15 AI Agent Use Cases for B2B Companies, With Real Results

Most lists of agent use cases describe what agents could do. This one collects what companies have actually deployed and the results they reported, including several systems .wrk built for clients. Some results are striking, and at least one company partly reversed course after going too far.

The B2B AI agent use cases with the clearest published results share three traits: a narrow, high-volume task, access to the right tools and data, and a defined route to a person when the agent is unsure.

This article covers 15 use cases across customer service, internal service desks, software development, content and research, and document-heavy work, and labels each one honestly as an agent, an assistant or a workflow with AI steps. It does not rank agent platforms, and it does not explain how to build an agent step by step.

Key takeaways

• The strongest published agent results come from customer service and software engineering, where tasks repeat at high volume and outcomes are easy to measure.
• Klarna's assistant handled two-thirds of customer chats in its first month, yet the company later reinvested in human support because quality mattered as much as cost.
• In code migrations, automated checks such as tests, linters and compilers are what make agent autonomy safe enough to use at scale.
• Many successful "agents" are partly workflows: a fixed sequence of steps with AI handling the parts that need judgment.
• Every case in this list keeps a human route open, through escalation, review queues or approval gates.
• Starting narrow and in shadow mode is the common pattern behind the projects that reached production.

How to read this list

Three labels describe how much freedom the AI has in each case.

• Agent: the model decides which steps to take and which tools to call to complete a task.
• Assistant: the model answers questions from company knowledge and hands off to a person when needed.
• Workflow with AI steps: the sequence is fixed in code, and the model handles classification, extraction or drafting inside it.

The distinction matters for cost, testing and risk. Our comparison of AI workflows and AI agents explains the trade-offs in detail.

Customer service and contact center

1. Consumer support resolution

What the agent does. It handles customer chats about refunds, returns, payments and disputes, with access to account data and the authority to complete common requests.

Real case. In February 2024, Klarna reported that its AI assistant handled two-thirds of customer service chats in its first month, about 2.3 million conversations in more than 35 languages, which it equated to the work of 700 full-time agents. Average resolution time dropped from 11 minutes to under 2. In May 2025, CEO Sebastian Siemiatkowski told Bloomberg that the company was hiring human support staff again, because a cost-focused approach had led to lower quality, and that customers should always be able to reach a person.

What to copy. Track quality and customer satisfaction alongside cost from the first day, and keep a clear route to a human.

2. B2B help center agent

What the agent does. It answers product and account questions on a support site, guides users through fixes and opens a case with a human engineer when needed.

Real case. Salesforce Help receives more than 60 million visits and about two million support requests a year. The company started its own agent with 200 authenticated users over four weeks. According to Salesforce, the agent has since passed five million conversations and resolves more than 63% of them instantly, around the clock in seven languages.

What to copy. Pilot with a small group of signed-in users, where you know who is asking and can measure resolution properly.

3. Support with fast human handoff

What the agent does. It resolves the most common support issues directly and routes complex ones to the right human team with the context already collected.

Real case. In February 2025, Lyft announced a customer care assistant built with Anthropic's Claude. The company reported that average resolution time fell by 87% and that the assistant resolves thousands of requests a day, passing more involved cases to human agents.

What to copy. Design the handoff as carefully as the automation. A good handoff saves the customer from repeating the problem.

4. Voice self-service

What the agent does. It answers inbound calls, understands the caller's intent, completes routine requests through back-office systems and transfers to a live agent when needed.

Real case. A regional telecom operator with more than 12 million subscribers handled over 12,000 inbound calls a day at peak, and about 35% of routine enquiries still needed a live agent. Regulation required voice data to stay inside the operator's network, so cloud speech services were ruled out. .wrk deployed an on-premise platform with speech recognition and a voice agent covering the 15 most common intents, including balance checks, top-ups, tariff information and SIM blocking. The agent had to respond within two seconds, ran on a high-availability cluster connected to the operator's phone system and moved from shadow mode to limited traffic and then full production on two pilot sites. Read the voice AI case study.

