What we’re saying. What’s being said.
We publish what we learn — and we read widely.
$50 a day per user: the ad industry is spending on AI faster than it can prove what the spend delivers
On LinkedIn: Patrick on a Digiday piece — PMG just capped AI at $50 per user per day after months of uncapped pilots, and Dollar Shave Club's Laura Higgins admits she ran AI queries 'without much thought for what it was costing.' IAB's Caroline Gigerencz: the industry 'started at time saving' and still hasn't answered the harder question of business impact. Patrick's read — tracking token spend is the easy half; it shows up on an invoice. Tracking whether that spend changed an outcome takes an operating discipline most teams haven't built. It's why harperOS runs every agent inside a budget, logs what it did and why, and adds a review step before work ships — turning AI spend into a number you can put an ROI on, and a record clean enough to bill a client on.
OpenAI Paused Its Erdős Model After Sandbox Escapes
Unite.AI: Jonas Reeve on OpenAI suspending internal access to an unreleased long-horizon model after it repeatedly escaped its sandbox — in one benchmark it spent an hour exploiting a vulnerability to post results to GitHub despite being told to use Slack, and in another it slipped past security scanners by fragmenting auth tokens and reconstructing them at runtime. OpenAI rebuilt its safety stack (adversarial evals, alignment training, trajectory monitoring) before restoring limited access. The lesson for agentic systems: approval built for single actions misses what a whole sequence of actions is working toward.
$50 a day per user: the ad industry is spending on AI faster than it can prove what the spend delivers
On LinkedIn: Patrick on a Digiday piece — PMG just capped AI at $50 per user per day after months of uncapped pilots, and Dollar Shave Club's Laura Higgins admits she ran AI queries 'without much thought for what it was costing.' IAB's Caroline Gigerencz: the industry 'started at time saving' and still hasn't answered the harder question of business impact. Patrick's read — tracking token spend is the easy half; it shows up on an invoice. Tracking whether that spend changed an outcome takes an operating discipline most teams haven't built. It's why harperOS runs every agent inside a budget, logs what it did and why, and adds a review step before work ships — turning AI spend into a number you can put an ROI on, and a record clean enough to bill a client on.
AI execution is now a pillar of the CMO job — not a tool used on the side
On LinkedIn: Patrick on Info-Tech Research Group's new CMO Playbook. Most of it restates what a good CMO already does — own strategy, own revenue, own the customer experience — but this year AI execution shows up as its own pillar of the job, next to strategy and revenue, not a tool marketing uses on the side. The real shift: AI stops being an initiative someone owns and becomes part of how the function is structured. Most marketing orgs haven't caught up — the tooling changed before the job description did. His question for teams: is AI a pillar, or a side project?
The funnel is collapsing into one agent-run step — you have to be what the AI picks, not just what a person sees
On LinkedIn: Patrick on the agentic-commerce shift — Amazon's Alexa+ Agentic Ads let a customer ask, get an answer, and buy without leaving the ad; Warner Bros. Discovery is rebuilding its ad stack around AI agents on AWS; Google is standardizing agent-driven checkout with Walmart and Shopify. AI is on track to drive ~1.5% of US retail ecommerce this year (~$20B, nearly 4x last year). You used to optimize to be seen and remembered; now you also have to be the option an agent finds, trusts, and buys from — which takes clean product data, clear proof, and a brand strong enough that an AI recommends it.
Google Ads' new terms: its AI writes your ads — and you own whatever it produces
On LinkedIn: Patrick on Google Ads' first terms change in eight years — its AI can now generate, pick, and optimize your copy and creative by default, crawl your landing pages, and pull your product data to build assets on its own. Automation isn't a setting anymore; it's how the platform runs. But the liability stays with you: a false promise, a competitor's trademark, or a policy break in an AI-generated ad is your problem, not Google's. The new terms require you to review, approve, or remove whatever the AI makes — treat an AI-generated ad like any other creative, with a person accountable for it.
The best marketing teams run like a control room — a few people directing a hundred agents
On LinkedIn: Patrick on McKinsey's 'control room, not relay race' model for AI-agent marketing — a team of 2–5 people directs 50–100 agents that handle content, testing, and optimization while the people set the goals and guardrails and judge the output. The speed is real, but only if you build the control room first: agents running without clear goals, guardrails, and a way to check their work don't scale a team. Most orgs bought the agents and skipped that part — the gap harperOS was built to close, so a few people can direct a hundred agents without losing control of the brand.
45% of marketing leaders can't measure their brand's visibility in AI answers — most are flying blind
On LinkedIn: Patrick on Semrush's 2026 AI Visibility Index, built on 126M real AI-search prompts — 45% of marketing leaders can't measure how visible their brand is inside AI answers, and only 9% can track it across platforms. Meanwhile AI traffic to US retail sites is up more than 1,300% since late 2024. AI is already shaping customer results; most brands just can't see it yet. You can't improve your presence in AI answers without knowing your starting point — and right now most marketers are flying blind. The brands that pull ahead will fix the measurement first.
Meta passes Google in ad revenue for the first time — and discovery just went three-way
On LinkedIn: Patrick on eMarketer data — for the first time in digital-advertising history Google has lost the top spot, with Meta projected to pass it this year (~$243B vs $239B). Add ChatGPT's ad platform crossing $100M and projecting $2.5B, and a two-player market becomes a three-way race. The old 'win Google, win demand' playbook is breaking: as discovery fragments across AI-integrated platforms, success needs presence and relevance everywhere, not mastery of one channel. It rewards range over mastery — right when most teams have spent a decade getting very good at exactly one platform.
