1Password has launched AI Spend and Consumption Management, a new feature within its SaaS Manager platform designed to give IT and finance teams real-time insights into AI service usage and spending from providers like Anthropic, Cursor, and OpenAI. This move expands 1Password's portfolio beyond password management into enterprise AI cost control, addressing the unpredictable and consumption-based pricing models of AI, which traditional software budgets struggle to handle. The tool consolidates token-level data across vendors, offers budget controls, and breaks down usage by team, user, and model, aiming to prevent unexpected spikes in AI spending caused by agentic workflows. This new capability reflects growing demand as enterprises face rapid growth in AI token consumption, with forecasts predicting a 24-fold increase by 2030. 1Password integrates this feature with its established identity security and SaaS governance foundation, ensuring visibility into who is spending and whether the expenditure delivers value. The launch is timely, as AI costs become a significant concern in enterprise budgets, echoing past challenges seen with cloud infrastructure costs. Currently available in public preview, the product will gain broader availability in fall 2026.
Slopsquatting is an emerging security threat fueled by AI coding assistants producing hallucinations—fake software package names that attackers can register with malicious code. Unlike traditional typosquatting, where misspelled packages are targeted, slopsquatting exploits AI-generated fictitious package names that seem plausible, making detection difficult. When developers unknowingly incorporate these fake packages, malware can be injected directly into their codebases.
This new risk arises because large language models (LLMs) tend to generate the most statistically likely output rather than verifying accuracy, resulting in frequent hallucinations. Proprietary AI models show lower hallucination rates compared to open-source ones, but no system is immune. As AI-assisted coding grows more common—with many developers integrating AI into their daily workflows—the risk surface for slopsquatting expands.
To reduce risk, developers should verify package existence against official repositories, and organizations must deploy automated checks and threat intelligence to catch malicious packages early. This evolving threat underscores the importance of vigilance and security measures in AI-supported software development.
— Zac Amos, Features Editor at ReHack
Model routing, a crucial part of enterprise AI, optimizes task assignments to AI models for better speed and cost efficiency. Traditional routing relies on static rules or trained classifiers, which can’t adapt to real-world changes or learn from outcomes. A novel open-source framework called Agent-as-a-Router overcomes these limits by using a dynamic, memory-driven agent. It employs a Context-Action-Feedback (C-A-F) loop to learn from each task's success or failure, improving routing decisions on the fly.
ACRouter is the practical implementation of this dynamic routing approach. It features three key components—Orchestrator, Verifier, and Memory—working together to select the best AI model for each task based on past performance and real-time feedback. The system integrates with actual execution environments to verify outcomes and update its strategy continuously.
Benchmark tests on coding and agentic tasks showed ACRouter outperforms static routing methods and expensive default-to-premium strategies. It balances cost and performance effectively, achieving a 2.6x cost reduction compared to always using the costly Opus model. This makes ACRouter a powerful tool for enterprises to harness diverse AI models smartly without blindly paying for top-tier models every time.
While ideal for verifiable tasks like coding or data retrieval, the framework is less suited for subjective or low-volume tasks. The code and models are openly available on GitHub and Hugging Face, supporting compatibility with popular AI models like Claude Code and Codex.
This advancement offers enterprises a self-optimizing AI routing system that adapts to user behavior, foundation model evolution, and shifting workloads, pushing the boundaries of cost-effective AI operations.
DeepSeek recently cut prices on its V4-Pro model by 75%, a move that initially appeared to be a major win for enterprise AI vendors and developers. However, this price reduction has not guaranteed improved margins because agent workflows consume tokens far faster than the rate at which inference costs are declining. Unlike traditional chatbots that process one user query with one model call, AI agents involve multiple costly operations such as planning, retrieval, tool use, and verification, leading to a significant increase in token consumption — often 100 times more per user query.
This token amplification creates a challenge for the current AI business model, which relies on seat-based SaaS pricing. Heavy users of agentic workflows can incur infrastructure costs that exceed their subscription fees, resulting in negative gross margins. Enterprises adopting these AI agents face mounting expenses, forcing a rethinking of cost structures and product architectures.
