The Agentic Enterprise
The term Agentic Enterprise refers to organizations using AI-driven agents to autonomously perform tasks, including decision-making, and, being context aware, adapting to complex workflows. These agents go beyond handling repetitive tasks—they can also be capable of managing heterogeneous processes with minimal employee involvement. The goal of an Agentic Enterprise is to enhance efficiency, streamline operations, and enable employees to focus on higher-value, creative tasks while leaving routine work to agents. The number one challenge companies face in deploying agents is data reliability and accuracy. But addressing that challenge alone will not ensure success.
Much of the work people do takes place outside the primary system of record. For example, in updating a customer order, an account manager may need to refer to emails, texts, phone calls, not to mention PDFs, spreadsheets, and various documents–all of which exist outside their CRM tool. There are any number of applications and services required to support today’s work activities, so in order to create an effective agent, one capable of autonomy, you have to go beyond the “happy path” of your sales platform and really understand what people do to get the work done. This is especially true for business processes where multiple employees are working asynchronously to complete the task. This will require organizations, and their Agentic AI vendors, to expand their training data.
LLM’s capture the variances and patterns within a data set, enabling prediction and the successful generation of outputs. Currently organization’s LLMs are limited to the data within their enterprise and SaaS applications, but truly effective agents will require additional training data; data that reflects the sequence of activities humans perform with those systems and with each other, to complete their work.
The activity stream of each employee, like a document, is unique. Which app do they open first? Do they click the link or copy and paste? What was the sequence of actions they took between multiple applications; messaging apps, search tools, etc. When combined with hundreds or thousands of instances performed over time, it is possible to create predictive interaction models that represent optimal workflows and interactions to enable more compelling and more successful agents. And it will be Design’s responsibility to ensure those models are comprehensive and inclusive as well as respectful of users’ privacy.
Decades ago this work was done by anthropologists like Julian Orr, carefully observing people at work, and building models of their activities and interactions to inform designs for office equipment. Today this information can be gathered by local digital sensors, log files, and clickstream analytics. In fact, I co-founded two start-ups that did exactly this, one focused on AI driven automation and the workforce automations. Modeling the activity stream is not the hard part, it was overcoming the collective skepticism both from employees (no one wants “big brother” looking over their shoulder), and designers over the invasiveness and ethical concerns. At both start-ups we invested heavily in designing in trust by providing transparency and end-user control over what data we collected and how it was used. We also provided control over the scope and level of autonomy for the resulting automations. Such assurances were necessary to gain admittance, adoption, and continued engagement. However once the users saw the results, their fears were quickly set aside, as they recognized the gains to be made by leveraging their newly minted agents.
Once users saw the benefits, their skepticism faded. Four key benefits today’s agents’ offer include:
Workflow Automation: Agents are great at handling repetitive tasks, freeing employees from the mundane so they can focus more time on strategic, creative work. At the same time with the development of new foundation models, tasks once thought too complex or heterogeneous can now be automated.
Adaptability: While the business as a whole operates consistently, these adaptive agents can make your employees more effective by responding to each user's unique working style; providing an experience that is both tailored and intuitive for each individual.
Autonomous Learning: Agent's ability to refine and/or expand its capabilities. Based on the varied datasets it encounters and user feedback, these agents update the underlying data model without needing external intervention.
Faster Support & Resolution: Agents enable quicker decision-making, whether in customer support, sales, or other processes, while automatically updating systems and tracking issues. Again adapting and applying those refinements to everyone’s benefit.
As agents play a larger role, users may face challenges:
Trust the autonomy
As enterprises shift more decision-making to AI agents, users may feel disconnected from the process and start questioning the value of their own contributions. To address this, it's important to provide user-friendly, contextual explanations of how the agent reaches its decisions. This could include decision rationales, contributing factors, or data sources used by the agent. Additionally, employees should have the opportunity to give feedback on the outcomes. The goal is to refine both the agent’s rules and the employees' role in the process to ensure alignment and maintain trust while also fostering accountability and a sense of control.
