Good evening. Here is what matters in AI today, and how to put it to work.
Agent adoption tripled in 2025 and real ROI is emerging, but today's MLflow vulnerabilities and Gemini's default data access remind us that trust infrastructure is not keeping pace.
~4 min read · last 12 hours
In today's issue
01
Business adoption of AI agents tripled this year as measurable ROI emerges
02
Amazon Bedrock AgentCore lets autonomous agents pay for APIs and web content
03
GitHub: Canvases make agentic workflows visible, steerable, and cost-efficient
04
Grab cuts mechanical analytics work from 44% to 30% with AI agents
05
Groq raises $350M at $3.5B valuation, pivots from chips to neocloud
Main story
Business adoption of AI agents tripled this year as measurable ROI emerges
Salesforce's Agentic Enterprise Index finds agent adoption has tripled in 2025, with industries now identifying the strategies that deliver real returns rather than just pilots.
Why it matters: Tripling adoption in a single year means your competitors are no longer just experimenting, so the question shifts from whether to deploy agents to how to do it safely and at scale.
What to watch next: Watch for the Salesforce index to segment ROI by agent autonomy level: the next meaningful signal will be whether fully autonomous agents outperform human-in-the-loop designs on cost-adjusted outcomes, or whether steerability consistently wins.
We are watching three distinct signals converge: real enterprise adoption data, new infrastructure for agent payments, and a practical workflow pattern for keeping agents steerable, all arriving on the same day and collectively confirming that agentic AI is moving from experiment to operational discipline.
Grab's share of mechanical analytics work, cut by AI agents in months · InfoQ
Watch · On the feeds
You can just keep the work moving
OpenAI
Sal Khan discusses MIT OpenCourseWare
MIT OpenCourseWare
The Signal
Agent adoption is no longer a leading indicator: it is a lagging one. By the time tripling shows up in an enterprise index, the early movers are already optimizing for cost and governance, not still debating deployment. The practical frontier has shifted to three harder problems: keeping agents steerable when they go wrong, giving them safe access to money and data, and locking down the ML infrastructure they run on. Today's MLflow critical vulnerability and Gemini's default-on data access are not isolated incidents. They are symptoms of a broader pattern where AI systems accumulate privileges faster than security and compliance teams can audit them. Engineering and product leaders should treat trust infrastructure as a first-class delivery item, not a post-launch concern.
All the best, the KYFEX team
“We have some really ambitious plans to help you work with AI in Chrome to get things done, and I'll have more to share soon”
TechCrunch
Quick hits
Agentic AI hits production: adoption, tooling, and infrastructure
Amazon Bedrock AgentCore lets autonomous agents pay for APIs and web content
A new AWS pattern connects AI agents to a payment wallet with spending guardrails, letting them transact with paywalled APIs, MCP servers, and web content using the x402 protocol.
Why it matters: Giving agents a budget and enforcing spend limits is a foundational governance primitive, and teams building autonomous workflows should design for it now rather than bolt it on later.
GitHub: Canvases make agentic workflows visible, steerable, and cost-efficient
A GitHub Copilot engineering post argues that chat alone loses agent work in the scroll, and that canvas-style interfaces restore human oversight by making intermediate steps inspectable and redirectable.
Why it matters: Steerability is an underrated production requirement for agents: if operators cannot see and correct what an agent is doing mid-task, costs and errors compound silently.
Grab cuts mechanical analytics work from 44% to 30% with AI agents
Grab deployed AI agents across analytics workflows and reduced the share of low-value mechanical analyst work from 44% in February to 30%, freeing analysts for higher-order tasks.
Why it matters: This is one of the cleaner published ROI numbers in enterprise AI so far, and the pattern of using agents to absorb repetitive analytical toil is directly replicable in most data-heavy organizations.
Model and compute infrastructure: new options, new money
Two significant infrastructure moves landed today: a production-ready open model tuned for agentic throughput, and a major funding round that signals how the compute layer beneath AI workloads is being restructured.
Groq raises $350M at $3.5B valuation, pivots from chips to neocloud
Groq raised $350 million and is repositioning from an AI chip company to a neocloud provider, expanding an Nvidia-powered data center footprint rather than betting solely on its own silicon.
Why it matters: The pivot signals that even purpose-built AI chip companies are finding the neocloud service model more defensible than hardware sales alone, which has implications for how teams should think about inference vendor lock-in.
Audit AI tool data-access permissions across your SaaS stack
I am an IT or security lead reviewing default AI data-access settings across our SaaS tools. For each tool I name, list: (1) what data it can access by default, (2) the specific setting or policy name to restrict that access, and (3) the blast radius if a prompt-injection or account-compromise attack exploited that default access. Tools to review: [list your tools, e.g. Google Workspace Gemini, Microsoft 365 Copilot, Salesforce Einstein, Slack AI]. Flag any where the default is broader than read-only on user-owned data.
Why it helps: Google Workspace's default-on Gemini access reported today is a reminder that default permissions in AI assistants are a governance gap most teams have not formally closed.
Before you ship it
The risk
AI assistants with default-on access to email, calendar, and documents can exfiltrate sensitive business data through prompt injection in a malicious attachment or meeting invite, and most organizations have not inventoried which tools hold this access.
Do this
Run a quarterly permissions audit across every AI assistant in your SaaS stack: document what data each can read by default, set access to the minimum required for each user role, and log the change in your compliance record.
Ready to ship AI, not just read about it?
KYFEX designs and builds production AI for teams that need it working, not just demoed. Tell us what you're working on and we'll bring the engineering.
