- Home
- AI agents
- Enterprise AI agent examples
- Customer service
Plan · Enterprise AI agent examples
AI agents in customer service: what companies report and what sits behind the numbers
A customer service agent answers routine questions, checks order and account status, completes small transactions within policy such as a size exchange or a password reset, and hands the rest to a person with a summary. For messaging, published figures run from two-thirds to 89% of contacts handled without a person, each defined differently; voice runs far lower.
In this guide: 83 pages
Key takeaways
- Four companies publish figures from live deployments: Canada Goose (89% of routine messaging resolved), Chime (resolution up from 50% to over 70%), Eneco (67% of chats without handoff, up from 40%) and Klarna (two-thirds of chats in its first month, in 2024). All are self-reported or vendor-reported.
- The measured gains come from scope: routine questions, status checks and a short list of transactions, with the handoff written as carefully as the answer.
- Since 2 August 2026 the EU AI Act’s transparency duty applies, and the Commission’s own examples include an AI agent: customers must be told they are dealing with an AI.
- Resolution, containment and deflection are different measures. Eneco counts conversations without handoff, Chime counts resolution, and Microsoft restates Eneco’s 67% against 40% as "70% more". Read the definition before the number.
What do AI agents do in customer service?
They handle the contacts that have a known answer or a known transaction: delivery timing, order and account status, password resets, size exchanges, payment questions, cancellations and disputes within policy. Anything unsure, sensitive or outside policy goes to a person with a summary of what was said.
Customer service tasks by level. "Answers" means it replies from content; "assists" means it routes or drafts and a person acts; "acts" means it completes the transaction.
| Task | Level | What the agent touches | Seen in |
|---|---|---|---|
| Delivery timing, order status | Answers | Order management and carrier systems, read only | Canada Goose |
| Direct deposit status, account access | Answers, with the customer verified | The banking core, read only | Chime |
| Password reset | Acts | The customer identity system | Chime; Wiley (pilot) |
| Size exchange, returns, cancellations | Acts within policy | Order management, with limits set in policy | Canada Goose; Klarna |
| Payment questions and disputes | Acts within policy | Payment and dispute records | Klarna |
| Multilingual website chat | Answers; first transactions added | Knowledge content; account systems for the first workflows | Eneco |
| Handoff with an AI summary | Assists | Live chat or the contact centre, with the conversation attached | Eneco; Chime; Canada Goose |
| Phone: identify the need and route | Assists | Telephony and the contact centre queue | The Home Depot (pilot); Canada Goose |
Voice is the newest channel and the least measured. Canada Goose reports 15% of calls resolved autonomously against 89% of messaging, and The Home Depot’s phone agent is quoted only for identifying the caller’s need within 10 seconds in a pilot (Google Cloud use-case list, updated 22 April 2026).
What have companies published about customer service agents?
Four organizations publish figures from production and one from a pilot. Every entry is a vendor customer story or a company release, and none has been independently audited.
Publicly reported customer service agents, as published by each source and opened on 2 November 2026. None of these companies is a QueryNow client.
| Company | What the agent does | Published figure | Source and date | How to read it |
|---|---|---|---|---|
| Canada Goose | Agentforce handles routine questions such as delivery timing and size exchanges over messaging and phone, so staff focus on personal shopping | 89% of routine messaging inquiries and 15% of calls resolved autonomously | Salesforce News, 1 October 2026 | Vendor claim quoting the head of customer experience; the specific system actions are not listed |
| Chime | Sierra agents answer routine member questions around the clock (direct deposit status, account access, password resets) and hand off when needed | Resolution rate up from 50% to over 70% | Sierra customer story, 23 October 2025 | Vendor claim quoting Chime’s COO; "resolution" is not defined on the page |
| Eneco (Belgium) | A multilingual website agent replaced the old chatbot; it hands unsure cases to live chat with an AI summary, and its first transactional workflows are live | 24,000 chats a month against 10,000 before; 67% of conversations handled without handoff against 40% before; intent recognition above 95%; built in about three months | Microsoft customer story, 5 March 2025 | Vendor customer story on Copilot Studio; Microsoft’s summary restates the 67% against 40% as "70% more" |
| Klarna | An AI assistant handles service chats in many languages and manages returns, payment issues, cancellations and disputes | 2.3 million conversations in the first month, two-thirds of service chats; resolution time under 2 minutes against 11; 25% fewer repeat inquiries | Klarna press release, 27 February 2024 | A 2024 company release; the profit figure in the same release is an estimate, so it is not quoted here |
| Wiley (pilot) | Agentforce answers from the knowledge base, resolves account access issues and resets passwords, and triages registration and payment issues for students | Outperformed the previous chatbot "by over 40% case resolution in the first few weeks" | Salesforce customer story, undated, read 2 November 2026 | An early pilot; the 213% ROI on the same page belongs to the wider Service Cloud rollout, not to the agent |
Two banking deployments are worth knowing without quoting their numbers as current: ABN AMRO’s "Anna" on Copilot Studio, published in December 2024, and Commerzbank’s "Bene" on Google Cloud, which Google reported at over two million chats and 70% of inquiries resolved in October 2025.