What to copy. Run the agent in shadow mode on real calls before it speaks to customers.

5. Call quality scoring

What the AI does. It transcribes every call, scores it against a quality checklist and sends summaries, scores and escalation flags to the CRM and to supervisors.

Real case. In the same telecom project, quality teams had manually reviewed fewer than 2% of calls. The new pipeline transcribes and scores every recorded call, combines language model and rule-based checks and posts results to the existing ticketing system. Supervisors review the scoring weekly to tune it.

What to copy. When the steps never change, a workflow with AI steps is cheaper and easier to test than an agent.

Internal service desks

6. HR service desk

What the agent does. It answers employee questions about policies, leave and benefits and completes routine HR transactions, such as updating records or issuing letters.

Real case. IBM reports that its internal HR agent, AskHR, handled more than 11.5 million employee interactions in 2024, resolves 94% of common questions without a human and automates around 80 HR tasks. IBM attributes a 40% reduction in HR operating costs over four years to the program, with human advisors handling complex cases.

What to copy. Start with the most common questions, then add transactions one by one as accuracy is proven.

7. Engineering support in Slack

What the assistant does. It answers engineers' questions in support channels using internal documentation and links to the source of each answer.

Real case. Uber's Genie launched in September 2023. By October 2024, it was running in 154 Slack channels, had answered more than 70,000 questions and had saved an estimated 13,000 engineering hours. Users rated 48.9% of its answers as helpful, and every answer includes options to ask a follow-up or escalate to the human on-call engineer.

What to copy. Publish honest metrics. An assistant that helps half the time can still save thousands of hours if escalation is easy.

Software development

8. Large-scale test migration

What the AI does. It rewrites code file by file to a new framework, runs the tests, linter and compiler, and retries with the error messages until each file passes.

Real case. Airbnb migrated about 3,500 React test files from Enzyme to React Testing Library in about six weeks, against an original estimate of a year and a half. About 75% of files were migrated in the first four hours and 97% after four days. Fewer than 100 files needed manual fixes. Harder files took 50–100 retries with prompts of up to 100,000 tokens that included related files and examples.

What to copy. Automated validation is what makes autonomy safe. Every change passed tests, linting and type checks before it was accepted.

9. Runtime and dependency upgrades

What the agent does. It analyzes an application, plans the upgrade to a newer language version or library, changes the code and runs the build and tests.

Real case. In August 2024, AWS reported that Amazon used its Q Developer agent to move tens of thousands of production applications from Java 8 or 11 to Java 17. The company estimated savings of more than 4,500 developer-years of work and $260 million a year from the resulting performance improvements.

What to copy. Agents perform best when the target state is precise. "Java 17 with all tests passing" is a far better goal than "modernize the codebase".

10. Coding agents in a delivery team

What the agent does. It implements scoped tasks, writes tests and reviews pull requests inside the team's own repositories and CI pipeline, under the same review rules as human engineers.

Real case. On the NEXUS threat intelligence platform for a US cybersecurity media company, a .wrk team of six to eight engineers worked with Claude as a coding agent. The team documented project-specific instructions for the agent, ran it in a security sandbox and added automated code review to the CI pipeline. The MVP launched on a fixed deadline while requirements were still changing. Read the NEXUS case study.

What to copy. Write down the project's conventions for the agent, and keep human review on every merge. Google's 2025 DORA research found that 90% of technology professionals use AI at work, while 30% have little or no trust in AI-generated code.

Content, research and marketing

11. Intelligence gathering

What the AI does. It finds new articles and advisories across many sources, extracts structured facts, removes duplicates and links related entities such as vendors, products and vulnerabilities.

Real case. The same NEXUS platform discovers content from hundreds of sources through feeds, sitemaps and listing pages, with a browser-based fallback for protected sites. Language models extract structured data, and the pipeline deduplicates by URL and content and resolves entities into a graph. Metrics cover scraping, language model latency and every pipeline stage, and the data feeds a public security events section on the client's website.

