Short answer
Do not begin by asking AI to produce more sales copy. Connect enquiries, first meaningful replies, conversations, quotations, follow-up and outcomes by buying stage and objection type. AI can structure unstructured language and surface repeated points of loss; front-line people validate the context and act. The first output should be a living “non-purchase map”, not an autonomous messaging machine.
Won business tells only half the story
Many CRM systems end with two clear states: won or lost. Won work has a value. Lost work may carry “price”, “no response” or nothing at all. Companies study successful customers while treating a larger body of near-miss conversations as noise.
I do not believe non-purchase is purely random. “Let me think”, silence after a proposal, or an enquiry that disappears during a WhatsApp-to-email handover may belong to a recurring pattern. The evidence is simply spread across systems and described differently by each salesperson.
AI’s first job is not to persuade the customer. It is to turn fragmented language into a pattern the company can examine.
Map stages and objections
An objection alone lacks context. “Too expensive” at first contact may indicate unclear positioning. After a proposal, it may mean value and risk were not made visible. Immediately before signature, it may indicate budget procedure or a missing decision-maker.
Place each conversation on two axes:
Buying stage: initial enquiry, discovery, proposal, quotation, internal approval, won or lost;
Objection type: price, timing, trust, capability, internal capacity, comparison, no urgency or unknown.
Add three facts: time to first meaningful reply, whether the next action was explicit and who last acted. The map may show that the problem is not weak selling but a customer group repeatedly losing direction at the same handover.
What AI should and should not do in CRM
Useful AI tasks include extracting need and timing, normalising objection language, summarising history, flagging commitments and missing next steps, and identifying phrases that recur at the same stage. It can prepare a reply draft or meeting brief.
It should not permanently label a customer “low value”, set discounts, invent commitments or send at scale without review. CRM history already reflects human bias. Training automation on old outcomes can turn yesterday’s choices into faster automatic choices.
Keep a source excerpt and confidence indication beside an AI label, and let sales correct it. Correction is not friction; it is some of the company’s most valuable new data.
Hypothetical example: “price” hides an unclear next step
This is a teaching scenario, not a Smark Global client case.
A Hong Kong business-services provider sees “price” across many lost opportunities and considers a broad discount. The team samples pre- and post-proposal conversations. AI helps classify stage, objection and next action; salespeople validate the result.
Few customers explicitly ask for a lower fee. A more common pattern is an overly broad proposal: customers cannot see the people they must commit, when the first result appears or whether they can stop after an initial stage. Sales has been using “price” as shorthand for the ensuing silence.
The first test is not a discount. It is a four-week diagnostic and contained workflow test with client input, deliverables, decision criteria and next-stage choices made explicit. The company measures reply after proposal, movement into a decision meeting and completeness of loss reasons.
A four-week implementation
Week one: define the language. Sales jointly defines stages, objection types and a meaningful reply. Keep “other” and “unknown”.
Week two: sample and calibrate. Two colleagues classify the same recent conversations and compare differences. Fix definitions before asking AI to assist at scale.
Week three: connect daily work. AI drafts a summary, stage, objection and next action after an interaction. The salesperson confirms before CRM is updated.
Week four: examine patterns, not a staff leaderboard. Find repeated loss by stage, customer group, product and handover. Select one process hypothesis to test. Never use uncalibrated AI labels as automatic performance judgement.
Consent, bias and front-line review
Sales conversations can contain identity, contact details, budgets and confidential plans. Establish purpose, access, retention and a suitable tool environment. Recording and transcription require the appropriate notice or consent.
Sample classifications across customer groups, and give staff a visible correction path. The system must not trade customer choice, honesty or brand trust for a higher reply rate.
Non-purchase map checklist
- Did front-line staff help define stages and objections?
- Are unknown and multiple causes allowed?
- Can each AI label be traced to the conversation?
- Are meaningful first reply and explicit next step recorded?
- Can staff correct classifications, and are corrections retained?
- Does the analysis lead to one contained workflow test?
- Are uncalibrated results kept away from automatic staff or customer judgement?
Further reading
Want the full version of this approach?
To continue exploring the patterns hidden inside what customers do not say, view Harry Lui’s book on Amazon or register for preview material and enterprise AI notes.
Harry Lui's book works through the same path in full: from business context to company data to a reusable operating system.
Source: Smark Global editorial, paraphrasing ideas from Harry Lui's book
Last updated: 2026-09-13（first published: 2026-09-13）
