Plan · Topic 02 of 14
Enterprise AI agent examples: how companies use AI agents in production
Enterprise AI agent examples are the deployments companies and vendors have published, with what the agent does, the figure they report and who reported it. This page sorts the ones with numbers by business function and industry, shows the pattern the measured ones share, and lists the controls each one needed. None is a QueryNow client.
In this guide: 83 pages
Key takeaways
- Of 119 sourced example entries we checked across 13 functions and 10 industries (some deployments appear under both), 64 carry figures from 2025, 2026 or a current undated page, and 89 rest on a vendor customer story. Treat every number as the publisher’s claim.
- The agents with measured results act inside one scoped workflow: an IT fix, an invoice, a freight quote, a requisition check. The broad "digital co-worker" announcements publish scale or intent, and rarely an outcome.
- Four things repeat in the ones that worked: a scoped workflow, defined data access, an approval point for exceptions, and a published before-and-after comparison.
- Projections, "up to" figures, targets and vendor self-use are flagged in every table on this page and its sub-pages, because they travel further than results.
- The sources almost never describe the controls. Identity, allow lists, evaluation gates and monitoring are what you add when you copy an example.
How should you read a published AI agent example?
Ask four questions of every example: who published it, what was measured, whether the number is a result or a projection, and what the agent was allowed to do. Most published examples are vendor customer stories, which the customer approves and the vendor edits, and that is fine as long as you read them that way.
Source types behind the 119 entries we checked (some cite more than one source), and what each can tell you.
| Source type | Entries | What it tells you | What it leaves out |
|---|---|---|---|
| Vendor customer story | 89 | A named company runs the agent; the figure the customer agreed to publish | Method, baseline, whether the figure covers the agent alone |
| Company release or blog | 19 | Scale and intent, sometimes internal figures | Independent checking; vendor self-use describes the vendor’s own product |
| Press | 8 | Context, quotes, earnings-call numbers | Deployment detail; often a pilot or a plan |
| Government release | 5 | Trial results with a method (an F1 score, a randomised trial) | Comparability with a commercial setting |
| Consultant case study | 1 | A result on the consultant’s own platform | The client’s name |
Three more labels matter. "Acts" means the agent changes something in a system; "assists" means it drafts and a person acts; "answers" means question and answer only. A projection ("anticipates", "expects", "goal", "target") is a plan, however precise the number, and an "up to" figure is a ceiling rather than an average. Vendor self-use (Microsoft, ServiceNow, IBM and Salesforce describing their own agents) is a product demonstration with real scale behind it.
What are AI agents doing in each business function?
In every function, the agent with a measured result does one narrow job inside an existing system. The table gives one publicly reported example per function, with the figure as published, its source and date, and the level the agent runs at.
Publicly reported examples by business function, as published by each source and opened on 2 November 2026. Every figure is the publisher’s own claim; none of these companies is a QueryNow client.
| Function | Publicly reported example | Published figure | Source and date | Level |
|---|---|---|---|---|
| IT service desk | Commonwealth Bank of Australia: ChatIT answers IT questions in Teams and runs automated fixes | More than 12,000 automated fixes in six months; 2 minute fix against a 17 minute call | Microsoft Source Asia, 1 July 2025 | Acts |
| HR and onboarding | IBM: AskHR answers HR questions and runs about 80 HR tasks such as vacation requests | 94% containment of common questions; more than 1 million HR transactions in 2024 | IBM case study, undated (vendor self-use) | Acts |
| Finance and accounting | Westbury Street Holdings: AppZen routes invoices from inbox to payment and stops duplicates | 38% less invoice processing time; GBP 774,000 of duplicate invoices stopped | AppZen customer story, undated | Acts |
| Procurement | SUEZ UK: a Pactum agent on Coupa checks every requisition against policy and corrects it | 90% shorter requisition cycle time; 80% of manual reviews removed | Pactum customer story, undated | Acts |
