2025 in Review: What AI Actually Delivered for Customer Support
AI customer support in 2025 was a year of hype, genuine progress, and a few quiet reckonings. Let me give you the honest version.
At the start of the year, the narrative was intoxicating. Every major platform — Salesforce, Zendesk, Intercom, ServiceNow — had an “AI-first” customer support story. The pitch was roughly the same: AI agents handling 80% of tickets automatically, human agents freed for complex work, CSAT scores through the roof. The demos were impressive. The deployments were messier.
MIT Technology Review called it “the great AI hype correction of 2025.” I’d call it something more specific: the gap between AI capability and AI deployment closed — but slower than anyone admitted.
What Actually Worked in 2025
AI did genuinely transform certain support categories. Routine inquiries — order status, account info, password resets, billing questions — got handled faster and at lower cost. Companies with well-structured knowledge bases and clear resolution paths saw real deflection rates. The technology worked for structured, predictable problems.
Automated routing improved dramatically. The AI triage systems available in late 2025 are materially better at reading customer intent and routing tickets than they were 18 months ago. Routing accuracy directly impacts handle time and customer satisfaction.
AI-assisted agents — the “copilot” model — proved more durable than full AI replacement. Give a human agent AI suggestions, next-best-action guidance, and automated documentation, and you genuinely improve their performance and reduce their cognitive load. That’s a win that scaled.
What Disappointed
Full AI replacement for complex support interactions largely failed to live up to its billing. The moment you leave structured territory — product defects, edge-case configurations, emotionally charged situations — AI still struggles. Customers notice. Escalation rates tell the story.
Product support specifically remains stubbornly hard for AI to fully automate. When someone’s device isn’t working and they can’t figure out why, text-based AI conversations often generate more frustration than resolution. The issues that require seeing what the customer is looking at are the issues AI chatbots handle worst.
Visual support — actually seeing the customer’s product environment — still requires human or semi-human intervention. We wrote about how visual AI would transform product support this year. The infrastructure improved. But the deployment lagged the infrastructure.
The Number That Tells the Story
Salesforce CEO Marc Benioff announced plans to eliminate approximately 4,000 customer service positions, citing AI’s capability to handle the work. That’s a confident statement. But Salesforce generates more customer service revenue than almost any company on earth. They have both the incentive and the capability to make AI work at scale. Most companies don’t have their resources or their data.
For the average mid-market SaaS company or hardware brand, the math is different. AI still requires human supervision, careful training, and regular intervention. The people savings aren’t as dramatic as the headlines suggest.
The Honest 2026 Outlook
Here’s what I think actually happens in 2026. The companies that used 2025 to build — to train their AI on their specific domain, establish feedback loops, integrate AI into their actual workflows rather than just deploying chatbots — those companies will pull ahead. The gap between AI-mature support orgs and AI-surface-level ones will widen.
The commodity AI support tools will commoditize further. Differentiation will come from the quality of training data, the sophistication of escalation design, and — critically — the human experiences your team delivers at the moments that matter most.
2025 delivered a foundation. 2026 is when we find out who actually built on it.
For us at Viewabo, the lesson is clear: visual context in support isn’t optional anymore. If you’re still resolving product issues through text alone, you’re leaving resolution quality on the table. Visuals aren’t just better — they’re faster.