Revival Group
Back to Blog

General

Agentic Experience: Why the Next Software Advantage Won’t Come From a Better Interface

UX taught us to make tools easier to click. AX is about designing software you can delegate to and trust. Here is why that distinction is now a business problem, not a design one.

Reshad Noorzay•6/27/2026•10 min read
Agentic Experience: Why the Next Software Advantage Won’t Come From a Better Interface

UX taught us to make tools easier to click. AX is about designing software you can delegate to and trust. Here is why that distinction is now a business problem, not a design one.

TL;DR

  • Agentic Experience (AX) is to AI agents what User Experience (UX) was to graphical interfaces: a new design paradigm. The shift is from designing screens that humans operate to designing the behavior, trust, and judgment of software that reasons and acts on a user’s behalf. Late Checkout’s framing captures it well: the center of design gravity has moved “from tool to relationship.”
  • This matters to leaders, not just designers, because adoption, retention, and differentiation in the AI age are decided by the quality of that relationship. MIT’s NANDA initiative found 95% of enterprise generative AI pilots delivered no measurable P&L impact despite an estimated $30–40 billion in spend, and Gartner expects over 40% of agentic AI projects to be canceled by end of 2027. The common failure mode is not the model. It is the experience and integration around it.
  • The practical agenda is concrete: design for delegation and oversight, treat trust and memory as first-class materials, build for non-deterministic systems that are sometimes wrong, and measure outcomes rather than outputs. A great agentic experience is precisely what turns an AI investment into a system that actually works, gets adopted, and compounds.

Key Findings

  • The interaction model has fundamentally changed. Jakob Nielsen of Nielsen Norman Group argues generative AI is the third UI paradigm in computing history and “the first new paradigm in more than 60 years”: a move from command-based interaction to “intent-based outcome specification,” which he says “completely reverses the locus of control” that defined UX for four decades.
  • The unit of design changed with it. Classic UX designed screens, pages, components, and flows. AX designs conversations, autonomy boundaries, permissions, memory, and trust. Salesforce’s design leadership describes designers shifting “from interface architects to experience orchestrators.”
  • Trust is now the scarce resource and the growth lever. Qualtrics found AI-powered customer service failing at nearly four times the rate of AI use in general, and data-privacy misuse is now the top customer concern with AI. Consumer comfort with autonomous agents is volatile. The products that win earn the right to act on the user’s behalf.
  • The business stakes are explicit. McKinsey’s 2025 research shows nearly eight in ten companies use generative AI yet roughly the same share report no significant bottom-line impact, a pattern it calls “the gen AI paradox.” High performers are distinguished by workflow redesign and defined human-in-the-loop validation, not by access to better models.
  • The contrarian implication for leaders: most businesses do not need an “AI strategy.” They need AI systems that actually work, and “works” is now an experience property. A generic chatbot bolted onto an existing product is a UX artifact. A system someone trusts to act is an AX one.

Details

From tool to relationship

For about forty years, software design optimized one thing: how efficiently a human could operate a tool. We reduced clicks, clarified labels, smoothed flows, and removed friction. That discipline, user experience, won markets. It assumed a deterministic machine that waited for human input and did exactly, and only, what it was told.

Large language models broke that assumption. Software can now reason in plain language, keep track of context, anticipate needs, and take initiative. As the team behind the book “AX: The Rise of Agentic Experience,” authored by Theo Tabah with Ed Landon at Late Checkout / LCA (the design firm co-founded by Greg Isenberg and Theo Tabah), puts it, the day models learned to reason “the centre of design gravity slid from tool to relationship.” That book is one of the clearest articulations of the concept, and it is not alone. Netlify CEO Mathias Biilmann coined a parallel term, “Agent Experience,” in a January 2025 post, defining it as “the holistic experience AI agents will have as the user of a product or platform.” Salesforce, Microsoft, Google, IBM, Apple, and independent practitioners like Emily Campbell have all been building the vocabulary.

The naming is still contested. Late Checkout uses “Agentic Experience” to describe designing the relationship between a human and an agent. Biilmann’s “Agent Experience” focuses on designing for agents as a new class of user. Both point at the same tectonic shift: software is no longer only a tool we operate. It is increasingly a collaborator we delegate to. This article uses AX in the relationship sense, because that is where the business stakes for most leaders sit today.

Six concrete differences leaders should internalize

The gap between UX and AX is not cosmetic. It runs through six dimensions.

1. Interaction model: manual input versus goal delegation. UX is built on direct manipulation. The user performs each step. AX is built on what Nielsen calls intent-based outcome specification. The user says “book me a flight to Chicago for the client meeting” and the system reasons about the steps. Nielsen’s own caution matters here: he expects a hybrid interface, not the death of the GUI, combining natural-language intent with familiar controls that let users fine-tune and verify.