| Legal and contracts | Proof: Ironclad’s Contract Review Agent gives a first-pass NDA review against the playbook | 50% to 75% less time on NDA reviews | Ironclad customer story, undated | Assists |
| Sales | Siemens Smart Infrastructure: two Agentforce agents engage and qualify every inbound lead across 132 countries | About 2,500 leads a month answered; response from days to minutes; 6% qualification rate | Salesforce customer story, 14 September 2026 | Acts |
| Customer service | Canada Goose: Agentforce handles delivery timing and size exchanges over messaging and phone | 89% of routine messaging inquiries and 15% of calls resolved autonomously | Salesforce News, 1 October 2026 | Acts |
| Marketing compliance | Prudential Financial: Norm Ai agents review marketing content line by line against the rules | No measured result; "almost 100%" accuracy is a quote without a method | Norm Ai, 6 January 2026 | Assists |
| Supply chain and logistics | C.H. Robinson: agents quote ocean and air freight, book shipments and process customs documents | More than 80,000 quotes; ocean quotes from two to three days to two to three minutes | C.H. Robinson press release, 26 October 2026 | Acts |
| Field service and maintenance | Textron Aviation: TAMI answers technicians’ questions across about 60,000 pages of manuals | Troubleshooting from up to 20 minutes to 1 to 2 minutes | Microsoft customer story, 20 March 2025 | Answers |
| Software engineering | Thomson Reuters: AWS Transform agents ran .NET modernization from discovery to refactoring | 1.5 million lines of code modernized a month; no baseline given | AWS case study, undated | Acts |
| Security operations | eSentire: a multi-agent system runs threat investigations and contains attacks under senior analyst review | More than 120,000 autonomous investigations in 12 months, averaging 6 minutes | Anthropic customer story, undated | Acts |
| Knowledge search | Wells Fargo: a Teams app gives 35,000 bankers guidance on 1,700 internal procedures | Response times from 10 minutes to 30 seconds | Microsoft Cloud Blog roundup, 24 July 2025 | Answers |
Two more agents act at scale with a published figure: Honeywell’s negotiation agents, more than 2,500 autonomous supplier negotiations (Pactum customer story, undated), and Danfoss’s order agents, up to 80% of transactional decisions with people handling pricing exceptions (Google Cloud customer story, undated).
What do the examples that worked have in common?
Four things, and none of them is the model. The deployments with measured results scoped one workflow, named the systems the agent reads and writes, put a person at the exceptions, and published a before-and-after comparison.
Scoped workflow. CommBank’s agent fixes a slow PC; it does not run the service desk. Westbury Street Holdings’ agent moves an invoice from inbox to payment. C.H. Robinson’s agents quote a shipment or clear a customs file, and the company says more than a dozen agents can work on one shipping task, each with one job. SUEZ’s agent checks a requisition against policy and nothing else. The broad announcements, a digital co-worker for every profession or agents in half of all workflows by 2028, publish intent and scale.
Data access and approval. Every measured example sits inside a system that already held the data: Teams and the ticketing system, Coupa, Dynamics GP, Sales Cloud, an ERP. The agent reads from it under an identity and writes to a short list of objects, and it routes the rest to a person with the reason. Danfoss describes the shape exactly: agents read customer emails in any format, interpret orders and quote requests and trigger the matching actions in the ERP, while people handle exceptions such as pricing. The same handoff appears at CommBank, eSentire and Eneco.
Measurement. The convincing numbers compare a time before with a time after on one task: an average 2 minute fix against a typical 17 minute call at CommBank, 10 minutes to 30 seconds at Wells Fargo, two to three days to two to three minutes at C.H. Robinson, up to 20 minutes to 1 to 2 minutes at Textron. Hours-saved totals and percentage productivity gains are weaker because the method is rarely stated. Siemens publishes the full funnel for its lead agents (engagement rate, qualification rate, satisfaction), which is the standard to copy.
Which examples exist by industry?
Every industry we checked has at least one named deployment and most have figures; healthcare, manufacturing and banking are the deepest, pharma, energy and professional services the thinnest. The table gives one example per industry, and the industry pages linked at the foot of this page show how we apply each pattern.
