User-Centricity and Empathy
The most technically sophisticated AI agent is worth nothing if the people it is designed to serve do not use it, do not trust it, or cannot interact with it effectively. User-centricity is the principle that places actual user needs — not technical capabilities — at the center of every design decision. It is both the starting point of agent design and the standard against which every design choice must be measured.
This principle matters more for agentic AI than for any prior generation of enterprise software. Agents take autonomous actions on users’ behalf. An agent that misunderstands a user’s intent, communicates in jargon the user doesn’t understand, or requires workflow changes the user resents will not just be underutilized — it may actively harm the user by taking wrong actions confidently.
Design for the Actual User, Not the Assumed User
Effective user-centric design begins with research, not assumptions. Different user populations in the same organization have fundamentally different needs:
Role-appropriate behavior — An agent assisting a compliance officer needs to surface regulatory references and flag risk language. The same agent assisting a relationship manager needs to surface client context and recommended actions. Both may use the same underlying system, but the output format, depth, and default focus should adapt to the user’s role and responsibilities.
Expertise-calibrated communication — An agent interacting with a data scientist can use technical terminology and reference statistical methods. The same agent interacting with a business executive needs to convert those outputs into implications and recommendations in plain language. Agents should detect or be informed of user expertise level and adjust accordingly.
Contextual awareness — The agent should understand what the user is trying to accomplish in their broader workflow, not just respond to the immediate request. A user asking “what are the Q3 numbers?” in the context of preparing a board presentation needs a different response format than the same question asked during a routine financial review.
Integration, Not Interruption
User-centric agents live where users already work. The principle of integration over interruption means:
- Agents surface in the tools users are already in — Microsoft Teams, Slack, Salesforce, ServiceNow, the EHR — not in separate applications that require context-switching
- Agent interactions fit the natural flow of the user’s work, appearing at relevant moments rather than requiring users to seek them out
- Outputs match the format the user needs — a structured table for a spreadsheet user, a brief summary for an executive, a draft document for a writer
Research consistently shows that the highest adoption rates for enterprise AI tools come from embedding capabilities in existing workflows, not from deploying standalone AI portals. JPMorgan Chase’s engineering productivity tools and clinical decision support agents that operate within the EHR are examples of this principle in practice.
Eliminate Friction, Address Pain Points
The clearest user-centric design criterion is whether the agent eliminates a specific pain point that users experience today. Agents designed around demonstrable friction — not hypothetical use cases — earn user trust quickly. Examples of friction-eliminating design:
- An agent that compiles a report from four systems that a finance analyst currently spends three hours assembling manually
- An agent that drafts the first version of a case summary that a case worker currently writes from scratch after every intake interview
- An agent that answers the “where do I find…?” questions that currently require a colleague search or a support ticket
The test of user-centricity is not whether users could use the agent — it is whether they prefer to.
Feedback Loops as Design Requirements
User feedback is not just a nice-to-have; it is the mechanism through which the agent learns what user-centricity means in practice. Every agent interface should include:
- Easy mechanisms to flag incorrect or unhelpful outputs (thumb down, “this is wrong”)
- Mechanisms to indicate that an output was useful (which improve retrieval and generation over time)
- Periodic structured feedback for high-value use cases (brief satisfaction surveys, usage analytics)
These feedback signals should feed directly into improvement cycles — logged, reviewed, and acted upon by the agent operations team.
Make It Your Own
Key questions to ask in the context of your organization:
- What research have you done with actual users — not assumed users — to understand their primary pain points, preferred communication styles, and existing workflow constraints?
- Which specific friction points in each user population’s daily work would deliver the most immediate, tangible value if eliminated by an agent?
- Which existing applications and workflow environments should agents be embedded in, versus which would benefit from a new dedicated interface?
- How will you calibrate agent behavior to different expertise levels and roles across your user population?
- What feedback mechanisms will you build into every agent interface, and how will feedback data flow into your improvement cycle?
- How will you measure user adoption, not just system utilization — distinguishing users who genuinely prefer the agent from those who use it reluctantly or not at all?