To manage this, companies must adopt strategies like cost-aware routing, prompt caching, context trimming, and speculative decoding to control inference expenses. Treating inference cost as a first-class metric, budgeting carefully, and negotiating volume commits are essential steps to sustain margins. The future success of AI-native companies hinges not on the cheapest models but on intelligent agent orchestration that balances capability with cost awareness.
The 100x problem highlights a critical turning point where architecture decisions directly impact financial outcomes, marking a new phase in enterprise AI economics.
Enterprise AI teams are granting agents increasing autonomy just as confidence in automated testing methods is waning. According to a June 2026 survey of 157 enterprise respondents, half of enterprises have deployed AI agents or large language model features that passed internal checks but still caused customer-facing failures, with a quarter experiencing multiple such incidents. Despite this, 66% of companies allow some deployments without human review or plan to within the year, even though only 5% fully trust these automated evaluations. This gap—where agent autonomy grows faster than verification capabilities—is leading enterprises to ship agents first and develop control systems later.
Traditional software testing falls short for AI agents, which take varied actions and outcomes can differ each time. The biggest reason enterprises distrust automated evaluation is poor alignment with real-world results, followed by bias, explainability, and privacy concerns. Experts emphasize the importance of repeatability, running multiple tests and evolving evaluation sets based on incidents. While some low-risk tasks can tolerate more autonomous agents, high-risk actions require stricter oversight. Larger companies tend to push for zero-human review faster but also face more failures. The key takeaway for leaders is that removing humans from the loop increases uncertainty unless strong evaluation and regression testing become a priority. The market drives increasing autonomy, but success will favor those balancing speed with reliability.
A recent survey by VentureBeat Research involving 573 technical leaders from enterprises with 100+ employees reveals a striking trend: 86% report that their GPUs for AI workloads operate at 50% capacity or less. This inefficiency coincides with enterprises deploying AI agents faster than they can implement control measures, such as identity verification, output evaluation, cost tracking, context management, and orchestration. Many companies are now retrofitting their systems with control layers and budgeting for vendor changes within the next year. Despite the rapid adoption of AI agents, the majority are single-prompt chatbots rather than true autonomous multi-step agents, impacting necessary control requirements and costs.
The survey also highlights risks, with over half of companies experiencing security incidents linked to credential sharing between agents, and 57% recording errors due to inconsistent or missing business context. Enterprises are increasingly prioritizing agent technology portability, motivated in part by regulatory impacts and new open-weight AI models promising greater control. Vendor lock-in concerns and the dominance of provider-native solutions shape spending and procurement strategies. The report underscores the need for enterprises to optimize hardware utilization, validate AI results against real-world outcomes, and govern the data definitions agents rely on to improve AI deployment effectiveness and security.
An increasing number of enterprises have experienced AI agents providing answers with full confidence that turn out to be incorrect due to missing or inconsistent business context. According to a recent survey, 57% of businesses with over 100 employees reported this issue, often traced back to outdated or incomplete context given to AI agents rather than model errors. Many enterprises still rely heavily on document retrieval systems to supply context, which often leads to inaccuracies after deployment due to prioritizing ease of integration over retrieval accuracy.
The solution lies in adopting a governed context layer—a unified, consistent model of business data that AI agents reference instead of independently deriving context each time. However, 75% of enterprises have yet to implement this solution, with 41% not having started on building one at all. Interestingly, companies that have experienced confident yet wrong AI answers are more proactive in developing or running such context layers.
Major tech vendors are developing their own versions of these context layers, though architectures vary widely. Some focus on catalog metadata and query behaviors, others on real-time business ontologies or edge-based memory, illustrating the market's fragmentation. Analysts agree that governed, current, and low-latency context is essential for AI agents to move beyond errant guesses and perform reliably in production.
