Agent Experience Design (AX): Why AI Agents Need Their Own Experience Design

A leadership inflection point from copilots to autonomous work

Enterprise AI has entered its “everywhere, but uneven” phase. In McKinsey & Company’s 2025 global survey, 88% of respondents report “regular AI use” in at least one business function, yet the majority are still in experimenting or piloting phases, with only about one‑third saying their organisations have begun scaling AI programmes.

At the same time, leaders are moving beyond copilots into AI agents—systems that can plan and execute multi‑step work. McKinsey defines AI agents as systems based on foundation models that can act in the real world by planning and executing steps in a workflow, and notes 23% of respondents report scaling an agentic AI system somewhere in their enterprise, with an additional 39% experimenting with AI agents.

Independent research points in the same direction. In a 2025 Boston Consulting Group brief, BCG reports that (in a study conducted with MIT Sloan Management Review) 35% of organizations say they are already using agentic AI and another 44% plan to do so soon.

This shift creates a new executive challenge: deploying agents is not merely “adding AI.” It is designing an enterprise environment where AI can reliably do work—with clear boundaries, trusted context, and observable outcomes. That discipline is Agent Experience (AX).

1. Why AI adoption is not translating into enterprise value

If this feels like a paradox—high adoption but low measurable impact—you are not imagining it. McKinsey reports that while respondents see use‑case‑level benefits and innovation signals, only 39% report enterprise adoption en say they have begun experimenting with AI agents. But use of agents is not yet widespread.

One reason is that much of what has scaled so far is horizontal AI (copilots, chat interfaces, broad assistance) that produces diffuse productivity gains that are harder to measure and harder to compound into end‑to‑end operational outcomes. McKinsey explicitly frames this as an imbalance between fast‑scaling copilots and more transformative, function‑specific “vertical” use cases that remain stuck in pilot mode, and argues that agents are one path to breaking out of that “gen AI paradox.”

OpenAI’s enterprise guidance reflects the same practical reality: real agent value comes when systems can both retrieve the necessary context and take actions through tools—while being designed with safe boundaries and escalation paths so they can operate reliably in production conditions.

So the executive takeaway is not “we need a better model.” It is: we need an environment fit for agentic work—the data, permissions, workflow pathways, and governance that allow agents to execute rather than merely advise.

2. What Agent Experience means in practice

AX is often misunderstood because it sounds like a rebrand of UX. It is not. AX starts from a different premise: an AI agent is a user with different needs and failure modes.

Salesforce’s Chief Experience Officer defines agent experience (AX) design as the development and optimization of digital environments so that AI agents can operate efficiently and effectively—and orchestrate human‑centred outcomes. Salesforce also emphasises that AX includes both designing for agents and designing of agents, because both are required to ensure agents prioritise people’s goals.

McKinsey similarly defines an AI agent as a software component that has the agency to act on behalf of a user or a system, and describes structured ways organisations may deploy agents—from copilots to workflow automation platforms to AI‑native operating models. This matters because as you climb that ladder, the agent’s “experience” becomes less about interfaces and more about operating conditions: clarity of tasks, access to trusted data, and permitted actions.

A practical way for executives to hold the distinction:

This framing aligns with Salesforce’s definition of AX and with OpenAI and McKinsey’s emphasis on tools, actions, controls, and governance for production deployments

3. The AX blueprint for B2B leaders

To make AX actionable without turning it into an engineering conversation, treat it as five leadership questions. Each one maps directly to why pilots fail and why scalable programmes succeed.

Trusted context: can agents reliably access the “truth” of the business?
Agents are only as reliable as the context they can retrieve. McKinsey’s 2025 reports repeatedly stress that scaling is hard work, and that organizations must redesign workflows rather than bolt agents onto legacy processes. That redesign starts with trustworthy, current data flows.

This is why enterprise platforms increasingly emphasise “zero‑copy” or live connectivity patterns. For example, ServiceNow positions Workflow Data Fabric and “Zero Copy Connectors” as a way for workflows and AI agents to run on real‑time contextual data rather than proliferating duplicates; ServiceNow’s materials explicitly describe accessing external data without copying it into the platform.

Autonomy boundaries: where can the agent act, and where must it ask?
OpenAI’s agent deployment guidance highlights the need to design workflows with appropriate controls, including human involvement for higher‑risk steps, and to architect tool use and orchestration so actions are reliable and reviewable.

This aligns with McKinsey’s broader argument that agents introduce new risks (from uncontrolled autonomy to lack of observability) that cannot be solved by “plugging agents into existing workflows,” but require reimagining task flows and governance with agents at the core.

