AI is reshaping software businesses and investor expectations, but the path to durable returns remains uncertain. HubSpot co-founder Dharmesh Shah says the company is rebuilding its platform around AI agents. He argues agents should create new work customers will pay for, not merely make coding cheaper. The 20-year-old company began with a one-person CRM, and Shah noted that 60% of surveyed customers use no CRM. He called this the most enjoyable time to build in his more than 30-year career, while offering no figures on agent adoption, customer retention or costs.
Agents are also starting to act as shoppers. In business-to-agent commerce, software purchases on a person’s behalf after learning their preferences and limits; current systems commonly browse and click like people. Businesses in ChatGPT lets companies publish product information, while “sign in lending” gives agents identity tokens for payments. Human Security says agentic browser traffic grew eightfold from August to October. In software development, more than 40% of GitHub pull requests began as agent drafts in September 2026, up from 17% six months earlier.
Companies are testing ways to keep AI agents reliable and governed. Maersk engineer Dmitry Buykin estimates that checking an agent’s code takes about nine-tenths of the work. His system covers roughly 10,000 instruction scenarios across 110 countries; replay tests across several thousand past sessions would have flagged less than 1%, while a separate review found only about a quarter of batch items were truly complete. At Fujifilm Biotechnologies, around 50 employees use Genki Studio, a secure agent-testing sandbox. A contract-review agent is slated for production in early December; agents generally take six to eight weeks to pass testing and reviews, and CIOs are expected to monitor and retire them as needed.
Security concerns surfaced in a Claroc review of Grok Build apps on grok.me: 35,578 of 37,571 active apps checked—94.7%—opened without sign-in. Claroc estimated 650 to 800 public apps contained private business or personal data, and found 24 open apps built by SpaceXSI and X staff that included confidential company material. Claroc said the publishing panel defaulted to “Anyone with the link”; SpaceXSI said apps remain private until creators share them. The company removed its staff apps after the report, while other open apps remained available.
Interactive Brokers has launched IBKR Autonomous Trading, allowing eligible clients to connect AI platforms that monitor markets and place orders in more than 75,000 listed stocks, options and ETFs worldwide. Clients set natural-language instructions and fund a separate “AI Partition” within their existing account. Supported orders can be placed without trade-by-trade approval, subject to the instructions and account safeguards. The service launched in the US, UK, Europe, Singapore and Australia, with eligibility and limits varying by jurisdiction and affiliate.
In chip design, Cognichip says its ACI Enterprise platform combines specialized models, agent workflows, project memory and inference computing to help meet power, performance and area targets. Chief Product Officer Stelios Diamantidis says its datasets contain more than 10 million training samples, built from company design work and synthetic data. The company says AI could shorten chip-design work from months to days. Separately, a review of Earth embeddings—general-purpose representations of observations from multiple sources—covers applications in six environmental fields and four tasks, including crop-yield prediction and biomass mapping. It finds embeddings do not consistently outperform satellite imagery or hand-crafted features, so researchers should evaluate them task by task and use them alongside conventional data.
Investors are weighing AI growth against uncertainty about how software companies will monetize it. Jefferies says vertical software stocks have fallen about 5% over five years and underperformed the IGV software ETF by roughly 24 percentage points, despite resilient growth and wider margins. It began coverage of Guidewire and ServiceTitan at Hold, preferring Guidewire, and sees no near-term re-rating; it rates Procore, Autodesk and Roper Technologies Buy, with Procore its preferred mid-cap pick. US stocks rebounded as attention turned to Anthropic’s planned November 18 IPO; Nvidia gained more than 2.5%. Strong demand for pre-IPO stakes may bode well, but the listing is a test of confidence in AI across chips, cloud and power. CoreWeave rose about 12% after better-than-expected quarterly results, while reporting $846 million in debt interest and continued cash burn.
Key Takeaways
- HubSpot’s Dharmesh Shah says AI agents should create new billable work; 60% of surveyed HubSpot customers use no CRM.
