HAPPY FRIDAY,
Mid-July, which means Q3 is fully underway, summer Fridays are doing their thing, and somewhere a marketing leader is staring at a half-built agentic AI stack wondering if they moved too fast. Turns out, statistically speaking, they probably did. A new study across 107 enterprises found that more than half have already had a confirmed AI agent security incident or near-miss, which is the kind of number that tends to land differently when you realize most of those organizations are still letting agents share credentials like it's 2019.
In This Week's AI Fridays:
- The agent security gap is already a real number, and it's 54%
- Salesforce's Agentforce problem is everyone's agentic AI problem
- Why AI search citations and AI search recommendations are not the same thing, and why that matters
- Where martech is going once you look past the AI noise
- Quick hits: Netflix drops $587M on Ben Affleck's AI filmmaking startup, and other things that happened
The agent security gap: 54% of enterprises have had an AI agent incident
More than half of enterprises surveyed have already experienced a confirmed AI agent incident or near-miss, and the study that produced that number only looked at 107 organizations. The core problem is straightforward: companies are giving agents real access to real systems while security controls remain roughly where they were before agents existed. Shared credentials, minimal oversight, and a general assumption that the agent will probably be fine.
The Highlights:
- 54% of the 107 enterprises surveyed have had a confirmed AI agent incident or near-miss
- Most organizations are still allowing agents to share credentials rather than issuing agent-specific access controls
- Security frameworks have not kept pace with deployment timelines, which is a polite way of saying governance got skipped
- The gap is widening as agent deployments scale faster than the policies meant to manage them
The Takeaway:
Everyone focused on what the agent can do, and fewer people asked what happens when it does something it shouldn't. That question has now answered itself for 54% of the sample.
Salesforce's Agentforce struggles reveal a broader marketing AI problem
Salesforce's Agentforce adoption has been slower than anyone at Salesforce would prefer to admit, and the reason is not the product. The reason is that the organizations buying it are not operationally ready, and the data quality sitting underneath those deployments is not where it needs to be to make agents useful. It is a pattern showing up across marketing stacks broadly.
The Highlights:
- Agentforce adoption is lagging despite significant Salesforce investment in the product
- Poor data quality is consistently cited as the primary blocker, not the AI itself
- Marketing organizations are attempting to deploy agentic tools on top of infrastructure that was not built to support them
- Operational readiness, meaning process design and clean data, turns out to matter more than the technology layer
The Takeaway:
The bottleneck for agentic AI in marketing is not the AI. It's the decade of accumulated data decisions that went unfixed because there was always something more urgent.
AI in Marketing
How To Measure AI Search Visibility
The tools being sold for AI search visibility are mostly measuring citations, which sounds useful until you realize a citation and a recommendation are two different things. Getting mentioned in an AI Overview is not the same as being the answer the model trusts, and the gap between the two is getting wider as AI search matures. The article makes a reasonable case for shifting what you measure: not whether you appeared, but how you appeared and with what framing. For anyone running SEO or content strategy right now, the distinction is worth sitting with, as the old metrics will keep going up while the meaningful ones stay flat.
Look past AI to see where martech is going
AI is dominating the martech conversation to the point where everything else has gone somewhat overlooked, which is unfortunate timing because several of the other things happening in martech are worth paying attention to. Stack rationalization is accelerating, modern measurement is getting a genuine rethink, and modular content approaches are starting to replace the monolithic content production model that most organizations are still running. None of it is as easy to write a press release about as agentic AI, but taken together it is probably reshaping day-to-day marketing operations more than most of the agent announcements will. Worth reading if you are the kind of person who has to make actual budget decisions rather than just follow the narrative.
Enterprise AI Must Prove Its Value Beyond Deployment
The conversation in enterprise AI is shifting from "we deployed it" to "what did it actually do," which is a more uncomfortable conversation for a lot of organizations than the deployment phase was. The focus is moving toward measurable business value, workflow redesign, and the governance structures needed to scale without things going sideways. The deployment announcement was the easy part. Proving ROI on an ongoing basis while also redesigning the workflows the AI is supposed to improve is a different kind of project, and most organizations are still figuring out what that looks like in practice.
Australia moves to curb automated AI decision-making in government
Australia's new national AI plan puts limits on automated government decision-making and pairs that with a Labor push for digital duty of care legislation. It is one of the more substantive policy moves in a while, in that it addresses a specific use case rather than gesturing broadly at AI risk. For marketers and agencies, the direction of travel matters: governments limiting automated decisions tend to eventually expect the private sector to apply similar thinking, particularly in areas like personalization, targeting, and anything that affects consumer outcomes in a meaningful way. Worth watching even if you are not Australian.
Quick Hits
- Netflix paid $587M for Ben Affleck's AI filmmaking startup InterPositive, which is a sentence that would have landed differently five years ago. Read more
- OpenAI's CFO introduced an AI scorecard measuring useful work, cost per successful task, dependability, and return on compute. A reasonable framework for anyone who has been nodding along to ROI conversations without a clear way to run them. Read more
- Enterprises are buying AI compute infrastructure faster than they can track what it costs, which is a pattern that tends to resolve itself eventually, usually at board level. Read more
- A new AI company called Thinking Machines is positioning directly against Anthropic's intellectual brand identity. Competitive positioning in AI is getting interesting now that there are enough players to have positioning at all. Read more
- Chinese startup Moonshot AI released Kimi K3, a 2.8 trillion parameter open weight model. Genuinely impressive technically; whether US-based organizations will use it is a separate question with geopolitical dimensions that have nothing to do with the model's quality. Read more
- Google is rolling out Top Stories inside AI Overviews, live now on US mobile. If you care about news visibility in search, this one has implications. Read more
This article was obviously generated with AI but curated by a human, don't be weird about it.
Compiled by
Pete Bishop
Chief Innovation Officer, ZGM Modern Marketing Partners
Pete Bishop has spent the last two decades helping brands adopt new technology without losing the plot. He hosts Artificial Breakdown, a podcast and weekly newsletter that translates AI news into practical marketing decisions, and scans 50+ sources each week to write this issue.
Get this in your inbox
One email a week with the AI and marketing news worth your time.