Action paths: can the agent complete the job, not just recommend?
In practical terms, an agent that only drafts recommendations is still a copilot. OpenAI’s guide explicitly separates “data” capabilities (retrieving context) from “action” capabilities (interacting with systems to take steps such as updating records or sending messages), and treats both as foundational to real agent workflows.

Agent definition also centres on agency “to act on behalf” of a user or system, reinforcing that execution—not just advice—is core to agent value.

Observability and ownership: do you know what agents are doing and who is responsible?
Governance becomes more important as autonomy increases. ServiceNow’s AI Control Tower materials emphasise maintaining an AI asset inventory (connected to enterprise services and assets) to gauge risk and apply governance at scale—an example of the “control plane” idea executives need for agent sprawl.

This governance narrative is consistent with the National Institute of Standards and Technology AI Risk Management Framework (AI RMF), which is intended to help organisations manage AI risks and incorporate trustworthiness considerations across the AI lifecycle.

Workflow redesign: are you building agent‑centric processes or automating yesterday?
Both McKinsey and BCG stress that the real upside comes when organisations rethink workflows and value creation—moving beyond initial productivity gains toward differentiation. BCG explicitly notes that productivity gains are often the initial benefit, but argues the “real prize” in the agentic age is differentiation and sustaining advantage.

4. How to implement AX without betting the company

AX is best implemented as an operating discipline, not a one‑off project. The research strongly supports starting small, measuring outcomes, and scaling with governance.

McKinsey’s survey shows agent deployments are most often scaled in only one or two functions, and that scaling at the individual function level is still uncommon. That is a signal to lead with focused “lighthouse” workflows, not enterprise‑wide mandates.

A practical staged approach (consistent with OpenAI’s deployment guidance and McKinsey’s “reimagine workflows from the ground up” message) looks like this:

Select a bounded workflow with clean success criteria (cycle time, resolution time, coverage, rework rate) and a manageable risk profile.
Define autonomy boundaries explicitly: what the agent can do unaided, what requires review, and what must always escalate.
Instrument outcomes and failure modes, then expand autonomy only when reliability is proven in your environment.

This is how AX protects credibility: it replaces “AI theatre” with controlled, measurable operating gains—while preserving trust and compliance expectations.

5. Evidence from early adopters

Public examples—when used carefully—help leaders see what “AX in the real world” looks like. The key is to use verifiable, attributable claims and to label them appropriately.

ServiceNow publishes “Now on Now” internal outcomes for HR service experience, reporting 20× faster resolution to HR queries, 410,000 hours saved annually through AI‑powered search and virtual agent capabilities, $17.7M in annual cost avoidance from self‑service HR services, and 81% employee digital experience satisfaction (eSAT). These are vendor‑reported internal metrics, but they are concrete and tied to specific workflows and measurement categories executives understand.

Salesforce describes how Accenture expects agents to accelerate bids through the pipeline, aiming for 100% bid coverage, up from 25% as of December 2024. This is reported as part of Salesforce’s story about Agentforce and Accenture, so it should be read as a stated goal and programme narrative rather than an independently audited outcome—but it is still a verifiable, attributable claim from a named source.

Walmart provides a strong pattern: the company describes consolidating multiple agents into four “super agents”—distinct entry points for customers, associates, partners/suppliers, and developers—because many separate agents can become “overwhelming and confusing.” That is a pure AX insight: experience design becomes the discipline of reducing cognitive and operational load as agents proliferate.

On the productivity side, CIO Dive reports that Walmart said in earnings call that AI coding assistance and completion tools saved developers about 4 million hours in a year, supporting the idea that well‑embedded AI (with clear workflow integration) can generate measurable operational gains.

Across these examples, the pattern is consistent with McKinsey and BCG’s research: early wins come when organisations pick specific workflows, redesign the environment, define boundaries, and govern scale—rather than treating agents as a feature to “deploy.”

Conclusion: AX is the operating system for the agentic enterprise

Leaders should treat AI agents as a new kind of workforce capability: powerful, fast, and increasingly accessible—but only as reliable as the environment you design around them. McKinsey’s 2025 data shows adoption is high, scaling is hard, and agentic deployment remains early across most functions.

AX is the discipline that closes that gap. It operationalises what the best research now repeats: agents unlock value when organizations redesign workflows (not bolt on tools), provide trusted context, define explicit autonomy boundaries, and implement governance that makes speed sustainable.

For entrepreneurs and executives, the message is straightforward: competitive advantage will not go to the company with the flashiest demo. It will go to the company that builds agent‑ready products and operations—so agents can reliably execute outcomes with traceable, governable behaviour.