Our analysts track the weekly developments reshaping AI and how it reaches users. Brands building AI products can partner with vetted AI companies to bring their ideas to life.
Five developments this week show implementation deciding the AI race:
- Anthropic and Blackstone debuted Ode, a $1.5B AI implementation firm, as Microsoft launched Frontier Company with 6,000 engineers.
- Bunkerhill's Carebricks runs 20+ live AI agents in one hospital system, proof that the implementation gap can be closed.
- 85% of companies run AI with no formal strategy or owner, per a Supermetrics survey of 435 marketing leaders.
- Moonshot AI released Kimi K3, the largest open-weight model yet, as Google's Gemini 3.5 Pro slipped months behind.
- Google shipped three cheaper Gemini Flash models, with its flagship 3.5 Pro still delayed months past the target.
Below, we start with what these events mean for the brands and agencies deploying AI, then recap how each one played out.
Our Take: Are AI Vendors Still Selling Software?
The biggest AI companies are now paying thousands of engineers to sit inside their customers' businesses and make the technology work.
We think that the product these companies actually sell is the engineering hours it takes to make AI work in your business.
What you buy now is the deployment itself, the integration into your systems, and the upkeep as these systems change.
So a demo and a benchmark score tell you almost nothing, because the real cost and risk sit in everything that happens after you sign.
Always ask who staffs your integration and who owns the outcome, because the answer decides whether the AI ever works.
Implementation Becomes the Product
Anthropic, Blackstone, and Hellman & Friedman launched Ode after Blackstone's portfolio companies kept needing consultants to run their AI.
Fractional AI, the boutique that the venture acquired in May 2026, now forms Ode's operational core alongside Anthropic's engineers.
Ode targets mid-sized companies that have the AI tools but can't run them well day to day.
"As [they] move from experimenting with AI to building it into their operations, they need partners with real implementation depth," Anthropic Head of Forward Deployed Engineering in the Americas Garvan Doyle said in a statement.
Microsoft placed the same bet two weeks earlier, embedding 6,000 engineers inside clients through its new Frontier Company.
Both companies read a market where capable models are common, and the integration work is the scarce part.
This scarcity makes implementation part of product strategy, the differentiator for winning the mid-sized buyers that everyone else overlooks.
Healthcare Shows AI in Action
Bunkerhill Health raised a Series B that pushed its total funding to $55 million, after growing revenue 20x over the past year.
The University of Texas Medical Branch alone runs more than 20 live agents on its Carebricks platform, across clinical, operational, and administrative work.
One agent flagged a life-threatening heart risk in its first month and routed the patient to a triple bypass. Another handled urgent lung findings 80% faster.
Both wins came from implementation, connecting AI to the right data, permissions, and process inside an existing workflow.
AI in healthcare has the least room for error out of any field, as it should be when lives are on the line.
So if hospitals trust agents with decisions this high-stakes, no brand can seriously claim its own work is too complex for AI.
Open Models Hand Companies Control
China's Moonshot AI released Kimi K3 this week, a 2.8-trillion-parameter open-weight model with a 1-million-token context window.
Moonshot calls it the world's first open 3T-class model, with full weights due July 27.
The release extends a pattern of Chinese labs closing the gap with U.S. frontier models on price and, increasingly, on raw capability.
Open weights give brands full control, so any company can download Kimi K3, modify it, and run it without a usage fee or someone else's roadmap.
This control ends the vendor lock-in of closed APIs, leaving the company with AI infrastructure it actually owns.
Google Gemini 3.5 Pro Stalls
Google released three new Gemini models on July 21, all cheaper and faster, with no sign of its delayed flagship.
The lineup covers Gemini 3.6 Flash for coding and agent work, a low-cost 3.5 Flash-Lite, and 3.5 Flash Cyber for security work.
Gemini 3.6 Flash cuts output tokens by 17%, which makes it cheaper to run than the model it replaces.
The flagship Gemini 3.5 Pro is months late, falling short of internal coding goals with no launch date after missing a May target.
In the same stretch, OpenAI shipped GPT-5.6 while Anthropic shipped Opus 4.8, Sonnet 5, and Fable 5.
China spent the same week on offense, as Xi Jinping opened its World AI Conference with a pitch for open, shared access.
A stalled flagship is a real risk for any brand's AI strategy, because the vendor you commit to can miss its own deadlines while rivals ship past it.
The Numbers Expose an Accountability Gap
Only 15% of organizations have an AI roadmap with success metrics, per a Supermetrics survey of 435 marketing leaders, so most AI budgets run on faith.
The same gap shows up at the top, where the Future of Life Institute's 2026 AI Safety Index gave no lab a grade higher than a C+.
Anthropic took the top C+, OpenAI and Google DeepMind scored a C, Meta a D+, and three labs failed outright.
No lab has settled who takes the blame when an agent gets a decision wrong.
And this fact leaves accountability as the soft spot in most companies' AI governance.
The model rarely decides whether AI works, so judge a vendor on the three things that do
- Ask how they do staff implementation. Ode and Frontier Company exist because model quality stopped being the bottleneck.
- Pick one workflow and measure it end-to-end before scaling. Healthcare's wins came from narrow, well-instrumented use cases, not broad rollouts.
- Write down who owns the AI outcomes now. The 85% without a roadmap will pay for that gap later, one way or another.
The models are converging, so the advantage now comes down to who runs them well.
The companies that put a name against every AI outcome will keep pulling away from the 85% who are still hoping that a purchase pays off on its own.
For a breakdown of Apple's OpenAI lawsuit and the first fully autonomous ransomware attack, check out last week's AI roundup.
If your AI vendor can't tell you who staffs your integration, do you actually know what you're buying?
These top AI companies help brands move from generic tools to specialized models trained on their own data.