2. Control: doing versus overseeing. In UX, the user is in control of every action. In AX, the user delegates and then supervises. The World Economic Forum’s design guidance frames this as a move from “human-in-the-loop” to “human-on-the-loop,” where people govern strategy while agents handle execution. That only works if the human can see what the agent did, challenge it, and disengage easily.

3. The design unit: screens versus conversations, memory, and autonomy boundaries. A UX designer ships wireframes and components. An AX designer ships conversation design, permission models, memory rules, escalation paths, and explicit limits on what the agent may decide or do. Salesforce describes the day-to-day work changing “from static wireframes to dynamic orchestration rules.”

4. Predictability: deterministic versus probabilistic. A button does the same thing every time. A generative system can produce different outputs from the same input, and it is sometimes confidently wrong. Apple’s Human Interface Guidelines for generative AI state the design problem plainly: unlike classic programming, small changes to inputs, or even identical inputs given multiple times, often produce very different outcomes. AX therefore designs for error: transparency, recoverability, confidence display, and graceful failure become core features rather than edge cases.

5. Success metric: outputs versus outcomes. UX optimized task completion and ease of use. AX optimizes whether the goal was achieved and whether trust was earned. Nielsen Norman Group calls the emerging discipline “outcome-oriented design,” where designers define goals and constraints rather than crafting every interface element. Outputs are what the software produces. Outcomes are what the business actually wanted: the quote sent, the issue resolved, the hours saved.

6. Trust and memory as first-class materials. In UX, trust was a byproduct of reliability and clarity. In AX, trust is the material you design with directly. Google’s People + AI Guidebook reframes the goal from “building trust” to “calibrating trust,” helping users rely on the system when it is right and stay skeptical when it is not. Memory is the other new material. An agent that forgets context feels like a stranger every session. The WEF notes that “trust erodes when AI behaves erratically,” and that authenticity and memory must be built into agent design.

The patterns are already being written down

This is not abstract. A practical pattern library for AX already exists across credible institutions, which means leaders can hold their teams to a standard rather than a vibe.

Microsoft’s HAX Toolkit codifies 18 guidelines for human-AI interaction, organized around four moments: initially, during interaction, when the system is wrong, and over time. They synthesize more than two decades of research and were validated in a study where 49 design practitioners tested them against 20 popular AI products. Google’s People + AI Guidebook adds six chapters including Explainability and Trust, Feedback and Control, and Errors and Graceful Failure. IBM’s Carbon for AI makes transparency a component: an “AI label” marks AI-generated content and provides an in-context path to explainability, on the principle that “imperceptible AI is not ethical AI.” Apple’s guidelines tell designers to ask for confirmation before irreversible actions and to make results easy to refine or revert. Emily Campbell’s Shape of AI catalogs emerging interaction patterns. The recurring AX patterns across all of them are clear: show your work, set expectations about what the system can and cannot do, design human checkpoints, make correction easy, and fail gracefully.

Why this is a business issue, not a design-team issue

Here is the part leaders cannot delegate. The economics of the AI age are being decided by experience quality, and the failure data is stark.

MIT’s 2025 “The GenAI Divide: State of AI in Business” report, from its NANDA initiative, found that 95% of enterprise generative AI pilots delivered no measurable impact on profit and loss, with only about 5% achieving rapid revenue acceleration, despite an estimated $30–40 billion in enterprise spending. The finding is based on 150 leader interviews, 350 employee surveys, and analysis of 300 public deployments. The authors attribute the failure not to model quality but to a “learning gap,” the inability to integrate AI into real workflows. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls; Senior Director Analyst Anushree Verma warns that most current projects are “early stage experiments or proof of concepts that are mostly driven by hype,” and Gartner estimates only about 130 of the thousands of agentic vendors are real. McKinsey’s research reports the paradox bluntly: nearly eight in ten companies report using generative AI, and just as many report no significant bottom-line impact, what McKinsey labels the “gen AI paradox.” Only 39% report enterprise-level EBIT impact, most of it under 5%.

Notice the common thread. These are not model failures. They are experience and integration failures. The pilots that stall are the ones bolted onto existing processes without redesigning the relationship between human and system. McKinsey’s high performers are distinguished by two things that are pure AX: they fundamentally redesign workflows, and they build defined human-in-the-loop validation processes.

The customer-facing stakes are just as direct. Qualtrics’ 2026 Consumer Experience Trends Report, based on more than 20,000 consumers across 14 countries, found AI-powered customer service failing at nearly four times the rate of AI use in general, with nearly one in five users seeing no benefit. It also found that misuse of personal data is now consumers’ single biggest concern about AI, cited by 53% and up 8 points over the prior year. Practitioner and survey evidence on chatbots is consistent: users do not resent AI itself, they resent wasted effort, dead-end loops, and being blocked from a human. When the experience signals that the company values its support budget over the customer’s time, trust collapses, and trust is the most expensive thing to rebuild. Research on AI customer support describes “algorithmic aversion,” where one bad AI experience makes users distrust AI broadly and approach the next interaction expecting failure.

