Adept’s action models and Salesforce’s agentic maturity roadmap

Action models replace passive information retrieval

Adept AI raised $350 million in its Series B, which gives the San Francisco-based startup a post-money valuation of at least $1 billion. Its ACT-1 model turns text commands into actions rather than only responding to them, whereas traditional Large Language Models (LLMs) remain passive. LLMs act as expert librarians that summarize content or answer questions, but they cannot independently take action inside business systems. These models often require external code or wrappers to simulate decision-making, which can be brittle or difficult to scale. Large Action Models (LAMs) function as digital teammates that execute work across applications. Uniphore’s ActIO, an enterprise-grade LAM, uses a planner agent to break user goals into subtasks and a UGround engine to interpret GUIs through computer vision. This technology allows the system to work within software environments by identifying screen elements based on appearance. Using SAIL technology, the system adapts to new interfaces with zero manual configuration. Customers using LAMs for invoice processing reported a 60% reduction in handling time and a 90% improvement in accuracy. The transition from basic chatbots to complex multi-agent orchestration requires businesses to identify specific use cases where reasoning capabilities improve outcomes through established governance frameworks and autonomous task performance across various software environments and organizational hierarchies.

Salesforce targets agentic maturity

Salesforce targets higher levels of AI maturity through its Agentforce platform, which has 5,000 customers, including 3,000 paying users. The company plans to acquire Convergence.ai to improve its ability to handle complex tasks and adapt to changing conditions. Convergence employs engineers from Google DeepMind and Meta to build tools that perform complex tasks in real time. Salesforce defines the progression of AI maturity through four stages: chatbots and co-pilots, information retrieval agents, simple orchestration and single domain, complex orchestration and multiple domain, and multi-agent orchestration. To move between these stages, companies like Alpine Intel and Wiley use the Salesforce framework to identify where chatbots fail and where reasoning can improve results. This progression aims for full interoperability between agents and data environments. The company aims to align data and workflows across its clouds using Salesforce Data Cloud customer data. You should observe how these agents move from being tools that follow orders to agents that achieve goals. Salesforce provides several ways to pay for these services.

The measurement gap and financial reality

The industry faces a disconnect between massive investment and realized value. AI startups collectively raised $305.6 billion according to Forbes 2026 lists, yet MIT NANDA research found that 95% of generative AI pilots deliver zero measurable P&L impact. Only 5% to 8% of these pilots achieve measurable at-scale ROI. This gap persists because most enterprises allocate over 50% of AI budgets to sales and marketing, while the highest ROI sits in operations and back-office applications that receive less than 10% of the total budget. 21.7% of IT leaders cited direct financial impact as their top success metric in 2026, a figure that doubled from the prior year. To measure this impact, enterprises must look at different P&L lines. Direct cost reduction involves absorbing work that would otherwise require headcount or vendor spend. Revenue lift comes from better recommendations or faster sales cycles, though this requires a counterfactual like a holdout group to defend the number. Risk avoidance covers losses from fraud, errors, or regulatory action. Productivity gains, which used to dominate the conversation, only provide value if they convert into headcount neutrality or reduced overtime.

Why do companies continue to fund unmeasurable pilots?