Conversation overload
Important requirements could be buried inside a long back-and-forth conversation.
Back to case studies
Confidential client case study
AI Sales Copilot for Faster, Safer Customer Replies
A specialty retailer needed a faster way to understand long customer conversations, prepare product recommendations, and respond consistently. I built a browser-based assistant inside the team's existing inbox that creates review-ready replies and quotations while keeping every customer message under human control.
IndustrySpecialty retail
TeamSales and customer support
StatusWorking internal pilot
Overview
From a long conversation to a review-ready response.
The copilot reads the conversation already open on screen, prepares a useful response or quotation, and places it into the reply box for the employee to review. It never sends automatically.
The business problem
Good sales replies took too many manual steps.
Employees had to reread long threads, remember what the customer already said, assemble a suitable product recommendation, check the quotation, and write a clear response while other conversations continued to arrive.
Repeated writing
Employees repeatedly drafted similar explanations, recommendations, and follow-up questions.
Quotation risk
Product names, prices, compatibility, and totals needed verified business data—not AI guesses.
Inconsistent tone
Reply quality and clarity could vary depending on workload and employee experience.
Partial context
The inbox could load only part of a conversation, creating a risk of assuming details agreed earlier.
Automation risk
A wrong response must never reach a customer without an employee reviewing it first.
Challenges
The assistant had to be fast without taking control away from the team.
- Fit the existing workflowThe team should not need to leave the inbox or copy full conversations into another tool.
- Use verified informationProduct details, prices, availability, compatibility, and totals must come from trusted business data.
- Fail safelyIf conversation roles or context are uncertain, the assistant must warn the employee or ask a clarifying question.
- Respect customer privacyThe extension reads only the open conversation and avoids broad browser permissions.
- Support natural communicationSuggested replies need to match the customer's English, Tagalog, or Taglish.
- Keep humans accountableThe employee remains responsible for checking, editing, and sending every response.
Solution
A human-reviewed sales assistant inside the existing inbox.
I built a browser extension that reads the customer conversation already visible to the employee, identifies the recent context, and sends a structured request to a private workflow backend.
The backend retrieves verified product information, applies deterministic quotation and compatibility rules, then uses AI only to turn those facts into a concise, natural response. If verified information is missing, the system asks a question instead of inventing an answer.
Employees can generate, shorten, revise, summarize, or make a response more sales-oriented. The result can be inserted into the inbox composer, but the assistant never clicks Send. Human review is a permanent safety boundary.
System architecture
Verified data first, helpful language second, human approval last.
The workflow separates customer context, business rules, AI writing, and final approval so the model cannot control prices or customer-facing sends.
Open conversation
Context extraction
Verified product data
Quotation rules
AI reply draft
Employee review
Manual send
Working proof
The pilot proves the workflow without pretending AI is infallible.
- 72 automated testsThe current extension test suite verifies conversation extraction, quotation behavior, reply orchestration, browser permissions, and send-safety rules.
- Live workflow backendThe assistant sends structured requests to a working private workflow that returns review-ready replies and quotations.
- Human-controlled sendingThere is no automatic-send path; only the employee can send a customer message.
- Verified quotation logicProduct facts and arithmetic are handled outside the language model, reducing hallucination risk.
- Context warningsThe interface warns when only part of a conversation may be visible and encourages the employee to load more history.
- Fast refinementEmployees can revise a quote, shorten a response, summarize the customer, or adjust the sales tone without starting again.
Technologies
A lightweight layer over the tools the team already uses.
Browser extensionTypeScriptReactWorkflow automationVerified product dataDeterministic quotation logicAI-assisted writingHuman review
Gallery
The sales assistant in action.
These sanitized screens show the employee workflow without disclosing the client, customer identities, private URLs, or internal credentials.
Assistance inside the existing inbox
The employee opens the copilot without switching tabs or manually copying the full conversation into another tool.
Conversation-aware reply generation
The assistant reads the visible thread, identifies the customer's requirements, and prepares a response for employee review.
One-click response refinement
Employees can shorten, revise, summarize, ask a question, or adjust the sales tone without rewriting the response from scratch.
A simple three-step workflow
Open the assistant, generate a useful reply, then review and insert it into the conversation composer.
Human review before every send
The interface makes the safety boundary explicit: the assistant suggests and inserts text, while the employee decides what reaches the customer.
Why this matters
Sales automation should support judgment—not replace it.
This system removes repeated reading, quotation assembly, and response drafting while preserving the employee's responsibility for accuracy, tone, and the final customer message.
Want faster replies without unsafe auto-sending?
Book a free discovery call and we will identify where a human-reviewed assistant can reduce repetitive sales work.
Book My Free Discovery Call