resolution-quality evidence attached to the price. https://aiagentchronicle.com/articles/2026-11-06-dialpad-and-ada-combine-voice-ai-with-customer-service-agents-putting-retention/
  10. Beyond the Business Case: When AI Investment Actually Pays Off (conference-talk) — A first-hand account that a model correctly flagging churners changes nothing without an action playbook, the cleanest statement of the detection-action gap. https://thecioroom.com/blog/when-ai-investment-pays-off-cio-panel