People no longer use AI just to get answers. Increasingly, we use it to delegate work. That shift demands new design principles.
In 2024 we designed for an AI that responds. In 2026 we design for an AI that acts. These are the eight principles that separate one from the other.
Declare the scope, not the authorship
That the content is AI-generated is already assumed. What the user doesn't know is what it's allowed to touch. Declare capabilities, permissions and limits before the first action.
Disclosing it's AI
A visible note whenever the content was generated by a model.
AI-generated content may be incorrect.
Permissions declared
Declare capabilities, permissions and limits before the first action.
- Read your calendar
- Draft emails
- Send without your approval
A disclaimer that it's AI is no longer enough. You have to declare what it can do.
Show the plan before executing it
The agent already knows what to do. Expose the sequence as an editable artifact: the user corrects the plan, not the result.
Suggesting the next prompt
Examples of what to write next to steer the generative process.
Editable plan before executing
The agent exposes the plan before executing it.
- Search overdue invoices in the ERP
- Group by client and amount
- Draft a reminder for each
The user corrects the plan, not the result.
Ask for approval first, not opinion after
When an action moves money, sends messages or writes to a system, feedback arrives too late. Define what requires human-in-the-loop.
Collecting feedback after
A thumbs up or down on output already delivered.
Approval at the point of risk
What's irreversible stops and waits for a human decision.
Will send 240 WhatsApp messages.
Irreversible action, requires your approval.
A thumbs up or down on something that already happened isn't feedback, it's a record.
Make execution visible
An agent can run multiple steps over several minutes. Execution is part of the experience too. Show what it's doing, what tools it's using, what's already done and where it's still working.
Telling the behind the scenes
Communicating training runs and system updates.
NEW Model updated · v2.4 · March 2024
Execution visible live
Execution is shown step by step, as it happens.
- Grouping overdue invoices by client 0:04
- Drafting personalized reminders 0:41
- Sending WhatsApp messages (86 / 240) 1:12
A black box that takes two minutes to respond breeds more distrust than a visible error.
Show evidence, not just sources
A list of documents at the end is no longer enough. Every claim should be verifiable against the evidence that backs it. If that evidence doesn't exist or can't be checked, the system should say so. Silence communicates too: it's often read as certainty.
Listing the sources used
A block of citations at the end with the documents consulted.
- Portfolio_Q2.xlsx
- Collections_Policy.pdf
Evidence per claim
Every data point anchors to its origin, and estimates are marked as estimates.
The outstanding balance is $1,200, per invoice #1042 , overdue since August 3. The client will likely pay this week. No source · estimate
Citing a source isn't the same as showing the evidence inside it.
Design for reversible actions
When an agent modifies a system, an error shouldn't be permanent. Design mechanisms to review, undo or recover actions whenever possible. What's irreversible should require explicit approval.
Didn't apply
Generated text had no consequences. If it was wrong, it got discarded.
Reversible by design
An agent's mistake can change a real system.
240 WhatsApp messages sent
Undo · Available for 10sIf an action can't be undone, the decision to take it can't be the agent's alone.
Stop and hand off to a human, always
The stop control visible at all times, and one that actually stops things. The handoff carries the full context: the user shouldn't repeat what they already said.
Didn't apply
The response ended in two seconds. There was nothing to stop.
A real, always-available stop
The agent needs a real stop and a handoff with context.
A stop button that's slow to react is worse than not having one.
Show the limits, don't hide them
An agent with a voice, memory and initiative humanizes itself. The challenge is no longer avoiding a human feel — it's making its limits visible: what it knows, what it doesn't, what it can do, and when it's inferring rather than answering from evidence.
Don't humanize the AI
Avoid language and design that attribute human qualities to it.
Limits visible, not hidden
The challenge is making its real limits visible.
- I'm reviewing your case, give me a minute.
- Checking 3 systems · no access to history before 2023.
This principle hasn't changed since 2024. What changed is how much it matters.
In summary
A framework for designing agents people can trust.
Declare the scope — what it can do, and what it can't.
Show the plan — before executing it.
Ask for approval — before acting, not after.
Make execution visible — while it happens.
Show evidence — not just a list of sources.
Design for reversible actions — the irreversible ones need approval.
Stop and hand off — always available, always with context.
Show the limits — what it knows, what it doesn't, when it's inferring.
AI can help your business. But only you know your operation.
Frequently asked questions
What is agentic AI?
AI that doesn't just respond: it executes steps, uses tools and takes actions within your systems to complete a task, with or without supervision at each step.
What is human-in-the-loop?
The point in a process where a person must approve before an action executes, especially if it's irreversible or involves money, customer communication or changes to a system.
Do these principles replace good conversational UX practices?
No. They build on them and add a new layer: the one for an agent that acts. The 2024 conversational UX principles are still the foundation.
The content of this infographic takes UX for AI: Design Practices for AI Developers, with John Maeda, a course available on LinkedIn Learning, as a reference. Concepts derived from the course have been adapted and complemented with Entrópica's perspective, including the 2026 reading and examples.
Transparency: At Entrópica, we use AI as part of our research and editing process. The content reflects our knowledge, experience, and judgment.