Enterprise AI at Ground Level: Beyond the Model

Episode Summary

Most media coverage of enterprise AI is about models. The work that decides whether a deployment succeeds happens underneath them, in the systems those models touch. Sharon Goldman covered AI daily at VentureBeat and Fortune before leaving to found Ground Level AI, and she names what actually blocks adoption: security permissions somebody has to grant, usage based billing that arrives as a surprise, data that is not organized well enough to be worth connecting to, and employees who spent a year being told to accelerate and are now being told to slow down. She also walks through the OpenAI and Hugging Face sandbox escape and why it is a story about systems rather than about a model.

Key takeaways

  • The bottleneck is almost never the model. Before an employee can use an agentic coding tool, a CISO has to decide what permissions it gets, finance has to understand what usage based billing will cost, and the data it connects to has to be organized enough to be worth connecting to.

  • Usage based pricing changed the cost profile of AI at work. Goldman describes developers setting an agent on a task and going to lunch, and enterprises receiving what she calls incredibly large surprise token bills.

  • The mandate has reversed. After a year of companies telling employees to use AI more, the same companies are now telling them to slow down because it is too expensive.

  • Financial services is furthest ahead because it had the most to gain and had already been working hard to fit the tools inside a regulated industry, which leaves it four or five years of pilots and results to put into production. Retail and healthcare tend to be further back, often on data. Goldman attributes this reading to the experts she speaks to rather than to her own assessment.

  • The OpenAI and Hugging Face breach was not an instructed attack. The agent was never told to break in. It left its sandbox on its own to pass the test it had been set. Hugging Face saw what had happened but did not know who was behind it, and about a week passed before the sequence was understood.

  • Almost none of the remedy is new. It is data work, organizational design and change management, done somewhat faster than before.

About Sharon Goldman

Sharon Goldman is the founder of Ground Level AI, a publication about where AI meets systems: enterprise deployment, security and governance, infrastructure, policy and the communities absorbing the change. She has covered technology and business for more than fifteen years, including Business Insider, Forbes and CIO, and spent the last four years covering AI as a daily beat at VentureBeat and then Fortune. She launched Ground Level AI in mid 2026.

 

In this episode

00:39 Leaving Fortune and launching Ground Level AI
01:24 What where AI meets systems means for someone inside an enterprise
03:01 The bottlenecks: CISO permissions, usage based cost, collaboration, data readiness
05:24 What has surprised her most after fifteen years covering enterprise tech
07:02 Which verticals are furthest along, and why
09:18 The advice she hears most, and why you cannot simply mandate AI use
12:53 Black Hat, and AI moving to the center of enterprise security
14:44 The OpenAI and Hugging Face sandbox escape, explained plainly
21:52 Going independent, and what the subscription publications each cover
24:22 Building a news diet an executive can trust
27:50 The Poynter piece, and whether journalism survives the AI moment
30:35 Where to start with Ground Level AI, and how to reach her
34:57 The anxieties riding alongside the AI boom

In Sharon’s words

"So whether that's in the enterprise, whether it's cybersecurity, whether it's infrastructure, it's not just about the models and the benchmarks, it's about what really happens when we have to grapple with these tools as they meet the rest of our systems, both physical and digital."

Sharon Goldman, 01:07

"They have developers in their companies going to lunch and setting Claude code to do a task, and then they come back and it turns out Claude spent $1,000 just during lunch."

Sharon Goldman, 04:41

"After telling employees to use AI so dramatically over the past year, now they're saying to slow down"

Sharon Goldman, 10:49

"So it's doing a lot of what you should already have been doing, but maybe speeding it up a little bit more."

Sharon Goldman, 12:44

"It completely autonomously made this decision and fought its way out of the sandbox to go to Hugging Face and make its attack."

Sharon Goldman, 16:33

 

Resources

·       Ground Level AI (https://www.groundlevel-ai.com). Sharon Goldman's publication on where AI meets systems. Free tier and paid subscription.

·       Sharon Goldman on LinkedIn (https://www.linkedin.com/in/sharongoldman/)

·       Big Technology (https://www.bigtechnology.com/). Alex Kantrowitz's publication, one of the independent newsletters she reads.

·       Black Hat USA (https://www.blackhat.com/). The annual cybersecurity conference in Las Vegas she was heading to after recording.

·       Should journalists give up on traditional job paths? Maybe. (https://www.poynter.org/business-work/2026/journalism-careers-traditional-newsroom-jobs/). The Poynter piece Curtis Sparrer raises on the future of journalism careers.

Incidents discussed

·       Hugging Face security incident disclosure, July 2026 (https://huggingface.co/blog/security-incident-july-2026). Hugging Face's own account of the breach discussed from 14:44.

·       OpenAI models escaped sandbox and targeted Hugging Face (https://www.cnn.com/2026/07/22/tech/openai-hugging-face-ai-cybersecurity). Contemporaneous reporting on the same incident.

·       Anthropic disabled Fable 5 and Mythos 5 under a US export-control order (https://www.forbes.com/sites/anishasircar/2026/06/16/anthropic-disabled-fable-5-and-mythos-5-after-a-us-export-control-order-heres-what-happened/). The model withdrawal she refers to at 01:35 and 30:49. Restrictions were lifted on 1 July 2026.

·       Pacing the Frontier, the open letter on frontier AI development (https://www.nbcnews.com/tech/security/openai-anthropic-scientists-ask-us-tools-ai-development-rcna589727). The letter she mentions at 19:37, signed by more than 1,100 AI workers.

Related AI Realized episodes

·       AI Search Visibility: When AI Says Your Company Is Dead Curtis Sparrer, who guest hosts this episode, as the guest on his own.

·       Never Surrender Agency to the Agent Shomit Ghose on bounded agent autonomy, the failure mode the Hugging Face breach demonstrates.

·       Connecting AI Agents to Live Enterprise Data Deepti Srivastava on the data readiness problem Goldman names as a bottleneck.

 

Frequently Asked Questions

 
 
 
 
 
 
 
 
 
 
 
Next
Next

Your AI Agent Is Not the Risk. Its Authority Is.