What controls does a customer service agent need?
Verification of the customer, limits for the transactions, a disclosure, an evaluation set built from real conversations, and monitoring that measures resolution rather than deflection alone.
- Identity and verification. The agent acts for the verified customer only, and the verification step is yours rather than the model’s. Account reads happen only after it, and anything that changes an account carries the customer’s verified identity in the log. The agent itself runs under its own identity on each platform (Microsoft Entra Agent ID, AgentCore Identity, Google Agent Identity).
- Data access. Read order, delivery and account status through narrow connectors; write only the transactions on the list. Keep card numbers, identity documents and anything the agent does not need out of its reach, and redact them before telemetry is recorded. Bedrock Guardrails sensitive information filters mask personal data in prompts and responses, Model Armor has Sensitive Data Protection, and Foundry guardrails add a PII control in preview, and none of them covers every log field, so check the trace itself.
- Approval and limits. Exchanges, cancellations and payment adjustments run within amounts and conditions that a deterministic policy enforces outside the model: AgentCore Policy on AWS, Copilot Studio agent flows with approvals on Microsoft, Agent Gateway on Google. Above the limit, or on a dispute, the agent hands off.
- Disclosure. Since 2 August 2026 the EU AI Act’s Article 50 requires that people are told they are dealing with an AI system, and the Commission’s examples name a chatbot and an AI agent. Say it at the start of every contact, in every market, and make the route to a person visible.
- Evaluation. Build the test set from last quarter’s transcripts, with the hard cases included: angry customers, two requests in one message, a request outside policy. Score resolution, policy adherence and the handoff decision, and re-run the set before every model or prompt change.
- Monitoring. Track resolution as you define it, repeat contacts within a window you choose (Klarna reports 25% fewer repeat inquiries, without stating the window), escalations per intent, and satisfaction. Review a sample of handled conversations weekly. DKB does this for its banking assistant, with flagged sessions going to expert review queues ranked by severity (Langfuse customer story, 31 August 2026).
Prompt injection arrives through the customer’s message and through anything the agent reads, such as an order note. Input checks reduce the risk; the limits and the handoff cap the damage.
Where should you start?
Start with the five intents that make up most of your contacts, as answers with a visible handoff, and add transactions one at a time once the resolution number has held for a month.
- Pull a quarter of transcripts and rank intents by volume. Delivery status, order status and account access usually lead.
- Record the baseline per intent: resolution rate, handling time, repeat contacts, satisfaction. Eneco’s 40% without handoff is the kind of number you need before launch.
- Launch answers only, with the AI disclosure and the handoff with summary, on one channel. Measure for a month.
- Add the first transactions (password reset, exchange within policy) behind verification and a policy limit, and replay last month’s transcripts before release.
- Add voice last, with a lower target. Canada Goose’s 15% of calls against 89% of messaging shows the gap.
How do Microsoft, AWS and Google implement it?
Where the published customer service agents ran and what each cloud offers, with statuses as the vendor pages showed on 2 November 2026.
