AI Agent ROI: Real Numbers from Klarna, Salesforce, and Cisco
The ROI claims in AI are everywhere. "We saved $X million." "We eliminated Y% of support tickets." "Our AI agents do the work of Z humans." Some of these numbers are real. Some are marketing math. And some are real but require significant context to understand what they actually mean.
Let me go through the three most-discussed public AI ROI cases and give you my honest read on each.
Klarna: the headline everyone cites
Klarna is probably the most-cited AI success story of 2025 and 2026. Their claim, made in a February 2024 press release and repeated consistently since: their OpenAI-powered customer service agent handles the equivalent of 700 full-time customer service agents, managing 2.3 million conversations in its first month, with customer satisfaction scores equal to their human agents and resolution times dropping from 11 minutes to 2 minutes.
What's real:
The automation numbers are real. Klarna did deploy a large-scale AI customer service system and it does handle the majority of routine inquiries. The 700 FTE equivalence figure comes from dividing total automated conversation volume by average conversations-per-agent-per-month, which is a legitimate calculation. The resolution time improvement is real. Routine inquiries like "what's my balance," "when will my payment be processed," and "how do I return an item" genuinely resolve faster through AI chat than through phone or human chat.
The financial savings Klarna claimed ($40 million annual profit impact) are directionally plausible based on the headcount equivalent, though the exact figure depends on assumptions about overhead, turnover costs, and what those 700 FTEs were making.
What requires context:
Klarna did reduce its customer service headcount. But the customer service function it automated was largely handling high-volume, highly repetitive queries that had already been systematically scripted and structured over years. Klarna's customer base is almost entirely digitally native (their product is buy-now-pay-later for e-commerce). Their customer inquiries arrive in writing, through an app, with clean data attached. This is essentially the ideal setup for AI customer service automation.
A traditional bank with phone-first customers, complex regulatory disclosures, and a customer base spanning 70-year-olds with little digital comfort doesn't get these results. Klarna's context was unusually favorable.
The second contextual point: Klarna also hired senior technical staff while reducing customer service staff. The net headcount reduction is real, but the narrative of "AI replaces 700 people" is simpler than the underlying reality of "AI replaced high-volume routine work while the company reallocated budget to other functions."
The honest ROI: Real, significant, and reproducible in companies with similar customer service profiles (digital-native customers, high volume, structured data, routine query distribution). Not a universal template.
Salesforce: Agentforce and the $1B ARR claim
Salesforce launched Agentforce in late 2024 and has been pushing it aggressively through 2025 and into 2026. The claimed metrics from their investor communications include $1 billion in ARR from Agentforce by end of FY2026, 8,000 deals closed in the first two quarters after launch, and internal deployment results showing a 50% reduction in agent workload for certain case categories.
What's real:
The internal deployment results are probably the most solid part of the story. Salesforce runs on Salesforce, and they've deployed Agentforce internally for their own customer support. The 50% reduction in case handling work for routine technical support queries is consistent with what other companies have found when deploying AI triage and first-response automation.
The $1B ARR figure is real revenue, though it's important to understand what it represents. Many Agentforce "deals" are customers adding the functionality to an existing Salesforce contract, often as part of an upsell conversation. The incremental revenue per customer is real. Whether customers are seeing the promised ROI from that incremental spend is a different question, and one Salesforce is still building a customer evidence base around.
What requires context:
Salesforce's go-to-market around Agentforce has been aggressive, and some of the early deployments have been pilots and limited rollouts rather than full production deployments. The number of "closed deals" doesn't map cleanly to production deployments at scale.
The ROI narrative Salesforce presents in sales conversations often involves comparing AI-assisted service to fully-staffed human service with no automation whatsoever. Most enterprise companies already had some automation, scripted chatbots, IVR systems, self-service portals. The incremental improvement over existing automation is real but smaller than the comparison to zero-automation would suggest.
The honest ROI: Salesforce customers deploying Agentforce for high-volume, structured customer service workflows are seeing real productivity improvements, typically 20-40% reduction in first-level case handling time and meaningful improvement in first-contact resolution rates. The "10x the productivity of your agents" claims require cherry-picked use cases.
Cisco: AI in networking and operations
Cisco has been more measured in their AI claims than the flashier software companies, which makes their numbers more credible. Their most detailed public case data comes from their AI Network Analytics deployment in their own global network operations and in disclosed enterprise customer deployments.
The specific claims: mean time to resolve network incidents reduced by 30-40% in deployments where AI-assisted root cause analysis was used. False positive rate on network alerts reduced by 22% using ML-based anomaly detection. Internal IT operations team able to handle 15% more infrastructure scale with the same headcount after deploying AI-assisted ops tooling.
What's real:
Network operations is actually a good domain for AI ROI measurement because the outcomes are precisely measurable. Mean time to resolve (MTTR) is a standard metric. False positive rate on alerts is measurable. Cisco has been deploying these capabilities in their own operations for years before going to market with them, which means the internal data is genuine.
The 30-40% MTTR reduction is consistent with academic literature on AI-assisted incident diagnosis, which typically finds improvements in the 25-45% range when models are trained on historical incident data from a specific environment. The key phrase is "trained on historical incident data from a specific environment." The model has to learn your network's baseline behaviors and failure patterns. Deploying a generic model without environment-specific training doesn't get these results.
What requires context:
Cisco sells the AI capabilities embedded in their hardware and software products, which means the ROI conversation is really about whether the premium for Cisco's AI-enabled tiers is worth it versus buying Cisco's base products or competitors. The productivity improvement numbers are real, but they're the improvement of Cisco's AI-enabled suite versus a non-AI baseline, not versus other vendors' AI offerings.
The 15% more infrastructure with the same headcount figure is worth unpacking. This wasn't from a single deployment. It's an average across Cisco's internal IT operations over two years of AI-assisted tooling deployment. Infrastructure complexity grows non-linearly as organizations scale, so holding headcount while growing 15% more infrastructure is a real operational win, but the baseline matters. A smaller or simpler network might see no benefit; a large, heterogeneous enterprise network sees more.
The honest ROI: Network operations AI delivers genuine operational efficiency in large, complex environments. The numbers are more modest than customer service automation (20-40% efficiency gains rather than "10x throughput"), but they're more consistently reproducible across different enterprise environments.
What separates real ROI from marketing math
After looking at enough of these case studies, a few patterns separate the credible numbers from the spin:
Specificity about the baseline. Real ROI claims specify what they're comparing against. "Reduced MTTR from 4.2 hours to 2.8 hours" is real data. "Reduced MTTR by 35%" without telling you the starting number could mean anything.
Unit economics, not totals. "We saved $40M" is a number that needs division. $40M saved over how many customers? Over how much infrastructure investment? Per agent or total? Claims that survive the "divided by what" question are more credible.
Time to value transparency. Real enterprise AI deployments take 3-9 months to reach steady-state performance. Companies that claim ROI from week-one deployments are either measuring something trivial or measuring before steady state.
Acknowledgment of what didn't work. Klarna, Salesforce, and Cisco have all quietly walked back some earlier claims or narrowed the scope of their success narratives. That's honest behavior. Companies that claim everything worked perfectly are usually hiding something.
The ROI is real in many cases. But it's real in specific conditions, with specific baseline characteristics, after enough time for the system to learn. Anyone promising immediate, universal ROI across all use cases is not telling you the full story.