Source: Unsplash
Gartner predicts 40% of enterprise applications will have AI agents by end of 2026, up from less than 5% in 2025. Salesforce just shipped Agentforce Summer '26 with multi-agent orchestration. SAP launched an AI Agent Hub backed by a €100M fund at Sapphire 2026. The enterprise AI agent wave is real. But Gartner also says more than 40% of these projects will be canceled by 2027. The reason isn't the AI models. The reason is the interface layer, and the fact that almost nobody designing these products is thinking clearly about it.
I have spent the last three years working inside products that use AI to do things users used to do manually. Summarizing contracts. Routing support tickets. Writing first drafts of emails. Generating reports from raw data. Some of these products shipped well. A lot of them did not. And when I try to figure out what separates the ones that worked from the ones that quietly got pulled from roadmaps, the answer is almost never the model. It is always the interface.
This week has been a big one for enterprise AI agents. Salesforce released its Summer '26 update to Agentforce, bringing multi-agent orchestration where different specialized agents can work together as a coordinated team across end-to-end workflows. At the same time, SAP announced its AI Agent Hub at Sapphire 2026, a marketplace for discovering and deploying enterprise agents with over 680 partner submissions already in the pipeline and a €100M backing fund. Meanwhile, Deloitte's 2026 State of AI in the Enterprise report, which surveyed over 3,200 business and IT leaders globally, found that 74% of organizations plan to adopt agentic AI within the next two years.
So the momentum is real. The investment is real. The problem is also real.
"Over 40% of agentic AI projects will be canceled by end of 2027, primarily due to escalating costs, unclear business value, and inadequate risk controls."
— Gartner, June 2025 Press Release
Sit with that number for a second. We are heading into a market where nearly half of everything being built right now will get pulled before it ever reaches production at scale. And the reasons Gartner cites are not "the model was too slow" or "the API was too expensive." The reasons are unclear business value and inadequate risk controls. Both of those are design problems. Both of those are interface problems.
What "Agent Washing" Is Doing to Enterprise Products
Gartner introduced a term recently that I think is the most accurate description of what is happening in enterprise software right now: "agent washing." Vendors are rebranding existing products, chatbots, RPA tools, basic workflow automation, as AI agents without building actual agentic capability underneath. Gartner estimates only about 130 of the thousands of vendors claiming to build agentic AI products are real.
This creates a credibility problem for the entire category. When a product promises autonomous agent behavior but delivers glorified autocomplete with a chat wrapper, users learn to distrust the whole thing. The next time something actually works autonomously, they override it. They double-check every output. They stop trusting the system, and the system becomes useless.
I have watched this happen in real products. A user gets burned once by an AI agent that confidently did the wrong thing, and from that point forward they treat the agent like a broken calculator. Trust, once lost in a human-AI workflow, is almost impossible to rebuild through feature updates alone.
via GIPHY
The Five Things Enterprise Agent Interfaces Keep Getting Wrong
After watching products succeed and fail in this space, I have started to see the same five failure patterns repeat across different companies, different industries, different tech stacks. The model is rarely the problem. Here is where the interface consistently breaks down:
- No planning visibility. Users do not know what the agent is about to do before it does it. The agent just acts. In high-stakes enterprise workflows, that is terrifying. If I am an accounts payable manager and an agent is processing invoices, I want to see the plan before execution, not just the result after. Salesforce's Agentforce is building toward this, but most implementations ship without it.
- No honest confidence signaling. AI agents produce probabilistic outputs. Not every answer is equally reliable. But most enterprise agent interfaces present every output with the same visual weight and tone. There is no way for the user to know whether the agent is 95% confident or 55% confident. Designing for uncertainty is not optional in agentic UX.
- No real override pathway. When an agent does something wrong in the middle of a multi-step workflow, users often cannot cleanly undo or redirect without starting over. The interface needs graceful interruption patterns, not just a back button.
- Memory that is invisible. Agents remember context across sessions but users do not know what the agent remembers or why it is making certain assumptions. This creates a mismatch between user mental models and actual agent behavior. SAP's Joule Studio 2.0 is moving toward surfacing this, but it is early.