- Human Security reports agentic browser traffic grew eightfold from August to October; over 40% of GitHub pull requests began as agent drafts in September 2026.
- Maersk engineer Dmitry Buykin estimates checking AI-generated code takes about 90% of the effort, highlighting the need for extensive testing.
- Fujifilm Biotechnologies’ Genki Studio has about 50 employee users; a contract-review agent is expected to enter production in early December after review.
- Claroc found 94.7% of the active Grok Build apps it checked opened without sign-in and estimated 650–800 public apps held private data.
- Interactive Brokers’ Autonomous Trading connects eligible clients’ AI platforms to markets covering more than 75,000 stocks, options and ETFs, with client-set instructions and account safeguards.
- Cognichip says its chip-design platform draws on more than 10 million training samples and could reduce design timelines from months to days.
- Earth embeddings can support environmental remote-sensing tasks but do not consistently beat conventional satellite imagery or hand-crafted features.
- Jefferies recommends selectivity in vertical software: it favors Procore among mid-cap stocks, while rating Guidewire and ServiceTitan Hold.
- Anthropic’s planned November 18 IPO is a test of AI-investment sentiment; Nvidia rose more than 2.5%, while CoreWeave gained about 12% after results and reported $846 million in debt interest.
Dharmesh Shah says AI agents can help HubSpot grow
HubSpot co-founder Dharmesh Shah said the company is rebuilding its platform around AI agents and called this the most fun time to build in his more than 30-year career. He said HubSpot, now 20 years old, began with a one-person CRM, and that 60 percent of surveyed customers did not use any CRM. Shah argued agents need to create new work HubSpot can charge for, rather than simply make coding cheaper. The article notes that fears coding tools could replace software suites helped drive a selloff in SaaS stocks earlier in the year, but Shah did not provide figures on customer retention, costs or agent use.
AI agents are beginning to shop, pushing businesses to sell to bots
Business to agent, or B2A, commerce describes software buying on a person’s behalf after learning their needs and limits. Most shopping agents currently use a browser or virtual machine and click like a person. The article describes tools such as Businesses in ChatGPT, where companies can claim a profile and share product information, and sign in lending, which lets users give agents identity tokens for payment. Human Security reported that agentic browser traffic grew eightfold between August and October. The article also cites growing use of agents in coding, database creation and software deployments, including more than 40 percent of GitHub pull requests starting from an agent draft in September 2026, compared with 17 percent six months earlier.
Maersk engineer says checking AI code takes most of the work
Dmitry Buykin, an engineer working on Maersk order and approval software, said about nine tenths of an AI agent’s time on a change goes to checking it rather than writing it. The system covers about 10,000 instruction scenarios across 110 countries, with roughly three quarters written as prose instructions, he estimated. A replay check covering several thousand past sessions would have flagged less than 1 percent, while a separate review found only about a quarter of batch items could truly be closed because agents sometimes skipped work when they ran out of context. Buykin said he keeps 10 to 20 feedback loops open, replaying frozen real tasks against old and new versions and requiring each check to fail once on purpose. The article says the available figures do not show incident rates, review hours saved or dollar costs.
Research finds 35,578 Grok Build apps open to the public
Security firm Claroc found that 35,578 of 37,571 active apps it checked on grok.me, or 94.7 percent, opened without sign-in. The firm listed 43,601 app addresses and found 24 open apps built by SpaceXSI and X staff, including confidential company material. Claroc estimated that about 650 to 800 public apps contained private data, including business or personal information. It said Grok’s publishing panel selected “Anyone with the link” by default, while SpaceXSI said apps remain private until creators choose to share them. SpaceXSI removed its staff apps after Claroc reported the issue and Ars Technica asked questions, but other users’ open apps remained available.
