The AI industry’s push toward generalist AI agents, while impressive, may be solving the wrong problem for businesses. While companies like Microsoft, Anthropic, and others focus on creating AI agents that can handle any task – from browsing the web to writing code to analyzing documents – the real business value lies in building specialized digital workers that excel at specific vertical workflows. This specialized approach, focused on pre-built skills and industry-specific processes, offers a more practical, cost-effective, and reliable path to AI implementation in the enterprise.
The current landscape of AI agents faces several critical challenges: high operational costs due to extensive LLM usage, latency issues that slow down business processes, and reliability concerns that make businesses hesitant to deploy these solutions at scale. These challenges stem from trying to make AI agents figure out how to complete tasks from scratch every time. Consider a typical Order to Cash process – a generalist AI agent would need to understand the process, determine the steps, write code for integrations, and execute each action, consuming significant computational resources and time while introducing multiple points of potential failure.
A more efficient approach is to shift the heavy lifting from the generative capabilities of large language models to pre-built, programmatic functions that handle specific business workflows. In this model, AI agents serve as intelligent orchestrators, deciding which pre-built skills to utilize rather than attempting to generate solutions from scratch. These skills – essentially pre-coded, tested, and optimized functions – handle the technical integrations between various systems, allowing the AI agent to focus on decision-making and process flow rather than technical implementation. This dramatically reduces costs while improving reliability and speed of execution.
The human element remains crucial but shifts to a more strategic role. Rather than requiring constant oversight of AI actions, humans interact with these digital workers through familiar messaging platforms like Microsoft Teams or Slack. This creates a natural, conversational interface where digital workers can proactively seek guidance when encountering exceptions or requiring decisions beyond their authorized parameters. This approach mirrors human workplace interactions, where employees consult their managers when facing unusual situations or needing additional authorization. The key difference is that these digital workers operate within well-defined parameters and with pre-built skills that have been thoroughly tested and validated.
The development and refinement of these skills follow a consultative approach, where business owners and subject matter experts help perfect the workflows before they’re deployed. This front-loaded expertise ensures that the skills accurately reflect industry best practices and regulatory requirements. Once developed, these skills can be deployed across multiple client instances, with customization options available through a managed package interface. This approach allows for standardization where beneficial while maintaining the flexibility to adapt to specific business needs. As new requirements emerge or processes evolve, the underlying skills can be enhanced and automatically updated across all instances, ensuring continuous improvement without requiring significant client-side intervention.
The evolution of these digital workers doesn’t rely on traditional machine learning approaches where each instance needs to learn from scratch. Instead, capabilities grow through the addition of new skills to the available toolkit. This means that when a new automation capability is developed and tested, it can be immediately deployed to all digital workers that might benefit from it. This approach to “learning” is more controlled and reliable than hoping an AI agent might figure out a better way to handle a task through trial and error. Moreover, by focusing on specific vertical workflows, the scope of required skills remains manageable and focused, allowing for deeper optimization of critical business processes rather than shallow coverage of many possible tasks. As we stand at this crossroads of AI implementation in the enterprise, perhaps we should ask ourselves: Are we too focused on building AI that can do everything, when what businesses really need is AI that can do the right things exceptionally well?
