Stop Asking AI for Answers: Start with Business Context

Fast answers can send management in the wrong direction

I often hear the opening question: “How can we use AI to increase sales?” There is nothing wrong with the ambition. The problem is everything the question leaves out. Which offer? Which customer? Where do enquiries disappear? How does sales follow up today? Has the company already tried pricing, advertising or process changes, and what happened?

With almost no context, AI fills the gaps with what is common on the internet. The answer sounds complete—produce content, segment customers, automate follow-up, improve experience. Any of those may be sensible. None tells this company what it should do first.

My starting view of AI is simple: do not rush to treat it as an answer machine. Its more valuable role inside a company is to find order in fragmented, sometimes contradictory information. Producing an answer is quick. Understanding the business situation is the work.

The five-box context sheet

Before an AI project, I ask the team to complete one page. It is not a technical specification. It is a business context sheet:

1. Outcome: What observable result do we want to change? “Increase the follow-up rate for qualified enquiries” is better than “improve efficiency”.

2. People: Who touches the process—customers, front-line staff, managers, finance or partners—and what does each person know?

3. Current workflow: What actually happens from start to finish, across real systems and informal workarounds?

4. Previous attempts: What has changed before? Who supported it, who resisted it and why did it not last?

5. Results and evidence: Which numbers, conversations, records and exceptions have we reviewed? What remains an assumption?

The sheet stops a team from presenting its own theory as fact. When the evidence sits together, AI can help expose a useful contradiction: sales may blame price while most lost enquiries actually disappear before a meaningful first reply.

Turning opinions into testable hypotheses

Meetings are rarely short of opinions. The owner says the market is weak, sales says the leads are poor, marketing says follow-up is slow, and operations says the team is stretched. Each statement may contain part of the truth.

I ask AI to do three jobs. Separate evidence from inference. Identify conditions that recur across different records. Then propose a small test. That test needs a defined scope, owner, time limit and decision rule.

Instead of replacing the CRM in one move, review six weeks of enquiries and classify source, first meaningful response time, question type and final status. If slow, vague replies repeatedly coincide with loss, test AI-assisted triage and reply preparation for one product line. Transformation begins when the company learns something it could not see before—not when it buys the largest system.

Hypothetical example: more enquiries, no more business

This is a teaching scenario, not a Smark Global client case.

A Hong Kong professional-services firm increases web enquiries by 30 per cent after an advertising campaign, but completed sales remain flat. Management assumes lead quality has fallen and prepares to change targeting again.

The five-box exercise reveals that enquiries arrive through email, WhatsApp and forms. Staff copy details between channels and wait for a manager to assign ownership. The highest-value enquiries are often the longest and take most time to interpret, so they receive the slowest useful response.

AI does not need to replace the salesperson. It can structure each enquiry into customer context, need, timing, risk and next action, then place it in one queue. The commercial measures are not “words generated”. They are time to first meaningful response, the proportion of qualified enquiries reaching a meeting, and whether human time moves towards higher-value conversations.

A 30-day way to begin

Days 1–5: choose one problem. It should have a visible start and finish, recur every week and already leave some evidence. “Transform the whole company” is not a first problem.

Days 6–10: complete the context. Speak to the people doing the work, sample raw records and capture the gap between policy and practice.

Days 11–15: let AI organise, not decide. Begin with classification, summarisation, anomaly flags and pattern exploration, while preserving links to source evidence.

Days 16–23: run one contained test. Limit it to one team, customer group or product line. Write success and stop conditions in advance.

Days 24–30: review and standardise. Compare before and after, record failures and exceptions, and only then turn a useful method into workflow, access rules and training.

Privacy, evidence and human responsibility

Better context often contains more customer and operational data. That is not permission to pour a database into a public tool. Decide whether the data may be used, what must be de-identified, who may see outputs and who reviews errors.

AI can organise evidence; it cannot carry management accountability. Pricing, contracts, employment, credit, health and other high-impact decisions require an authorised human. If a conclusion cannot be traced back to evidence, it should not be treated as fact.