Google Cloud's catalog of customer AI stories was first published April 12, 2024 and last updated April 22, 2026. Over that time its framing has grown bolder. Google now says companies are "firmly in the era of the agentic enterprise." Read the entries looking for agentic AI in production, though, and they split into two kinds. Some name a company, describe what the agent does, and give a number. Many describe a capability and stop there.
It helps to pin down the word first. Anthropic separates two kinds of agentic system. Workflows are "systems where LLMs and tools are orchestrated through predefined code paths". Agents "dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks". Keep that line in mind for each example below: the question is how much of the job the system decides for itself.
Klarna: Two-Thirds of Support Chats in Month One
In its first month live, Klarna's AI assistant, powered by OpenAI, handled 2.3 million conversations, two-thirds of Klarna's customer service chats. Klarna announced this on February 27, 2024. The tasks matter more than the headline number. The assistant handles refunds, returns, payment-related issues, cancellations, disputes, and invoice inaccuracies. It also gives real-time updates on your outstanding balances and upcoming payment schedules. Those tasks act on account data, which is more than an FAQ bot does.
Klarna published these figures:
- Workload: the equivalent work of 700 full-time agents
- Speed: errands resolved in less than 2 mins compared to 11 mins previously
- Accuracy: a 25% drop in repeat inquiries
- Reach: 23 markets, 24/7, more than 35 languages
- Satisfaction: on par with human agents
Read the money line carefully. The assistant was estimated to drive a $40 million USD in profit improvement in 2024. Klarna made that forecast after one month. Customers can still choose to interact with live agents, so humans remain available as a fallback. CEO Sebastian Siemiatkowski also used the launch to call for a considerate, informed and steady stewardship of AI's effects on society. Press releases rarely say that.
Hershey's Agnes: A Small Job With a Clear Count
Hershey Entertainment & Resorts runs Agnes on Google's Customer Engagement Suite, and it handles over 25% of employee call-offs, saving eight hours of staff work per day. Agnes is an agentic AI Interactive Voice Response system. It identifies employees by caller ID, greets them by name, and processes call-off requests, late notifications and schedule questions.
This kind of AI agent use case gets too little attention. Nobody gives a keynote about shift call-offs. Still, the scope is tight, the inputs are structured (who is calling, and for which shift), and the result can be counted. It is also something a mid-sized employer could realistically copy. The limits are just as clear. Most call-offs still go another route, and the entry gives no cost or savings figure.
Honeywell: Named, but No Number
Honeywell's Smart Shopping Platform, developed in partnership with Google Cloud and 66degrees, runs on Gemini. It gives retailers real-time insights into in-store compliance, inventory accuracy, and associate productivity. The deployment is named and real, but the entry gives no performance figure. It also doesn't say what the system does by itself and what it simply flags for staff. Insights are not actions. Until Honeywell publishes before-and-after results, read this as an early project story, not proof.
Honeywell's own comments on the Smart Shopping Platform describe the problem it addresses and the benefit to retailers, saying it "creates a better experience for shoppers" and gives retailers an "out of the box" AI solution they can implement "without having to maintain a team of AI experts" quote, but no figure accompanies that statement.
Product Pitches Aren't Deployments
Other entries on Google's list describe what a product can do, not what a customer measured. The Corvic AI Agent promises "production-ready decision intelligence — no pipelines and no data engineering required." Aible is building long-running enterprise AI agents, running Nemotron 3 and NemoClaw agents on BigQuery and serving them on Cloud Run. Mistral AI offers Mistral Large 24.11 for complex tasks like agentic workflows on Vertex AI. Thoma Bravo gives its portfolio companies access to the Gemini Enterprise platform for agentic AI creation.
These entries describe tools, infrastructure and access. None of them reports a customer result with a number.
| Deployment | What it does | Published number | Source |
|---|---|---|---|
| Klarna AI assistant | Refunds, returns, disputes, balances | 2/3 of chats (2.3M conversations); work of 700 agents (klarna.com) | Klarna |
| Agnes (Hershey Ent. & Resorts) | Voice call-offs, late notices, schedules | >25% of call-offs; saves 8 hrs of staff work/day (cloud.google.com) | Google Cloud |
| Smart Shopping Platform (Honeywell) | In-store compliance, inventory insights | Not published | Google Cloud |
| Corvic AI Agent | Decision intelligence on company data | Product description only | Google Cloud |
| Mistral Large 24.11 | Model offered for agentic workflows | Product description only | Google Cloud |
The dates show why most of the hard evidence is old. The stack has kept changing since the first numbers came out:
| Date | Event | Impact |
|---|---|---|
| Feb 27, 2024 | Klarna publishes assistant results | First-month figures on record |
| Apr 12, 2024 | Google's list debuts at Next '24 | Named customer use cases |
| 2024 | Grounding with Google Search GA | Gemini outputs get fresh search data |
| Jun 2024 | Gemini 1.5 Flash generally available | 1 million-token context window |
| Apr 22, 2026 | List updated for Next '26 | "Agentic enterprise" framing |
Recommendation: Where to Point Your First Agent
Klarna's results show what is possible. Hershey's show what most companies can realistically copy. Klarna's case is the strongest published evidence that customer-facing agents can take on real volume. Even so, its profit figure is a forecast, and it kept human agents available. Agnes shows a narrower path that is easier to defend: one job, structured inputs, and a count of hours saved.
Choose a Klarna-style rollout if you have high support volume, repetitive tasks such as refunds, and clean account data. Choose a Hershey-style project if you need a measurable win inside one department before asking for a bigger budget. Treat Honeywell-style entries and product pitches as ideas to test until someone publishes a result.
Before committing, measure the task as it runs today. Without a before number, you can't show that the agent is working.
Sources
- anthropic.com – "Building effective agents | Anthropic" (2024)
- cloud.google.com – "cloud.google.com"
- klarna.com – "Klarna AI assistant handles two-thirds of customer service chats in its first month | Klarna International" (2024)
- cloud.google.com – "Vertex AI offers enterprise-ready generative AI | Google Cloud Blog" (2024)
- honeywell.com – "Honeywell Unveils AI-Enabled Technology to Personalize In-store Shopping with Google Cloud" (2026)