This is where Revival Group’s thesis becomes concrete. Most businesses do not need an “AI strategy” deck. They need AI systems that actually work, and in the agentic era “works” is an experience property. A system works when a service manager trusts it to draft and send the quote, when a customer trusts it to resolve the issue, when an operations lead trusts it to triage the inbox without supervision. That trust is engineered through AX. The differentiator is not a proprietary model, which everyone can rent. It is the operating model, the delivery, the speed, and above all the adoption that good agentic design produces.

Augmentation, not replacement, is an AX design choice

The most durable framing for leaders is that AI is a nail gun, not a pink slip. A nail gun lets a construction crew build far more, faster. It does not fire the crew. The same is true of a well-designed agent: it removes manual hours, compresses quote-to-close time, and lifts customer experience while keeping humans on the loop where judgment matters.

That framing is also an AX decision. Google’s guidebook advises automating tasks that are unpleasant or need scale, and augmenting tasks people enjoy or disagree on. The research on automation bias underlines why the human role must be designed deliberately and not assumed: studies show people can over-rely on confident automation, and that simply putting a human “in the loop” without real authority creates the illusion of oversight rather than the substance. One study of decision support found prescribing errors rose by 56.9% when an automated system gave erroneous guidance. Good AX designs the human’s role as an empowered overseer, with the visibility and the off-switch to intervene. Bad AX makes the human a rubber stamp, which is how automation bias turns into expensive mistakes.

Outcomes over outputs, in numbers

The clearest way to keep an AI program honest is to measure outcomes, not activity. “We deployed a chatbot” is an output. “We removed manual hours, cut response time, and improved retention” is an outcome. McKinsey’s data shows why this discipline matters: having AI “somewhere in the organization” is no longer a differentiator when nearly everyone does, and only a minority connect AI to enterprise-level financial impact. The ones who do are the ones who redesign the work and define where human judgment enters. That is an experience and operating-model achievement, not a procurement one.

Recommendations

For founders, CEOs, and operators deciding where to put attention, the staged agenda is concrete.

Stage 1, now: reframe the question. Stop asking “what is our AI strategy” and start asking “which specific workflow will an agent own, and what outcome will it produce.” Pick one high-value workflow, such as quote-to-close, support triage, or back-office processing, and define the business metric it must move. The benchmark that should change your plan: if you cannot name the outcome and the number, the project is not ready to fund.

Stage 2, in the build: make trust and oversight explicit design requirements. Require that any agent show its work, state its confidence and limits, log its actions for review, and offer a frictionless human handoff. Adopt an existing standard rather than inventing one: Microsoft’s 18 HAX guidelines and Google’s People + AI Guidebook are free and battle-tested. The threshold to watch: if users cannot see what the agent did and easily override it, do not expand its autonomy.

Stage 3, at adoption: measure outcomes and calibrate autonomy. Track hours removed, cycle-time reduction, resolution rate, and retention, not usage counts. Start the agent with narrow authority and widen it only as measured reliability and user trust grow, the “human-on-the-loop” progression. The benchmark that should expand autonomy: sustained accuracy plus rising user confidence. The signal that should pull it back: rising escalations, error patterns, or any erosion of customer trust.

Stage 4, ongoing: treat the experience as a living system. Generative systems drift and contexts change. Monitor performance, refresh knowledge sources, and watch for automation bias in your own people. The point is not to reach a finished state but to build a relationship that compounds, because the relationship, not the feature list, is the moat.

Caveats

  • AX is an emerging discipline and its vocabulary is unsettled. “Agentic Experience” (Late Checkout) and “Agent Experience” (Netlify’s Biilmann) describe related but distinct ideas, the former about the human-agent relationship and the latter about agents as users of products. Treat anyone claiming a finished, authoritative framework with appropriate skepticism.
  • Several headline statistics deserve careful reading. The MIT 95% figure has been criticized for a narrow six-month P&L definition of success that ignores efficiency and other gains, so it is best read as a signal about integration difficulty, not proof that AI does not work. Gartner’s cancellation forecast is a prediction, not an observed result. Widely cited triple-digit agentic ROI figures, such as PagerDuty’s, are executive expectations rather than realized returns.
  • Consumer trust data is genuinely mixed and moving. Adoption is rising even as trust softens in several surveys, which suggests habituation rather than satisfaction. Leaders should not read rising usage as a verdict that the experience is good.
  • The strongest, most rigorously validated material in this space, the Microsoft, Google, IBM, and Apple frameworks and Nielsen’s paradigm analysis, predates the current agentic wave and was written for AI features broadly. The autonomous, multi-agent case adds new questions, especially around permissions, memory, and liability, that the field is still working out in real time.

Sources

Enjoyed this article?

Explore more Revival Group perspectives on AI operating systems and operational transformation.

View All PostsBook a call