Prompts Are Not the Core: Turn Enterprise Data into AI Fuel

Prompts can be copied; operating experience cannot

People often equate learning AI with learning prompts. I understand why: a prompt produces an immediate result and makes an easy demonstration. From a business perspective, however, an instruction is not a moat. If you can write it today, a competitor can write it tomorrow.

What belongs to the company is the history of what customers asked, when they cancelled, which failures returned, which orders were delayed and which promises proved difficult to keep. The evidence may be spread across WhatsApp, email, spreadsheets, ERP systems, paper and people’s memories. Its disorder is precisely why organising it can create value that a competitor cannot copy overnight.

Where valuable data hides

Do not begin by asking whether the company has “big data”. Ask what traces the business leaves each day:

enquiries, call notes, WhatsApp and email;

quotations, discounts, wins, losses and refunds;

complaints, repair records, product images and inspections;

stock movements, returns, purchasing and supplier lead times;

routes, delays, proof-of-delivery exceptions and warehouse events;

training enquiries, attendance, follow-up questions and workplace use;

the steps staff take when a standard process fails, and the eventual result.

The most revealing source is often not a polished management report but an exception record. A normal process tells you how the company hopes to operate. Exceptions show what actually keeps it running.

The four-step data fuel map

List the sources. Start with one business question and identify the relevant systems, documents and conversations. Do not move everything.

Score four dimensions. Consider commercial value, usability, risk and refresh rate. Highly sensitive data may wait; a modest but well-structured source may be ideal for the first test.

Define fields and evidence. “Customer dissatisfaction” is vague. Complaint theme, product, purchase date, first response and outcome can be analysed. Keep a route from each conclusion back to source evidence.

Connect it to action. Who sees the result? What will it trigger? How often does it refresh? Without an owner and a next step, even an attractive dashboard is wallpaper.

My test question is: “If this result changes tomorrow, who will do something differently?” If nobody can answer, the project is not yet connected to operations.

Hypothetical example: not a shortage, but the wrong replenishment level

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

A Hong Kong distributor repeatedly runs out of popular products. Management blames supplier lead-time variation and considers raising safety stock across the range. The team aligns sales, stock-outs, returns, campaign dates and purchase orders on one timeline, with AI classifying recurring combinations and anomalies.

Another explanation appears. During campaigns, demand concentrates in two pack sizes, while replenishment remains calculated at product-family level. Total stock looks adequate, but the exact variants likely to sell out are not flagged early.

The first response need not be a vast forecasting platform. Connect the promotion calendar, variant-level sales and supplier lead time, then give purchasing a daily exception list. The return is not a magical percentage; it is whether the company prevents expediting, misallocation and risks it previously could not see.

Turning analysis into institutional memory

Many AI projects end after one useful analysis. Three months later, a staff change means classification logic, judgement and exception handling must be explained again. The knowledge still belongs to a person, not the company.

At minimum, preserve a data dictionary, source and refresh time, decision rules and versions, plus human corrections and exceptions. Each correction should improve the next cycle rather than patch only today’s answer. Institutional memory does not mean storing more documents. It means knowing why the company reached a conclusion.

Privacy, bias and data quality

Data is fuel, but more is not automatically better. Stale, duplicated or selectively recorded information lets AI amplify an old problem with confidence. For personal data, define purpose, retention, access and deletion. If anonymised or aggregated information is sufficient, do not use full identity data.

Record what is unknown. A blank loss reason does not mean the customer had no objection. A system containing only successful cases does not mean the process always succeeds. AI can surface patterns in what is present; management must ask who and what is absent.