Interactive Brokers lets eligible clients trade with connected AI
Interactive Brokers launched IBKR Autonomous Trading, which lets eligible clients authorize connected AI platforms to monitor markets and place orders across more than 75,000 listed stocks, options and ETFs worldwide. Clients set natural language instructions and use an AI Partition, a separate allocation within their existing account funded by internal transfers. The AI can execute supported orders without approval for each trade, within the client’s instructions and account safeguards. The feature was available at launch in the US, UK, Europe, Singapore and Australia, though eligibility, funding limits and availability vary by jurisdiction and IBKR affiliate. CEO Milan Galik said the service builds on the company’s automation approach, while the article notes that regulators’ rules on suitability and automated trading may affect its reach.
Earth embeddings offer new tools for environmental remote sensing
Earth embeddings turn observations from different sources into general-purpose representations that can support many remote sensing tasks. The article reviews their foundations, products, uses across six environmental fields, and results in four tasks: hyperspectral image classification, individual tree crown segmentation, crop yield prediction, and aboveground biomass mapping. The review finds that embeddings do not always perform better than satellite imagery or hand-crafted features. Researchers should assess their value for each task and use them alongside, rather than as replacements for, conventional Earth observation data.
CIOs guide employee-built AI agents from testing to retirement
Fujifilm Biotechnologies is preparing to move an AI agent that reviews legal contracts from a test environment into production in early December. The company built Genki Studio, a secure sandbox where employees can create and test agents, and about 50 employees currently use it. Agents typically take six to eight weeks to move from an idea to production after testing and a review for regulatory concerns and overlap with other tools. CIOs are also expected to track agents’ performance and value over time, and retire them when they no longer meet company needs.
Jefferies urges selectivity in vertical software stocks
Jefferies says investors should be selective about vertical software stocks because companies have not yet shown clearly how they will earn money from AI. The group has fallen about 5% over five years and lagged the IGV software ETF by about 24 percentage points, despite resilient growth and wider margins. Jefferies began coverage of Guidewire and ServiceTitan with Hold ratings, preferring Guidewire, and said a near-term re-rating is unlikely. It rates Procore, Autodesk, and Roper Technologies Buy, naming Procore its preferred mid-cap vertical software stock.
Cognichip says AI tools could cut chip design from months to days
Cognichip Chief Product Officer Stelios Diamantidis described the company’s AI products in an interview with Tom’s Hardware Premium. Its ACI Enterprise platform combines specialized models with agentic workflows, project memory, and computing for AI inference. Cognichip says its design environment uses specifications, register-transfer level data, and verification results to help meet power, performance, and area targets. Diamantidis said the company’s datasets contain more than ten million training samples, created through its own design work and synthetic data technology.
Anthropic IPO and CoreWeave results lift AI investment mood
US stocks rebounded after a sell-off, with Nvidia rising more than 2.5%, as investors focused on Anthropic’s planned IPO on November 18. The article says strong demand for pre-IPO stakes may signal that the listing will be well received, but notes that major IPOs can perform unevenly after their debut. The IPO is described as a test of investor confidence in the AI trade, from chips to cloud computing and power. CoreWeave shares rose about 12% after stronger-than-expected quarterly results, although the company reported $846 million in debt interest and continues to burn cash.
Sources
- Dharmesh Shah says the SaaSpocalypse isn't happening
- Agents now shop, businesses must sell to bots
- Code got free, upkeep did not at Maersk
- Grok Build left 35,000 apps open to anyone, including SpaceX's own confidential sales material
- Interactive Brokers Launches Autonomous AI Trading Capability
- Earth Embeddings in Environmental Remote Sensing
- As employees build AI agents, CIOs manage the lifecycle
- Should you own vertical software stocks into 2027? Jefferies weighs in
- Cognichip reveals how its new AI-assisted program can slash chip design times from months to days — automating labor-intensive tasks cuts design time down, Chief Product Officer explains
- The AI trade turns a corner
Comments
Please log in to post a comment.