One publicly reported example per industry, from the 57 industry entries checked on 2 November 2026.
| Industry | Publicly reported example | Published figure and source |
|---|---|---|
| Pharma and life sciences | Novo Nordisk: NovoScribe drafts clinical study documentation for staff review | "10+ weeks to 10 minutes" for clinical study documentation (Anthropic customer story, undated) |
| Manufacturing | Danfoss: agents read order emails and trigger the matching actions in the ERP | Up to 80% of transactional decisions automated; processing time halved (Google Cloud customer story, undated) |
| Financial services | Tata Capital: drafts credit appraisal notes from uploaded financials for credit managers | 25% shorter turnaround; 30% higher credit team productivity (Microsoft customer story, undated) |
| Insurance | ICICI Lombard: a claims copilot reads claim documents and summarises them for adjudicators | Claim processing from 20 minutes to under 10 (Microsoft customer story, undated) |
| Healthcare | Acentra Health: drafts Medicare appeal letters that nurses review and approve | Time per letter from about six minutes to about three (Microsoft customer story, 3 December 2024) |
| Retail | Grupo Falabella: an incident agent creates support tickets around the clock | Resolution from 2 hours to 30 minutes (Google Cloud use-case list, updated 22 April 2026) |
| Logistics | C.H. Robinson: agents quote, book and process customs documents | More than 80,000 quotes; customs time down 23% (company release, 26 October 2026) |
| Energy and utilities | AES: a three-layer agent system runs internal safety audits through to the report | Reports from about two weeks to roughly an hour (Anthropic customer story, undated) |
| Public sector | Citizens Advice with UK i.AI: Caddy finds guidance for advisers | Response time halved in a randomised controlled trial (i.AI, undated) |
| Staffing and professional services | The Adecco Group: Agentforce pre-screening and recruiter agents talk with candidates | More than 2.5 million agent-candidate interactions; no outcome figure (company release, 15 September 2026) |
| Legal services | A&O Shearman: Harvey agents assemble draft work product for loan review and filings | No measured figure; "minutes" against "several hours" is a partner’s quote (company release, 6 April 2025) |
| Accounting and audit | KPMG: agents in KPMG Clara perform substantive audit procedures with auditors in control | No performance figure; more than 95,000 auditors on the platform (KPMG US, 23 April 2025) |
Three areas publish no measured deployment at all in our set: energy grid and field operations, insurance carriers acting on their own, and marketing compliance review, where Prudential, New York Life and Moderna describe agents that assist and publish no cycle-time figure.
What controls did each example need?
The published sources describe the handoff and sometimes the threshold, and almost never the identity, the allow list, the evaluation gate or the monitoring. The table shows, for examples that act, what the source says and what the workflow needs regardless.
Controls the source describes against controls the workflow needs. The last column is our reading of the task, and says nothing about what the company did.
| Example | Risk of the action | Control the source describes | Needed regardless |
|---|---|---|---|
| CommBank ChatIT (IT fixes) | A fix on the wrong device or user | Hands off to a specialist when it cannot resolve the issue | An allow list of fixes; the agent’s own identity; a verified user; a logged result |
| Westbury Street Holdings (accounts payable) | A wrong or duplicate payment | Flags suspicious senders; stops duplicates before payment | A posting threshold outside the model; vendor master read-only; requester and approver separated; monthly sampling |
| SUEZ UK (requisitions) | A purchase approved against policy | Checks each requisition against category, quotes and rate cards | Policy rules a procurement owner edits; escalation above a value; an audit trail per correction |
| Siemens (lead qualification) | A prospect mishandled | Routes qualified leads to a seller with a summary and closes the rest | Consent and disclosure per market; a cap on follow-ups; satisfaction measured |
| Danfoss (order entry) | A wrong order in the ERP | People handle exceptions such as pricing | Write rights limited to order objects; a confidence threshold; a replay of past orders |
| eSentire (threat investigation) | A missed attack or a wrong containment | Senior analysts review critical findings; verdicts compared with experts | Containment gated by severity; every tool call traced; re-testing when models change |
The sub-pages for IT service desk, finance and accounting, customer service and knowledge search go through the five controls (identity, data access, approval, evaluation, monitoring) for each function, with the products that provide them on Microsoft, AWS and Google.
How do Microsoft, AWS and Google implement it?
Which platform the measured examples ran on, and where the gaps are, with statuses as the vendor pages showed on 2 November 2026.
