Noma Edeki
CASE STUDY 03 / 03
Concept · Prototype
AI PRODUCT · CONVERSATIONAL UX
HUMAN-IN-THE-LOOP
AI Customer Assistant — self-initiated concept

Designing an AI customer assistant that knows when to act — and when to stop.

Small businesses often don't have a dedicated customer support team. I explored how an AI assistant could help business owners manage customer conversations without taking important decisions out of their hands.

RoleProduct Designer
TypeSelf-initiated AI product concept
FocusAI product strategy · Conversation design · Decision systems · Human-in-the-loop UX
StatusConcept · Prototype
THE DECISION LOOP Understand Classify Decide Act Escalate AUTOMATED HUMAN REVIEWS
EVERY MESSAGE MOVES THROUGH THIS SEQUENCE BEFORE ANYTHING IS SENT
THE PROBLEM

For a small business owner, a customer message rarely arrives as a neat support ticket.

Behind every message, the business owner has to figure out: what does this person want? How urgent is it? Should I respond? Should I follow up? What should I say? When should I stop?

For a business with a small team, these decisions become repetitive and time-consuming. But automating them completely introduces another problem: what happens when the AI gets it wrong?

That became the core design question.

"How much?"
"Is this available?"
"I sent money yesterday and haven't received anything."
"I'll get back to you."
"Hi."
"An assistant that pretends to know more than it does isn't helpful — it's a liability."
— the principle that shaped every major decision
THE PRODUCT QUESTION

Not another
"AI chatbot for your business."

The more interesting question was: how might AI handle routine customer conversations while knowing when to hand control back to a human?

That changed the product from a simple chat interface into a decision-support and automation system.

The user I focused on: small businesses with roughly 1–7 people, where the owner or a small team may still be directly responsible for customer communication. No dedicated customer success team. No time to constantly monitor conversations. No complex CRM infrastructure. No one reviewing every AI response.

The experience had to be simple enough to use without training — while still giving the business owner visibility and control.

WHAT SHOULD THE AI ACTUALLY DO

Five decisions behind every response.

01
Understand
What is the customer actually asking?
02
Identify intent
Is this a question, enquiry, complaint, purchase signal, follow-up — or something else?
03
Determine urgency
Does this need immediate human attention?
04
Assess confidence
How confident is the AI that it understands the customer's intent?
05
Decide
Respond · Draft · Follow up · Escalate · Do nothing
DESIGNING CUSTOMER STATES

A customer is more than a "Lead."

I explored customer states based on what the conversation was actually indicating — so the business owner understands what needs attention, not just a list of conversations.

State

New enquiry

Initiated a conversation but hasn't expressed a clear intent yet.

State

Interested

Has shown meaningful interest in the product or service.

State

Needs follow-up

The conversation has paused, but there's a legitimate reason to follow up.

State

Ready to buy

Has demonstrated a strong purchase signal.

State · Urgent

At risk

The conversation contains signals of frustration, dissatisfaction or abandonment.

State

Resolved

The customer's need has been addressed.

NOT EVERY MESSAGE NEEDS A RESPONSE

I designed explicit
stop conditions.

Stop condition 01

The customer says no

If the customer says "not interested" or "stop messaging me" — the AI stops proactive communication. No clever re-engagement. No "just checking in."

Stop means stop.

Stop condition 02

The AI isn't confident enough

If the system cannot confidently determine what the customer wants, it shouldn't confidently invent an answer. Instead: ask a clarifying question, or escalate to the business owner.

Uncertainty becomes a visible product state — not a hidden failure.

Stop condition 03

The conversation becomes sensitive

Refund disputes. Complaints. Payment problems. Highly frustrated customers. Requests outside the business's defined rules.

The AI can identify the situation and prepare context for the owner. But the human makes the decision.

Stop condition 04

Follow-up has reached its limit

An interested customer shouldn't receive endless "just checking in 😊" messages. The business owner defines a maximum number of proactive follow-ups. Once that limit is reached, the AI stops.

