The practical answer
Start with the agent's authority, not its chat box. Define allowed actions, permission boundaries, approval points, and recovery paths before designing the conversation or progress display.
Key takeaways
- Uncertainty: uncertainty should be defined early enough to influence architecture, not added during visual polish.
- Autonomy: Treat autonomy as a testable product decision with an owner and a success signal.
- Delegation: Document delegation explicitly so design and engineering do not resolve it differently.
- Permissions: Use realistic content to validate permissions; placeholder data can hide important failures.
- Provenance: Connect provenance to user behavior and business risk rather than treating it as a style preference.
The core principles
1. Uncertainty
Provide undo, correction, and recovery paths. Instrument the relevant behavior before launch so the team can distinguish a successful release from a merely attractive one. A useful validation signal is approval reversals, but the number should be read alongside qualitative evidence so the team understands why behavior changed. One recurring failure mode is presenting probabilistic output as certain.
2. Autonomy
A stronger decision is to design visible checkpoints before irreversible actions. Separate what the team knows from what it assumes, then design the research around the riskiest assumption. A useful validation signal is time to recover from an AI error, but the number should be read alongside qualitative evidence so the team understands why behavior changed. One recurring failure mode is showing confidence without evidence.
3. Delegation
When the stakes are higher, teams should show sources or evidence when claims matter. Use research, production data, support evidence, and usability observation together rather than letting one signal dominate. A useful validation signal is rate of unnecessary confirmations, but the number should be read alongside qualitative evidence so the team understands why behavior changed. One recurring failure mode is hiding what the agent can access.
4. Permissions
Treat the first design as a hypothesis and keep a visible trail from evidence to decision. A useful validation signal is user trust calibration, but the number should be read alongside qualitative evidence so the team understands why behavior changed. One recurring failure mode is making irreversible actions without a checkpoint.
5. Provenance
The design consequence is to measure whether users understand the system's authority. Test with realistic content and edge cases; placeholder data hides many of the problems that appear in production. A useful validation signal is correction rate, but the number should be read alongside qualitative evidence so the team understands why behavior changed.
6. Recoverability
When the stakes are higher, teams should provide undo, correction, and recovery paths.
7. Human oversight
The design consequence is to map what the system can decide versus what needs approval.
A practical framework you can use
A useful framework for Designing AI Agents should help a team move from an ambiguous problem to a testable product decision. The sequence below is intentionally lightweight: it can fit a focused audit, a discovery sprint, or a larger redesign. Do not treat the steps as a rigid waterfall. Research can change scope, testing can reveal a missing requirement, and production data can force a team to revisit the initial diagnosis.
Step 1: Test failure states as seriously as happy paths. Use real constraints, representative content, and the closest available production data. Define a baseline for successful task completion when possible, or at least a clear qualitative success criterion when quantitative measurement is not yet available. Review the step with design, product, engineering, and the people who understand the operational edge cases. Record what changed, what evidence supports the change, and what remains uncertain; this makes later iteration faster and reduces design-by-opinion.
Step 2: Design visible checkpoints before irreversible actions. Define a baseline for correction rate when possible, or at least a clear qualitative success criterion when quantitative measurement is not yet available.
Step 3: Show sources or evidence when claims matter. Define a baseline for time to recover from an AI error when possible, or at least a clear qualitative success criterion when quantitative measurement is not yet available.
Step 4: Measure whether users understand the system's authority. Define a baseline for user trust calibration when possible, or at least a clear qualitative success criterion when quantitative measurement is not yet available.
Step 5: Provide undo, correction, and recovery paths.
Step 6: Map what the system can decide versus what needs approval.
Working on a real product? If you want an expert review of how these principles apply to your product, contact Osama Ali or send a WhatsApp message. I work across UX research, product design, AI/agentic UX, enterprise products, eCommerce, design systems, and Arabic/RTL experiences.