Capability, Microsoft, AWS, Google
| Capability | Microsoft | AWS | |
|---|---|---|---|
| Examples with figures on the platform | CommBank, Eneco, Textron, Bradesco, Wells Fargo and Capita (Copilot Studio, Azure OpenAI, Microsoft 365 Copilot) | Thomson Reuters (AWS Transform); eSentire (Claude on Amazon Bedrock); Novo Nordisk and Banner Health (Claude on AWS) | Danfoss, Grupo Falabella, SIGNAL IDUNA and Highmark Health (Google Cloud, Gemini Enterprise) |
| Current agent platform | Copilot Studio (low code) and Microsoft Foundry Agent Service (pro code); Microsoft Agent 365 as the control plane (GA 1 May 2026) | Amazon Bedrock AgentCore (GA 13 October 2025); Bedrock Agents Classic closed to new customers on 30 July 2026 | Gemini Enterprise Agent Platform (renamed from Vertex AI on 22 April 2026) and the Gemini Enterprise app |
| Identity per agent | Microsoft Entra Agent ID (GA) | AgentCore Identity (GA) | Agent Identity (API in preview since 18 June 2026) |
| Policy on tool calls | Conditional Access for agents; Foundry tool call guardrails (preview); Copilot Studio approvals | AgentCore Policy (GA 3 March 2026) | Agent Gateway (GA 18 June 2026) with Model Armor inline |
| Evaluation and observability | Foundry evaluators and tracing; Copilot Studio agent evaluation | AgentCore Evaluations (GA 31 March 2026); CloudWatch generative AI observability (GA 13 October 2025) | Agent Observability (GA 18 June 2026); Gen AI evaluation service |
Many measured examples ran on SaaS and specialist vendors rather than a cloud agent platform: Salesforce Agentforce (Siemens, Canada Goose, Adecco), Pactum (Honeywell, SUEZ), AppZen (Westbury Street Holdings), Moveworks (Broadcom), Sierra (Chime), Glean (DBS), Harvey (Dentsu, A&O Shearman), Ironclad (Proof) and LangChain (Toyota).
When should you not copy an example, and what goes wrong?
Do not copy an example when its number is a projection, when you cannot measure the same thing, or when the agent would act on a system you have not yet given it a safe identity in. Once you do copy one, these are the ways it goes wrong.
A projection becomes your target
Dow "anticipates" millions in freight savings from a 2024 pilot; Genentech’s research agent "is expected to" automate more than 43,000 hours; Goldman Sachs’s coding pilot is quoted for a productivity rise its CIO believes possible. Each is a plan with a precise number attached.
The definition travels without the number
Resolved, contained and deflected are three different things. Broadcom’s 88% is resolved, Microsoft’s 30% is tickets never created, Eneco’s 67% is chats without handoff. A target copied from one page and measured by another’s definition is wrong before launch.
Hours saved with no method
Capita’s 9,000 hours a month is credited to Copilot overall; DBS’s 10% of work hours is "up to"; Dentsu’s 4 to 10 hours per lawyer per week is described as an estimate. Use them as evidence of adoption, and measure time per task yourself.
Vendor self-use read as a customer result
Microsoft, ServiceNow, IBM and Salesforce publish figures from their own internal agents. The scale is real and the product is theirs. Read them as a demonstration of what a vendor can do with unlimited access to its own platform.
A pilot or an intent read as production
Goldman Sachs’s accounting agents were "in the early stages" in February 2026; FedEx’s agent workforce is a plan for 2028; Siemens’ supplier onboarding agents were announced in the future tense. Check the tense before you cite the company.
Copying the agent without the controls
Few sources explain how the agent was given an identity or which tools it may call, and most say nothing about testing or monitoring; eSentire’s verdict comparison and the Caddy trial are the exceptions. Copying the workflow without those four controls is how a pilot that looked like CommBank’s becomes an incident.
What should you set up before you copy an example?
The eight things to put in place before any of these patterns goes live, whichever cloud or vendor runs the agent.
- The scope in one sentence: which requests, which documents, which queue, and what the agent does not touch.
- The baseline, measured on the same task you will measure after: time per ticket, time per invoice, time to find, resolution as you define it.
- An identity for the agent on its platform (Microsoft Entra Agent ID, AgentCore Identity, Google Agent Identity) and a named owner and sponsor in the registry.
- Data access written down: the systems it reads, the objects it writes, the connectors, and nothing shared through a human account.
