Field Guide Vol. 4: Enterprise AI Agent Case Studies — What Actually Happened

This month's Field Guide turns from vendor roadmaps to the deployments themselves: four companies that put AI agents into real workflows, what they reported, and what they'd tell you to watch for. Three are vendor self-reported case studies (treat the specific numbers as marketing-adjacent, not audited figures) and one is a widely corroborated reversal. All four are worth reading for the operational detail, not just the headline stat.

A note on sourcing: every figure below comes from a company's own published case study or mainstream business press, with links. We skipped several "171% ROI" and "12 case studies with proven ROI" aggregator roundups that kept surfacing in search — none traced back to a verifiable primary source, so none of their numbers made the cut.

Salesforce: Agentforce as its own first customer

Salesforce's "Customer Zero" program put its own Agentforce agents into Slack starting in February 2025, and by its own account the results weren't immediate — it took roughly six months of testing and tuning before adoption took off. By September 2025, 86% of Salesforce employees said they relied on Agentforce inside Slack to get answers or take action.

The breakdown by function, per Salesforce's own report: a Techforce IT-support agent had already handled 64,000+ requests, returning 17,000 hours to technical teams — the one figure described as already realized rather than projected. A Sales Agent was projected to save 203,000 hours annually; an Engineering Agent was projected to save 275,000 hours annually, which Salesforce equated to roughly 130 full-time engineers; and a BaseCamp onboarding/HR agent reached over a third of the company (about 25,000 employees), with no hours figure attached. Salesforce rolls these into a headline "500,000 hours saved annually" — worth noting that figure blends one realized number with several projections.

The more durable lesson is about data hygiene, not hours: Salesforce said its BaseCamp agent launched against a 2,000-article knowledge base, and "quantity didn't equal quality" — conflicting and outdated policy articles hurt accuracy. Cutting the source library down to 600 vetted resources is what actually improved reliability. Salesforce also said adoption varied sharply by department until it led with a task everyone encounters (password resets), which built trust before rolling out more specialized, team-specific agents.

Source: Salesforce Customer Zero: Deploying Agentforce in Slack

Moderna: 750 custom GPTs in two months

Moderna's deployment, documented directly by OpenAI, is less about one agent and more about seeding many. Within two months of rolling out ChatGPT Enterprise, employees across the company had built 750 custom GPTs; 40% of weekly active users had created at least one, and the average user logged 120 ChatGPT Enterprise conversations per week. Moderna's legal team reported 100% adoption.

This followed an earlier internal chatbot, mChat (built on OpenAI's API, launched in early 2023), which had already reached over 80% employee adoption — so the ChatGPT Enterprise rollout was a second act, not a cold start. Moderna's CIO, Brad Miller, framed the gap between AI ambition and AI results industry-wide: "90% of companies want to do GenAI, but only 10% of them are successful, and the reason they fail is because they haven't built the mechanisms of actually transforming the workforce to adopt new technology." Moderna's own mechanism had three layers: individual training (including an AI tutor), a collective layer (an internal prompt contest that surfaced a volunteer corps of "Generative AI Champions," plus office hours by business line and an internal forum with roughly 2,000 weekly active participants), and executive sponsorship from the CEO down.

The use cases that stuck were narrow and specific rather than general-purpose: a "Dose ID" GPT that helps the clinical team cross-check vaccine dose selection with sourced rationale and charts (still described as a pilot under validation, not a finished tool); a contract-summary GPT; a policy-lookup GPT; and tools that turn biotech terminology into investor-facing language for earnings materials. Notably, OpenAI's write-up gives no time-savings, cost, or ROI figure, and no account of what didn't work — a reminder that even well-documented vendor case studies rarely publish their failure modes.

Source: OpenAI: Moderna

ServiceNow: running agents on its own IT desk

ServiceNow's internal account is the most granular of the four, and the IT service desk is where it gets specific. Across more than 950 business applications and over 29,000 employees, the company says AI agents now handle 90% of employee IT support requests at a 96% resolution rate, which let it redeploy 85% of human IT support staff to higher-value work (serving HR, finance, and customer support teams) while the remaining 15% moved into managing the agents, handling escalations, and VIP support.

Company-wide, ServiceNow reports its AI agents support 400,000 workflows a year, freeing 3 million hours of capacity and an estimated $0.5 billion in annualized value — though, as with the other vendor cases here, no baseline or measurement methodology is published alongside those totals. Elsewhere in the business, ServiceNow cites a 53% productivity gain in server patch scheduling and says security risk assessments became 66% more efficient; customer-support case workflows are 37% automated, with no resolution-time figures given for that one.

The operational lesson underneath the numbers: none of this worked until the underlying data was consolidated. ServiceNow points specifically to merging logs, performance metrics, and traces into a single view, and to having both human and AI agents draw from one shared configuration database (CMDB) rather than separate sources of truth. The company frames the staffing change as redeployment, not replacement — which lines up with what both Salesforce and Moderna describe, and stands in direct contrast to Klarna's experience below.

Source: ServiceNow: How is ServiceNow using AI agents?

Klarna: the reversal that's more instructive than the launch

Klarna is the one entry here that isn't a vendor's own success narrative, and it's the one with the clearest lesson. In February 2024 Klarna said its AI customer-service assistant was doing the work of 700 full-time agents and handling 75% of chats (about 2.3 million conversations) within a month of launch. The company had frozen hiring in late 2023 to invest in AI and held that freeze for over a year, during which headcount fell 22% to roughly 3,500, mostly through attrition.

By 2025, per Bloomberg (via Entrepreneur), CEO Sebastian Siemiatkowski said the AI chatbots were cheaper but produced lower-quality support than human agents, and Klarna reversed course: it is now actively hiring customer-service staff again, recruiting for fully remote roles aimed at students, rural populations, and dedicated Klarna users. Siemiatkowski's framing: "Investing in the quality of human support is the way of the future for us," adding that customers need to know "there will always be a human if you want."

The contrast with the other three cases is the point. Salesforce, Moderna, and ServiceNow each describe redeploying humans toward escalation, oversight, and complex-case handling alongside agents — never fully away. Klarna tried something closer to full replacement in a high-stakes, trust-sensitive channel, and the quality gap became visible enough to force a public reversal. If you're scoping an agent rollout, the operative question this case study raises isn't "can the agent handle the volume" — it's "what happens to trust and quality at the edges the agent doesn't handle well, and is a human still reachable there."

Sources: Entrepreneur: Klarna Is Hiring Customer Service Agents After AI Couldn't Cut It on Calls

What ties these four together

Across all four: adoption took months of tuning, not a single launch event (Salesforce: ~6 months; Moderna: a structured multi-layer change program). Data quality and consolidation showed up as a precondition for reliability, not an afterthought (Salesforce's knowledge-base cut; ServiceNow's unified CMDB). Three of the four describe redeploying humans rather than eliminating the role, reserving people for escalation, oversight, and edge cases. And the one case that pushed furthest toward full automation in a customer-facing, trust-sensitive channel is the one that had to walk it back.

Glossary addition

Employee Zero / Customer Zero — a vendor's practice of deploying its own product internally, at scale, before or alongside selling it to customers, and publishing the internal results as a case study. Useful context for reading vendor-published figures: the deployment is real, but the vendor also has a direct incentive to report a success story.

Full glossary: The Autonomous Edge glossary

Past volumes — adoption data, the vendor landscape, security and governance patterns — are all in the archive: https://buttondown.com/TheAutonomousEdge/archive

If this was useful, forwarding it to one colleague who'd care is the best way to help it grow. Subscribe: https://buttondown.com/TheAutonomousEdge