Data privacy
Agentic enterprises will depend heavily on behavioral data, capturing the tasks employees routinely perform. To ensure AI agents make informed decisions, organizations must gather data on both standard workflows and edge cases. However, employees may be wary, especially if they fear this data could be used punitively. To build trust, transparency is crucial. Employees should control what data is collected, with the option to review and edit their activities. And organizations must explain how workflows are anonymized, standardized, and included in the AI models
Loss of satisfaction
Some users find satisfaction in handling certain tasks themselves, especially in creative or problem-solving areas. For instance, someone managing personal finances may prefer the hands-on approach to budgeting, while AI automation in these tasks might feel impersonal or even frustrating. Taking a human-centered approach engaging both employees and business process experts to identify the most impactful tasks for automation will help ensure you're solving real problems and delivering genuine business value.
Reduced transparency
When AI agents take actions without user input, it can create a sense of lost transparency. This disconnect may lead users to feel less ownership over outcomes. For example, if an agent automatically schedules meetings or manages emails, users might lose track of important details or decisions, affecting their sense of control. It's important to remember that even though agents handle tasks, users still want to stay informed and be accountable for what’s happening.
Control over agent behavior
To strengthen users' sense of control, they should have control over how AI agents operate—such as setting limits on the agents' autonomy or choosing which tasks they can handle. Consider allowing users to “dial up” or “dial down” an agent’s autonomy according to personal preferences (like setting budgets) while allowing agents free rein in low-impact areas (like adjusting calendar availability).
To address these issues, users may want flexibility in controlling how much autonomy an AI agent has over specific tasks. However, giving agents too much independence without oversight could lead to discomfort or distrust. While users may understand the efficiency benefits, they might feel disconnected or "disempowered." Consider the following approaches:
Customizable Autonomy Settings: Design will need to provide adjustable settings that allow users to decide which actions agents can perform autonomously. For example, in customer service applications, users might allow an AI to handle routine queries but will want to be notified when more complex issues arise. This will also require design solutions for both differentiating between routine issues and “complex issues”, as well as how users train agents to autonomously address these complex situations in the future.
Layered Permissions: Design will need to define a layered permission system, allowing users and/or company policies to set boundaries for agent actions. A banking user, for instance, might permit an AI agent to perform low-risk transactions but require manual approval for larger transfers, while the bank may have additional policies in place as part of compliance and governance.
Notification and Consent: Design will need to provide scalable solutions to ensure the user is aware whenever an AI makes a significant decision on their behalf, and can intervene if necessary. As agents perform in novel circumstances, design needs to understand if the user requires real-time explanations or when and how agents ask for help from users to maintain oversight. This “in-the-loop” approach can help foster trust by keeping users informed and in control.
Fail-safes and Reversibility: Design will need to provide the means of overriding or reversing AI-driven actions when necessary. For instance, if an AI agent makes an error, the user should have clear options to correct or “undo” the AI’s action. Even making corrective adjustments to the underlying model to ensure the error doesn’t occur in the future.
On-Demand Agent Insight: Design also should provide a clear “explain why” feature, so users can ask the agent why it made a specific choice. Not only will this again help build trust by explaining the agent's reasoning it can make the agent's actions more predictable. Showing the steps behind their decision-making process (such as “I scheduled this meeting based on your availability and preferences”) will provide another means for users to query and correct their agents' rules.
Finally, there are two additional factors that have to be addressed to ensure your success in transitioning to an agentic enterprise.
Bias and Fairness: If agentic systems are being designed to make decisions based on algorithms it is critical that your system does not inadvertently harbor biases, otherwise your customers and your business can suffer from unfair treatment, especially in high-stakes areas like finance, insurance, and hiring.
Accessibility and Inclusivity: As enterprises transition to the new agentic paradigm, there’s a greater need to ensure these agents are accessible to employees and customers with disabilities, and that your agents can serve a diverse user base equitably.
Conclusion
The Agentic Enterprise promises to deliver greater customer value through customized experiences, faster and more flexible solutions, and greater agility in a rapidly changing market. However, maintaining trust, control, transparency, and fairness will be critical in balancing the benefits of autonomy with user empowerment.