- The approval point: the threshold or exception rule, who owns it, where the person sees the context, and what happens on timeout.
- An evaluation set replayed from real cases, with the pass mark signed by the owner before release.
- Monitoring of the outcome measure, the exception rate, errors and cost per run, reviewed weekly for the first quarter.
- The retirement rule: what ends the agent, who switches it off, and where the logs go.
Enterprise example · Publicly reported
CommBank: ChatIT fixes IT issues from Microsoft Teams
Publicly reported third-party example. Commonwealth Bank of Australia, as described by Microsoft Source Asia on 1 July 2025. Not a QueryNow client.
ChatIT is a generative AI IT support agent in Microsoft Teams, built on Copilot Studio and Azure. Employees ask it IT questions; it answers, and for a known set of problems it runs the fix itself, for example clearing caches or resetting a user session for a slow PC. When it cannot resolve an issue it hands off to an IT specialist.
The published figures: more than 2.3 million messages processed; more than 12,000 automated fixes in the six months after the automation function launched; an average fix time of 2 minutes against a typical 17 minute service desk call; nearly 2,500 hours saved; satisfaction of 4.63 out of 5 and an NPS of +79 from internal surveys. The figures are CommBank’s own, published through Microsoft.
The shape is what makes it worth copying. The agent lives in the channel people already use, it started with answers and added fixes as a separate function, the fixes are a defined set, the handoff is explicit, and the measure is a time comparison on one task, an average fix against a typical call. The story does not say how the agent was given an identity, which fixes it may run, how it was tested or what the team watches. The controls below separate what the source says from what the workflow needs.
- Owner
- Not named in the source. The workflow needs an IT support product owner who decides which fixes join the allow list.
- Identity
- Not described. A fix on a user’s device needs the agent’s own identity with scoped rights and the user verified first; on Microsoft that is a Microsoft Entra Agent ID.
- Data access
- The source names Teams as the channel and the service desk as the handoff. The workflow needs read access to the knowledge base and the ticket queue, and endpoint tooling only for the allow-listed fixes.
- Risk tier
- Answers are low risk; automated fixes that change a device are a higher tier and need the allow list and a logged result.
- Approval
- Routine fixes run without a person; the source describes the handoff to a specialist when the agent cannot resolve the issue. Access changes would wait for approval.
- Evaluation
- Not described. A replay of real tickets scoring the path chosen, the fix result and the handoff decision, before each new fix is added.
- Monitoring
- The source reports messages, fixes, fix time, hours saved and satisfaction. Add the reopen rate within seven days and the handoff rate per category.
- Retirement
- Not described. Each automated fix needs an owner who can remove it from the allow list, and the agent an identity that can be disabled.
Publicly reported by Microsoft Source Asia, 1 July 2025
How do you start?
- Pick the function where your own volume is concentrated and an example above already acts: IT service desk, accounts payable, customer service or knowledge search.
- Read the sub-page and the source behind the example, and write down what the source does not say about controls.
- Measure the baseline on the same task the example measured, before anything changes.
- Scope one workflow, give the agent an identity and an owner, and launch answers or drafts before actions.
- Add the first action behind an allow list or a threshold, with the replay signed off, and compare your number with the published one after a quarter.
- Re-read the examples after Microsoft Ignite and AWS re:Invent in late 2026, when platform names and statuses will change.
Frequently asked questions
How are companies using AI agents in production today?
Inside one narrow workflow in a system they already run: fixing IT issues from Teams, moving invoices from inbox to payment, quoting freight, checking requisitions against policy, qualifying inbound leads, answering procedure questions. The examples with measured results all share that shape.
Which enterprise AI agent examples have real numbers?
Of the 119 entries we checked, 64 carry figures from 2025 or 2026 or from a current undated page. The strongest before-and-after times come from CommBank, Wells Fargo, Textron Aviation, C.H. Robinson and Westbury Street Holdings, and Siemens publishes a full funnel for its lead agents. Every figure is still the publisher’s claim.
Are any of these companies QueryNow clients?
No. Every example on this page and its sub-pages is a publicly reported deployment attributed to the company or vendor that published it. Our own proof appears only where it is labelled QueryNow, such as the Rockwell Automation search deployment on the knowledge search page.
Why are most examples vendor customer stories?