For enterprises, this means relying solely on document retrieval for context is insufficient, and investing in semantic context layers is becoming critical. No single vendor dominates this space yet, so integration across solutions is expected. Notably, enterprises hit by repeated AI confidence failures are more actively seeking new solutions. The evolution of these context platforms will be a key topic at VB Transform 2026, highlighting the race to close the critical context gap in AI deployment.
A study evaluating 67 advanced AI models from 21 providers reveals a critical flaw in the common enterprise strategy of combining multiple AI models to reduce failure. Known as the "co-failure ceiling," this flaw shows that while models may fail on different questions individually, there exists a hidden rate where all models fail simultaneously, which is often underestimated by 2.25 times using traditional pairwise error correlation methods. This means multi-model orchestration often incurs high costs in complexity, latency, and governance without delivering the expected reliability improvements.
Common architectural approaches like model routers, cascades, or Mixture-of-Agents (MoA) assume that diverse model failures don’t overlap, but the study found that difficult prompts cause entire pools of models to fail together, limiting the effectiveness of routing or voting.
The research encourages developers to focus on matching model quality rather than simply diversity and suggests a $0 pre-deployment test using the Clopper-Pearson bound to estimate the worst-case failure rate, helping teams decide if adding multiple models is worthwhile. For tasks with clear, verifiable answers, relying on the single best model is often more effective than combining many.
This insight impacts enterprise AI deployment strategies, emphasizing the need for precise evaluation of multi-model systems to avoid unnecessary overhead and disappointing performance gains.
Sharing one API key among multiple AI agents means that if one is compromised, all associated workflows are exposed, making attackers' access widespread and difficult to trace. VentureBeat’s June 2026 Pulse Research across 107 enterprises highlights that 69% of organizations allow credential sharing among AI agents, raising serious security concerns. This situation has triggered significant enterprise security investments, with companies like Palo Alto Networks, CrowdStrike, and Cisco acquiring firms focused on AI agent security to address this vulnerable layer.
The research reveals that only 32% of enterprises assign unique, scoped identities to every AI agent, while many still rely on shared or borrowed credentials, weakening accountability and increasing risk. Larger companies face even higher incident rates, yet sandboxing and isolation measures, which can contain damage, are underfunded and underutilized.
Most enterprises depend primarily on security features provided by AI model vendors or cloud hyperscalers, which focus on filtering malicious activity but often lack identity management and isolation capabilities needed for robust protection. Despite high confidence in their current tools, many organizations have experienced incidents or near misses, and investment in agent security does not fully reflect these risks.
Security directors are advised to: (1) thoroughly inventory and eliminate shared or borrowed credentials; (2) prioritize sandboxing for high-risk agents to contain breaches; and (3) align security budgets with real incident exposure. The Q2 Agentic Security report, detailed at the upcoming VB Transform event, underscores that enterprises must urgently close these security gaps or face potential breaches that could compromise critical systems.
Five years after Salesforce acquired Slack for $27.7 billion, the two platforms are finally merging into a unified system. Slack has introduced a new integration connecting Slackbot, its built-in AI agent, to the full Salesforce platform. This includes CRM data, Tableau analytics, Data 360 profiles, and third-party apps—all accessible through simple chat commands in Slack. Powered by Salesforce's Model Context Protocol (MCP) servers and Headless 360 infrastructure, users can now retrieve customer deal histories, generate pipeline charts, update records, and send DocuSign approvals without leaving Slack. This integration enables team-wide visibility and collaboration, addressing the common issue of fragmented information across apps. Slack's CMO, Ryan Gavin, emphasizes "multiplayer AI" as the future of enterprise tools, where AI operates in shared channels rather than isolated chats, enhancing teamwork. While competitors like Microsoft Teams and Google Workspace embed AI into their offerings, Slack bets on openness and integration breadth to differentiate itself. The new capabilities democratize access to Salesforce CRM data across entire organizations, turning more employees into empowered decision-makers. Although some technical and strategic challenges remain, including performance and pricing clarity, this marks a major step in transforming Slack into a comprehensive AI-driven work platform.