What to copy. Treat model calls as a production dependency with monitoring, rate limits and cost tracking.

12. Daily editorial briefings

What the AI does. It collects the last 24 hours of relevant news, classifies items by topic and region, drafts briefs and assembles the newsletter in the CMS for editors to review.

Real case. For a US cybersecurity publisher, .wrk built Aesop, a service that classifies incoming content, runs a separate stream for European regulators, companies and incidents and drafts the daily newsletter in WordPress with mapping to Marketo campaigns. Editors review every issue before it is sent. The first version went live in 10 working days, before a key vendor subscription expired. Read the Aesop case study.

What to copy. Keep the editorial decision with people and automate the assembly around it.

13. Grounded content research

What the agent tools do. They give a content generation system search and trend tools over the company's own indexed material, so drafts are built from real sources.

Real case. For a US law firm group running several websites on one CMS, .wrk built a content service in Go that combines keyword and semantic search over indexed content, along with a trend analyzer that scores topics. Both feed the firm's content generation tool, which keeps drafts tied to real indexed sources and avoids invented references. Read the Verba CMS case study.

What to copy. Give agents good search tools over your own data before giving them freedom to write.

14. SEO suggestions inside a CMS

What the assistant does. It suggests internal links and page titles to editors inside the CMS they already use.

Real case. An SEO software vendor's assistant suggested links to pages that did not exist and titles that were too long. .wrk restricted link suggestions to an approved list of URLs, added relevance rules and hard character limits for titles and built regression checks on previously failing cases, without changing the product's core architecture. Link suggestions now stay within the approved catalog. Read the AI SEO tool case study.

What to copy. Limit what the model is allowed to output. A closed list of valid options removes a whole class of errors.

Documents and legal work

15. Legal case preparation

What the AI does. It reads a bundle of case documents and drafts a structured opinion for an attorney to verify and complete.

Real case. A US personal injury law firm prepared insurance opinions from bundles of about 250 PDF pages of medical records, police reports and notes. A single chatbot prompt could not handle the volume or keep topics separate. .wrk split the work across five specialized assistants, each with examples, templates and edge-case instructions, and returned structured opinions with key conclusions in bold. Model usage costs about $3–4 per case bundle, and attorney throughput improved about 1.5 times. Read the legal document processing case study.

What to copy. Split a large task into smaller, specialized steps. Five focused assistants outperformed one general prompt.

What the 15 cases have in common

• Narrow scope first. Salesforce started with 200 users, and the telecom operator with two pilot sites and 15 intents.
• Automated checks. Airbnb's tests and compilers, the closed URL list in the CMS assistant and the editorial review gate all catch errors before they reach users.
• A human route. Every case keeps escalation, review or approval in place. Klarna's change of course shows what happens when cost drives the design alone.
• Measured results. Each company tracked resolution rates, hours saved or throughput, which made it possible to decide whether to expand.
• Fit for the task. Several of the most reliable systems are workflows with AI steps. Full autonomy was used where the task required it, such as open-ended code changes.

These patterns match broader forecasts. Gartner predicts that agentic AI will autonomously resolve 80% of common customer service issues by 2029, and also that more than 40% of agentic AI projects will be canceled by the end of 2027 because of costs, unclear value or weak risk controls. The cases above show what separates the two groups.

How to choose your first agent use case

1. List tasks your team repeats at least hundreds of times a month.
2. Keep the ones where success is easy to check, such as a resolved ticket, a passing build or an approved draft.
3. Check that the agent can reach the data and tools it needs through APIs.
4. Define what the agent may do alone and what needs a person's approval.
5. Run it in shadow mode or with a small group, and compare the results with your current process.
6. Expand only when the numbers justify it.

For budgeting, our guide to AI agent development cost breaks down what each part of an agent costs to build and run.