- No audit trail in the UI. Especially in regulated industries. Finance, healthcare, legal. Every agent action needs to be traceable and explainable in the interface itself. Most products offload this to logs that only developers can read. That is not a governance model. That is a liability waiting to happen. Deloitte found that only 21% of organizations have mature governance models for autonomous AI agents right now.
Why This Is Happening Now, Not Two Years Ago
The AI model capabilities moved faster than anyone expected. GPT-4 class reasoning plus function calling made it technically possible to build agents that could execute real workflows with minimal scaffolding. Product teams rushed to ship. The interface design thinking did not keep up.
Part of the problem is that most product designers have not shipped agentic products before. The patterns do not exist in design systems yet. There is no equivalent of the Nielsen Norman Group heuristics for autonomous AI agents. Teams are building on instinct and copying whatever the leading AI demo showed on stage at a product keynote.
The other part is organizational. The teams building the model pipelines and the teams designing the interfaces are often disconnected. Engineers make implicit design decisions about how agents surface outputs, handle errors, and communicate uncertainty. Those decisions compound into a product that users do not understand and cannot trust.
I wrote about this pattern with conversational UX on my Medium a while back. The same mistake keeps showing up: teams design for the happy path, where the agent does the right thing every time, and completely ignore what happens when it does not. Failure states are where user trust is actually built or destroyed.
What Salesforce and SAP Are Getting Right
I want to be fair here. The major enterprise platforms are starting to think about this more seriously. Salesforce's Agentforce Summer '26 ships with what they are calling the Agentforce Trust Layer, a set of guardrails designed to protect enterprise data as agents interact with Tableau and Slack and other systems. The multi-agent orchestration model, where specialized agents hand off tasks to each other, is architecturally sound and gives product teams more control over what each agent is responsible for.
SAP's approach with Joule Studio 2.0 is also interesting because it is building the development environment for enterprise agents with native support for Python, ABAP, and LangChain. The fact that they are treating agent development as a first-class design problem and not just an engineering one is a good sign.
But these are platform-level decisions. Individual product teams still need to make hundreds of interface decisions that these platforms do not prescribe. And right now, most teams are guessing.
The Opportunity Hiding Inside the Failure Rate
Here is the flip side of Gartner's 40% failure prediction. If most products in this space are failing because of interface and governance problems, and not because the AI cannot do the work, then the teams that get the design right have a significant advantage.
Gartner's long-term projection shows agentic AI driving approximately 30% of enterprise application software revenue by 2035, surpassing $450 billion. Deloitte finds that up to 75% of companies will invest in agentic AI in 2026. The market is not going anywhere. But the products that survive the shakeout will be the ones where users actually trust the agent to do work on their behalf.
That trust is earned through interface design. It is earned through transparency about what the agent is doing and why. It is earned through honest uncertainty communication. It is earned through recovery patterns that work when things go sideways. It is earned through audit trails that compliance teams can actually read.
None of that is model work. All of that is design work. And most enterprise product teams are not staffed or structured to do it well right now. That gap is where the real opportunity sits, for designers, for product leaders, and for the companies willing to slow down slightly and build agent interfaces that people actually trust.
I keep seeing the same pattern across every enterprise AI product I review or work on. The underlying capability is impressive. The interface makes it feel unpredictable. Users disengage. Projects stall. Budgets get cut. And the narrative becomes "AI agents do not work yet" when the real diagnosis is "we designed them like they were always right."
They are not always right. Your interface should say so. That honesty is the feature.
I would genuinely like to know what you are seeing on your end. If you work in enterprise software, have you shipped an AI agent feature? What did the interface look like for handling failures or uncertainty? Drop a comment below. I read every one.
Sources: Gartner Press Release: "40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026" (gartner.com), Gartner: "Over 40% of Agentic AI Projects Will Be Canceled by End of 2027" (gartner.com), Deloitte: "SaaS Meets AI Agents" 2026 Report (deloitte.com), Salesforce Summer '26 Release Announcement (salesforce.com), SAP AI Agent Hub at Sapphire 2026 (savictech.com), XMPRO: "Gartner's 40% Agentic AI Failure Prediction Exposes a Core Architecture Problem" (xmpro.com)