AI Agent Adoption Just Tripled. The ROI Is Real.

Here’s the signal worth your attention today: the “are AI agents actually useful?” debate now has hard numbers — and the answer is yes. Below, what tripled, what it means, and the three risks that should temper your enthusiasm.

AI Agents

AI Agent Adoption Just Tripled — and the ROI Numbers Are Finally Here

What happened: Salesforce’s 2026 Agentic Enterprise Index — drawn from 400 businesses tracked over five quarters and ~5,000 survey respondents across nine markets — shows the average organization went from 5 AI agents (Feb 2025) to 13 (April 2026). Agent creation time dropped 53%, from 4 days to 1.9. Agentic actions grew from 0% to 15% of monthly actions. Crucially, 7 in 10 customer-service sessions are now handled autonomously by agents, up from zero in early 2024.

Why it matters: This is the first large-scale, vendor-backed dataset showing agents crossing from pilot to production — and doing so fast. The “so what” isn’t the headcount; it’s the direction: agent capabilities improved 350%, unique actions per agent doubled, and retail/consumer sectors saw output surge 60%. If your team is still treating agents as a novelty, the gap between you and early adopters is now measurable in quarters, not years.

What’s next: Watch the trust signal — weekly employee-agent sessions tripled year-over-year, approaching 8 per week, and 77% of shoppers report higher confidence when a “shopper agent” is present. The companies winning aren’t deploying more agents; they’re routing more real work through the ones they have and letting them take on adjacent responsibilities. That’s the playbook to copy.

Read the full data breakdown → ZDNet

OpenAI’s President Told CISOs to ‘Embrace Agents.’ Critics Noticed What Was Missing.

After the Hugging Face breach, Greg Brockman urged companies to hand security agents access to codebases and infrastructure — but didn’t address liability, blast-radius limits, audit trails, or rollback paths. Gartner’s Nader Henein put it bluntly: “I tend to recommend against taking advice from a party actively selling the solution to a problem they had a role in creating.” The takeaway for you: treat vendor security guidance as a sales pitch first, and demand accountability clauses before you hand an agent the keys. Source → Computerworld

Your MCP Servers Are Probably Leaking Secrets Right Now

The Model Context Protocol — the glue connecting AI assistants to your tools — holds real credentials, and security researchers have mapped five ways they leak: plaintext config files, credential sprawl, prompt injection, over-permissioning, and a supply-chain flaw (CVE-2025-6514) in a proxy downloaded 400,000+ times that enabled remote code execution. The sobering part: leaked MCP secrets let attackers act, not just read. Rotate credentials, enforce least privilege, and inventory every MCP server before an attacker does it for you. Source → The Hacker News

AI News

The ‘Free AI’ Era Is Ending — Get Ready for the Cost Shock

Harvard Business Review warns that enterprise vendors have been absorbing GPU, inference, and token costs to win customers — and that subsidized “unmetered” and “complimentary” AI features are about to flip to usage-based pricing. If your budget treated AI as free, it’s time to re-forecast before the invoice arrives. Source → HBR

ECB: An AI Market Correction Is ‘Likely’ — and It Won’t Matter If AI Succeeds

The European Central Bank compared today’s AI valuations to the dot-com era and concluded a sharp correction is likely, even if the technology delivers — because “overly optimistic investors bid up prices beyond fundamentals.” The message for operators, not speculators: build on durable unit economics, not on the assumption that AI hype keeps your runway warm. Source → Economic Times

Memory Prices Up 500% — and That’s Now an AI Problem

Zoho CEO Sridhar Vembu flagged that memory prices have risen 500% in 12 months, declaring the “free memory” era over — a squeeze that hits both AI training costs and the software developers building on top of it. It’s a quiet reminder that AI’s economics are tethered to hardware supply, not just model breakthroughs. Source → Business Today