User Trust and Transparency

Building an agent that is 99% accurate is an engineering feat. But if the user feels like the agent is 50% accurate, your product is a failure. Trust is a perception problem, not just a technical one.

In this lesson, we will explore the "Trust Paradox": Why showing an agent's mistakes can actually make the user trust it more.

---

1. The "Black Box" Anxiety

When an agent says "I have analyzed your portoflio and you should sell Stock X," the user's immediate thought is "Why?"

If the UI doesn't answer "Why," the user will assume the agent is hallucinating or biased.

---

2. Strategy: Show Your Work (Citations)

Citations are the single most effective trust-builder in RAG and Agentic systems.

• The Wrong Way: "According to the news, it will rain tomorrow."
• The Right Way: "According to [Bloomberg News - Jan 4th], it will rain tomorrow (80% probability)."

The React Implementation

Clickable "Footnotes" that open a side drawer showing the Raw Text Chunk or the PDF Page the agent read. This allows the human to perform "Spot Checks."

---

3. Strategy: Tool Intent Disclosure

When an agent calls a tool, it's not enough to say "Using Python." You should say what the agent is trying to achieve with that tool.

• UI Message: "I'm using the Python Calculator to verify the sum of your invoices because I want to make sure the total matches your bank statement."
• Why it works: Even if the calculation fails, the user understands the Intent of the agent, and can offer a correction.

---

4. The "I'm Not Sure" State (Confidence Scoring)

Most LLMs are "Convinced" of their own answers, even when they are wrong. You can use a Logit Bias check or a Self-Critique Node (Module 10) to generate a "Confidence Score."

• If Confidence < 70%: The UI should show a "Warning" icon: ⚠️ "I'm providing this answer based on limited data. You may want to verify this specific section."

Outcome: You have shifted the responsibility from the system to the user, preventing a "Silent Failure."

---

5. Transparency in Pricing

Agents cost money (Tokens). A user should never be surprised by a bill.

• The "Pre-Flight" Estimate: Before a long-running agent starts, show: "This task will involve ~15 search steps and cost approximately $0.40. [Proceed]."
• Real-time Counter: A small tally in the corner of the UI showing total tokens used in this session.

---

6. Managing the "Hallucination" Conversation

If an agent is caught hallucinating by a guardrail (Module 3.4), don't hide it.

Bad UX: The agent crashes or says "I can't help." Great UX:

• Agent: "I was about to claim that your warranty expires in 2026, but I realized I couldn't find a source for that date. Instead, I'll tell you that the warranty is usually 'Limited Lifetime' according to the FAQ."

The user now trusts the agent MORE because they see the agent actively checking its own facts.

---

Summary and Mental Model

Think of the Agent like a New Employee.

• If the employee says "I'm done" and walks away, you're worried.
• If the employee says "I'm done, here are the three spreadsheets I used, here is where I found the data, and I'm 90% sure about the third column," you feel confident.

Transparency is the antidote to the "Uncanny Valley" of AI behavior.

---

Exercise: Trust Design

1. Citations: Design a small UI component that shows a "Source" for a claim.
   • What happens when a user clicks it?
   • (Hint: Should it open a full URL or a specific PDF preview?)
2. Confidence: You have an agent that predicts house prices.
   • If it only finds 1 comparable house, what should the "Trust Bar" look like?
   • What "Helpful Hint" should the agent give the user to improve the search?
3. Disclosure: Why should you tell a user that an agent is "Running in a Sandbox" (Module 7)?
   • (Hint: Does it make the user feel more or less safe to upload their files?) Now you are ready to build the whole stack. Next Module: Full Stack Integration.