DESIGNING THE CONFIDENCE LAYER

Showing uncertainty without adding complexity.

Confidence becomes part of the system, not exposed AI terminology. The AI should not behave with more certainty than it has.

High confidence

"Customer is asking about pricing."

→ AI responds
Medium confidence

"Customer may be asking about pricing, but intent is unclear."

→ AI asks for clarification
Low confidence

"I don't have enough information to respond safely."

→ Human review
DESIGNING THE OWNER'S EXPERIENCE

The owner reviews exceptions,
not every conversation.

The dashboard answers: what happened? What needs my attention? What is the AI doing? Where do I need to step in?

Today
12

conversations handled by AI

Pending
3

customers need follow-up

Action needed
2

conversations need your attention

Priority
1

high-priority issue

Conversa AI customer assistant — home dashboard

The conversation view.

Instead of a chat transcript, the owner sees context. They can approve, edit, take over — or stop automation entirely.

CONVERSATION — SARAH
IntentProduct enquiry
ConfidenceHigh
Customer stateInterested
AI actionResponded with product information
Approve
Edit
Take over
Stop automation
THE PRODUCT ARCHITECTURE

Five layers,
one principle.

Layer 01Input
Customer conversations
Layer 02Understanding
IntentSentimentContext
Layer 03Decision
ConfidenceUrgencyCustomer state
Layer 04Action
RespondDraftFollow upEscalateStop
Layer 05Human control
ReviewEditOverrideTake over
DESIGNING FOR FAILURE

The system can fail even when the interface is working perfectly.

I explored failure states as core UX — not edge cases.

Failure mode

Wrong intent

The AI misunderstood the customer.

Design responseAllow the owner to correct the intent and update the conversation state.
Failure mode

Missing information

The AI doesn't have enough context.

Design responseAsk the customer for clarification, or escalate.
Failure mode

Conflicting information

The customer's message conflicts with known business information.

Design responseDon't guess. Flag for review.
Failure mode

Negative sentiment

The customer appears frustrated.

Design responseReduce automation and bring in a human.
Failure mode

Unwanted automation

The customer asks not to be contacted.

Design responsePermanent stop condition.
WHERE THIS STANDS

A concept, not a shipped product.

Problem framing
Complete
System logic
Defined
Prototype
In progress
Validation
Not yet run

This is a self-initiated concept. The value here is in the thinking behind the decision system — not production metrics. The next step would be testing the assumptions the system is built on.

WHAT I'D VALIDATE NEXT
Can business owners correctly understand the AI's decisions?
Do customer states actually help them prioritise?
Are the stop conditions understandable?
Do owners trust AI-generated responses enough to approve them?
Does the system reduce the amount of manual conversation management?
At what confidence level do users want human review?
WHAT I LEARNED

Designing AI products means asking different questions than designing traditional interfaces.

What does the AI know?
What doesn't it know?
How confident is it?
What is it allowed to do?
When should it ask?
When should it stop?
When should a human take over?

Those decisions became the product.

WHAT I BROUGHT TO THIS PROJECT
AI product strategy

Defined a workflow around understanding, confidence, action and escalation.

Conversational UX

Designed interactions around intent, not simple chatbot replies.

Human-in-the-loop

Created explicit intervention points for business owners.

AI safety & control

Designed uncertainty, escalation and stop conditions as core product states.

Systems thinking

Mapped the relationship between messages, decisions, business rules and human action.

Product design

Translated an ambiguous AI opportunity into a structured concept and interaction model.

MORE CASE STUDIES
01
MoveFinance→
Fintech · Immigrant Banking · Virtual Card
Designing a virtual card flow for immigrants navigating international payments and financial services.
02
Oloja→
Marketplace · Trust · Negotiation
A service marketplace designed around transparent offers, negotiation and secure transactions.
+
Other Projects→
Websites · Henna Place · Trian Energy
Selected website projects, including Henna Place and Trian Energy.
Noma Edeki
Product Designer · Remote
Work About Contact Resume ↓ LinkedIn
© 2026 Noma Edeki