Capability, Microsoft, AWS, Google
| Capability | Microsoft | AWS | |
|---|---|---|---|
| Where the examples ran | Eneco and ABN AMRO on Copilot Studio with Azure AI language services | Druva, quoted on the AgentCore product page (undated) as solving 68% of support issues without a person; no other customer service example with figures in the research set names AWS | Commerzbank on Google Cloud; The Home Depot and Humana on Gemini Enterprise for Customer Experience |
| Build the agent | Copilot Studio, including its Customer Conversation harness; Microsoft Foundry Agent Service for custom builds | Amazon Bedrock AgentCore (Runtime, Gateway, Identity GA 13 October 2025); Bedrock Agents Classic closed to new customers on 30 July 2026 | Gemini Enterprise for Customer Experience; Gemini Enterprise Agent Platform with ADK and Agent Runtime |
| Limits on transactions | Copilot Studio agent flows with approvals (multistage approvals in preview); Foundry tool call guardrails (preview) | AgentCore Policy, Cedar rules per tool call (GA 3 March 2026) | Agent Gateway policies (GA 18 June 2026) with Model Armor inline |
| Input and output checks | Azure AI Content Safety Prompt Shields (GA); Foundry guardrails PII control (preview); Copilot Studio built-in protection | Bedrock Guardrails: content filters, denied topics, sensitive information filters; the prompt attack filter does not evaluate tool results | Model Armor: prompt injection, sensitive data protection, malicious URL detection |
| Handoff to a person | Copilot Studio Transfer conversation node with context variables, marked Escalated in analytics | A Gateway tool into the contact centre; Step Functions callbacks for asynchronous approval | Not documented in the research set; build it as a tool call into the contact centre |
| Evaluation and monitoring | Copilot Studio agent evaluation; Foundry tracing and evaluators | AgentCore Evaluations (GA 31 March 2026); CloudWatch generative AI observability | Agent Observability (GA 18 June 2026); Gen AI evaluation service |
Canada Goose and Wiley run on Salesforce Agentforce, Chime on Sierra and Klarna on OpenAI. SaaS and vendor platforms carried four of the five measured examples, and the controls above apply the same way on each.
Frequently asked questions
What resolution rate should a customer service agent reach?
For messaging, the published range is two-thirds to 89% of in-scope contacts, each publisher defining resolution differently, and voice results sit far lower. Set your target per intent from your own baseline, and count a contact as resolved only when the customer does not come back within seven days.
Do customers have to be told they are talking to an AI?
In the EU, yes. The AI Act’s Article 50 transparency duty has applied since 2 August 2026, and the Commission’s examples include an AI agent. Outside the EU it is still the right design, because the handoff to a person only works if customers know they can ask for it.
Can the agent handle payments and disputes?
Klarna’s 2024 release says its assistant manages payment issues and disputes, and Canada Goose’s agent handles size exchanges. The condition is a limit the policy engine enforces and a handoff above it. Disputes with legal weight belong with a person, and the agent’s job is to collect the facts first.
Why are voice results so much lower than chat?
Canada Goose reports 15% of calls resolved against 89% of messaging, and The Home Depot’s phone pilot is quoted only for identifying the need within 10 seconds. Speech adds recognition errors, verification is harder, and callers often bring the cases chat could not solve.
What is the difference between Eneco’s 67% and "70% more"?
Eneco reports 67% of conversations handled without handoff against 40% before the agent: 27 percentage points, or a 67.5% relative increase, which Microsoft’s executive summary rounds to "70% more". Quote the share handled, which you can measure against your own baseline.
More in Enterprise AI agent examples
Sources
Each source was read on 2 November 2026. The date shown is the date on the page, or the page’s ms.date for Microsoft Learn; an undated page shows the date it was read.
- Getting to ROI: how brands turn cost centers into revenue engines with AI
- How Chime made banking support faster and more personal with AI
- Eneco moves to Microsoft Copilot Studio to better support over 1.5 million customers
- Klarna AI assistant handles two-thirds of customer service chats in its first month
- Wiley sees 213% return on investment with Salesforce
- ABN AMRO Bank migrates to Microsoft Copilot Studio to support over 3.5 million customer conversations
- Introducing Gemini Enterprise
- The Home Depot and Google Cloud launch agentic AI tools to help customers and associates
- 1,302 real-world gen AI use cases from the world’s leading organizations
- Amazon Bedrock AgentCore
- Safer and more transparent AI
- DKB (Deutsche Kreditbank) customer story
- Hand off to a live agent (Copilot Studio)
- Prompt Shields
- Prompt attack filters in Amazon Bedrock Guardrails
- Policy for Amazon Bedrock AgentCore is now generally available
- Model Armor overview
- Gemini Enterprise Agent Platform release notes
- Copilot Studio harnesses overview
Planning an AI program?
Talk to the team that builds it and runs it after launch.
- Built to SOC 2, HIPAA and GDPR standards
- EU AI Act-aligned delivery
- Microsoft Solutions Partner