Adept’s Fuyu-Heavy and the Salesforce Agentforce pivot

Adept Fuyu-Heavy and agent technology

Adept holds $415 million in total funding, including a $350 million Series B round led by General Catalyst and Spark Capital. This follows a $65 million Series A round in 2022. The company, founded by authors of the iconic transformers paper, develops autonomous agents that convert high-level objectives into actions. Its Fuyu-Heavy model outscores Gemini Pro on the MMMU benchmark for multimodal reasoning. This model uses DPO and focuses on user interface understanding. Adept agents perceive screens via pixels and execute actions using coordinates and keystrokes. This functionality allows for the automation of repetitive tasks such as form entry or navigating web applications. Fuyu-Heavy remains smaller than models like GPT4-V or Gemini Ultra, yet it matches or exceeds their performance on certain natural language tasks. Adept utilizes natural-language workflow authoring to create step-by-step plans. Fuyu-Heavy builds on the Fuyu-8B model, which Adept released last year. The development of Fuyu-Heavy involved tuning the training procedure to stabilize image model performance. Adept’s platform relies on computer vision and language to process multimodal data from the start.

Salesforce pivots toward Agentforce

Salesforce pivots toward an AI operating platform strategy through its Agentforce platform. Marc Benioff emphasizes a "beginner’s mind" approach as the company transitions from predictive AI to autonomous agents. Salesforce acquired Contentful to provide the structured knowledge and metadata that agents need to retrieve and reason over information. The company also purchased m3ter to manage usage metering and monetization for these workflows. While traditional SaaS platforms charge for users, the company targets a model that charges for specific agent actions. An organization that maintains a base of only 500 Salesforce users might deploy 50,000 individual AI agents that perform complex tasks continuously throughout the entire workday without needing any human intervention. Salesforce also integrates Data Cloud to provide the customer context necessary for these agents. To support this, the company acquired the AI agent startup Airkit in 2023 and invests in companies like Anthropic and Cohere. Agentforce includes a Trust Layer for secure data retrieval and toxicity detection. Does Salesforce have enough data to prevent customers from switching to AI-native companies?

Automation competition and enterprise results

Automation providers now compete on how agents interact with software environments. Microsoft’s Copilot Studio provides computer use capabilities that allow agents to operate applications by looking at the screen. This method works for systems without APIs where traditional screen-scraping bots struggle. Microsoft offers these services at a lower cost than UiPath unattended robots, which cost 3 to 4 times more. You should check if your current bots actually require a robot or just a scheduled agent. Many legacy bots in a UiPath estate can collapse into fewer flows or agents. Industry data from SpendHound shows UiPath pricing for smaller customers increased 73.95% year over year.

Case studies prove the value of these autonomous workflows. A Fortune 500 SaaS company with 12,000 employees and 80,000 customers achieved 75% ticket deflection and a $2.1 million annual saving by using a multi-agent AI system. Previously, 68% of its 140,000 monthly tickets were repetitive, but its First-Contact Resolution rate was only 38%. After implementation, the company increased First-Contact Resolution to 80% and reduced handle time from 4.2 minutes to 2.1 minutes. In manufacturing, Zochem reduced data entry errors by over 90% within 12 weeks using AI. Centracom lowered support costs by 80% after implementing AI voice agents. Openreach deployed NiCE Cognigy AI agents in April 2026 to manage 15 million customer journeys.