MENA, Arabic, and bilingual considerations
Even when Designing AI Agents is not specifically an Arabic UX topic, regional context can change the design. MENA is not one homogeneous market, so a Saudi product, an Egyptian consumer service, and a UAE B2B platform should not inherit the same assumptions by default. For Designing AI Agents: UX Patterns for Products That Act, Not Just Chat, separate universal product logic from locale, language, regulation, payment, identity, content, or behavior decisions.
Regional consideration — Arabic language quality can affect perceived intelligence and trust. Convert this into a concrete design or research question rather than leaving it as a general cultural statement. For Designing AI Agents: UX Patterns for Products That Act, Not Just Chat, ask which workflow, label, component, policy, or metric could change because of this constraint. Then validate it with the market and user segment you actually serve. This is more reliable than building a generic 'MENA persona' and treating it as evidence.
Regional consideration — Bilingual prompts and outputs need explicit testing.
Regional consideration — Local regulations and sector expectations may change permission design.
Regional consideration — Regional terminology should be grounded in user research.
Regional consideration — Arabic citations and source readability need attention.
Regional consideration — Products should handle language switching without losing context.
How to measure whether the design is working
Measurement for Designing AI Agents should match the user outcome and the business risk. With Designing AI Agents: UX Patterns for Products That Act, Not Just Chat, one number rarely tells the whole story: a shorter task can still be confusing, a higher conversion rate can hide regret, and lower support volume can mean users abandoned the task. Use a small metric set that combines behavior, quality, and operational impact.
Successful task completion: define the event or observation precisely, segment it where relevant, compare it with a baseline, and pair it with qualitative evidence before drawing a conclusion.
Correction rate: define the event or observation precisely, segment it where relevant, compare it with a baseline, and pair it with qualitative evidence before drawing a conclusion.
Approval reversals: define the event or observation precisely, segment it where relevant, compare it with a baseline, and pair it with qualitative evidence before drawing a conclusion.
Time to recover from an ai error: define the event or observation precisely, segment it where relevant, compare it with a baseline, and pair it with qualitative evidence before drawing a conclusion.
User trust calibration: define the event or observation precisely, segment it where relevant, compare it with a baseline, and pair it with qualitative evidence before drawing a conclusion.
Rate of unnecessary confirmations: define the event or observation precisely, segment it where relevant, compare it with a baseline, and pair it with qualitative evidence before drawing a conclusion.
Before launching a change to Designing AI Agents, write the expected direction of change and what evidence would make the team reject its own hypothesis. After launch, review Designing AI Agents: UX Patterns for Products That Act, Not Just Chat by meaningful segments such as language, market, device, role, new versus returning user, or traffic source when those segments are relevant. The purpose of measurement is not to prove that design was right; it is to learn whether the product now supports the intended behavior with less friction, error, or uncertainty.
Common mistakes - and what to do instead
Mistake 1: Presenting probabilistic output as certain. This usually happens when a team optimizes the visible interface before understanding the underlying decision or workflow. In Designing AI Agents: UX Patterns for Products That Act, Not Just Chat, the safer alternative is to state the assumption explicitly, connect it to a user need or constraint, and choose a test that can challenge the assumption. If the team cannot explain what evidence would change its mind, the design decision is probably being treated as preference rather than product reasoning. Document the resolution inside the AI & Agentic UX system so the same debate does not restart in every sprint.
Mistake 2: Hiding what the agent can access.
Mistake 3: Asking for confirmation on every trivial action.
Mistake 4: Making irreversible actions without a checkpoint.
Mistake 5: Showing confidence without evidence.
Mistake 6: Designing only the ideal response.
Implementation checklist
Define the primary user outcome for Designing AI Agents.
Identify the user segments, roles, languages, and markets that materially change Designing AI Agents: UX Patterns for Products That Act, Not Just Chat.
Map the end-to-end workflow before optimizing an isolated screen.
Use realistic content, data, errors, and edge cases in prototypes.
Record assumptions separately from known facts.
Test the highest-risk interaction before polishing low-risk details.
Include accessibility and recovery requirements in the definition of done.
Instrument the behaviors needed to judge the outcome.
Review results by relevant segments rather than relying only on an overall average.
Document decisions and exceptions so the product can scale consistently.