Because vendor stories are where companies agree to publish figures: 89 of the 119 rest on one. The customer approves the text, the vendor edits it, and nobody audits it. Company releases, press and government trials make up the rest, and the UK government cases are the only ones with a stated method.
What is the difference between an agent that acts and one that assists?
An agent that acts changes something in a system, such as posting an invoice or resetting a session. One that assists drafts or recommends and a person acts. One that answers only replies. The level sets the controls: an acting agent needs its own identity, an allow list and an approval point.
Which industries have the most published examples?
Healthcare, manufacturing and banking in our set, with five, four and four examples carrying figures. Pharma has one strong example (Novo Nordisk), professional services publishes adoption counts rather than outcomes, and nobody in our set publishes a measured deployment for grid operations or for an insurance carrier acting on its own.
Can we use these numbers in a business case?
As evidence that the pattern works somewhere, yes; as your expected result, no. Measure your own baseline on the same task, and put the published figure beside it with the publisher’s name and date. A business case that quotes Broadcom’s 88% as its own target has borrowed a definition it cannot check.
In this topic
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.
- CommBank’s ChatIT helps employees fix IT issues in seconds using generative AI
- Transforming HR support with agentic AI (IBM AskHR case study)
- WSH cut AP invoice processing time by 38% across a GBP 1.7B operation
- How Pactum’s AI agents support SUEZ in eliminating manual requisition checks
- How Honeywell is running autonomous procurement at enterprise scale
- Proof cuts NDA review times by 75% and gets contract answers in minutes with Ironclad AI
- Siemens takes lead response from days to minutes with Agentforce
- Siemens and Salesforce deepen AI partnership to redefine industrial sales and service
- Getting to ROI: how brands turn cost centers into revenue engines with AI
- Inside Prudential’s marketing transformation: a conversation on AI, compliance and scale
- New York Life partners Norm Ai to scale AI-driven compliance
- Veeva AI agents now available to increase productivity and customer centricity (Moderna quoted)
- How Chime made banking support faster and more personal with AI
- Toyota scales enterprise AI with Deep Agents and LangSmith
- C.H. Robinson debuts Lean AI for ocean shipping, air freight and customs
- Textron Aviation enhances maintenance efficiency with Azure AI
- How Thomson Reuters turbocharged .NET modernization with AWS Transform
- How eSentire runs expert-level threat investigations at scale with Claude
- AI-powered success with more than 1,000 stories of customer transformation and innovation
- Danfoss and Go Autonomous: reimagining order management at scale with Google Cloud
- Novo Nordisk Claude Code case study
- Tata Capital: automating credit appraisal
- ICICI Lombard: increasing productivity using a claims copilot
- Acentra Health boosts employee productivity with generative AI and Azure OpenAI Service
- 1,302 real-world gen AI use cases from the world’s leading organizations
- AES customer story
- Building better frontline services (Caddy)
- Adecco Group announces Salesforce’s Agentforce Coworker rollout across 40 countries
- A&O Shearman and Harvey to roll out agentic AI agents targeting complex legal workflows
- KPMG advances AI integration in KPMG Clara smart audit platform
- Dow reimagines productivity and supply chain efficiency with Microsoft 365 Copilot
- Genentech generative AI case study
- Goldman Sachs taps Anthropic’s Claude to automate accounting, compliance roles
- Goldman Sachs is piloting its first autonomous coder in major AI milestone for Wall Street
- FedEx plans agent workforce in over 50% of workflows by 2028
- How Broadcom powers its knowledge base with AI to help employees help themselves
- Eneco moves to Microsoft Copilot Studio to better support over 1.5 million customers
- Capita customer story: Copilot agents built by employees
- How DBS enables its staff to focus on meaningful work through AI integration
- How Dentsu’s legal and compliance team turned pressure into strategic advantage with Harvey
- Building trust in AI-powered support at Microsoft: our Employee Self-Service agent story
- Banner Health Claude Platform case study
- Microsoft Agent 365 now generally available
- Amazon Bedrock AgentCore is now generally available
- Policy for Amazon Bedrock AgentCore is now generally available
- Amazon Bedrock AgentCore Evaluations is now generally available
- Amazon Bedrock Agents Classic maintenance mode
- Gemini Enterprise Agent Platform release notes
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