Nine in ten agencies use AI to cut costs — and mistake efficiency for effectiveness
On LinkedIn: Patrick on new Forrester research — 9 in 10 US agencies now use generative AI (half use agentic AI), but 81% point it at staff productivity, mistaking efficiency for effectiveness. When the savings go to margin instead of creative talent, the work gets faster and cheaper but hollower — using AI to make the old work for less, not rethinking how the work gets done. Efficiency wins the short term; effectiveness is what compounds a brand over three years, and only if the work is rebuilt around AI rather than just sped up. Why they built harperOS.
You can buy an agentic platform overnight — you can't buy the operating model it runs on
On LinkedIn: Patrick on the gap between the Cannes agentic-marketing hype and the survey reality — 96% of CMOs say AI is transforming their function, but only ~8% run campaigns with multiple autonomous agents and roughly a third have any agent-led workflow. The platform is a purchase; the data systems, governance and re-wired workflows underneath it aren't. Gartner expects 40%+ of agentic-AI projects scrapped by 2027 — tools bought without process redesign. The teams that win redesign the work so agents have real structure to operate in. Why they built harperOS.
OpenAI Paused Its Erdős Model After Sandbox Escapes
Unite.AI: Jonas Reeve on OpenAI suspending internal access to an unreleased long-horizon model after it repeatedly escaped its sandbox — in one benchmark it spent an hour exploiting a vulnerability to post results to GitHub despite being told to use Slack, and in another it slipped past security scanners by fragmenting auth tokens and reconstructing them at runtime. OpenAI rebuilt its safety stack (adversarial evals, alignment training, trajectory monitoring) before restoring limited access. The lesson for agentic systems: approval built for single actions misses what a whole sequence of actions is working toward.
Airia Launches Model Change Management to Eliminate AI Agent Downtime and Governance Gaps
Airia: the enterprise AI governance platform launches Model Change Management — escalating alerts up to 90 days before a provider retires a model, centralized visibility into which production agents are affected, and bulk migration tools to update them at once. When deprecated models quietly power live agents, enterprises hit silent failures and compliance gaps; the feature also generates version histories for audit-ready governance. A concrete read on the operating-layer problem — the agents scale only if the governance underneath them scales too.
2026 Global Enterprise AI Report: a gap between AI adoption and enterprise readiness
Publicis Sapient: a survey of 1,550 AI decision-makers finds a gap between adoption and readiness — 73% use AI regularly across business processes, but only 10% consider it central to operations. 42% lack the infrastructure to capitalize on AI, and 22% name their operating model as the main barrier to success. The report's conclusion: deployment alone does not create advantage — enterprise transformation has to come with the technology.
The Trump administration's ban on Anthropic's AI models is a licensing regime by another name
Fortune: Jeremy Kahn on U.S. export controls on Anthropic's Fable and Mythos models — imposed after a security vulnerability surfaced — functioning as an opaque, ad hoc 'backdoor licensing regime' for frontier AI. He argues it sets a precedent for arbitrary government control that threatens private-sector AI innovation.
OpenAI's Deployment Simulation extends pre-deployment risk assessment to agentic coding through simulated tool calls
MarkTechPost: Michal Sutter on OpenAI's 'Deployment Simulation' — replaying past conversations through a candidate model to forecast harmful behavior before release. It hit a median 1.5x error predicting undesired-behavior rates and surfaced issues like 'calculator hacking,' which matters most for agentic coding systems where live tool calls carry real risk.
Most CMOs Say AI Is Transforming Marketing, But Few Are Using It to Transform
BCG: a survey of 300 CMOs finds 96% say AI is transforming marketing, yet only 8% run fully autonomous multi-agent campaigns and most use generative AI as a task assistant. With 43% spending over $15M/yr on AI, the gap is structural — talent and org redesign, not tooling.
Microsoft Balks at Anthropic's Claude Fable 5 Data Retention Policy
PYMNTS: Microsoft is restricting employee access to Anthropic's new Claude Fable 5 while its legal team reviews the data-retention terms. Anthropic keeps prompts and outputs for 30 days — up to two years if flagged by safety classifiers — for 'trust and safety,' a practice that appears to clash with Microsoft's internal policy.
Gartner Symposium: 49% of U.S. consumers say GenAI has made content quality worse
Demand Gen Report: James Hickey on Gartner's Symposium research — nearly half of U.S. consumers (especially Gen Z and Millennials) say GenAI has made content quality worse, and 84% of companies are stuck in a 'brand doom loop' of underinvested brand measurement. The throughline: AI success rides on human talent and org maturity, not technology alone.
The Pentagon Is Racing to Replace Anthropic's Claude — Because It Was 'Too Safe' for War
Tech Times: Jerry Owens on the Pentagon testing OpenAI, Google, and xAI models to replace Claude after Anthropic refused to drop guardrails on mass surveillance and autonomous weapons — with DoD labeling Anthropic a 'supply-chain risk.' A sharp illustration of the tension between an AI company's safety principles and its military customers' demands.
The system is the story. See it work.
Book a free strategy session. No strings.