Is your business ready for AI?

One concrete goal

Have we written down one specific AI goal for the next 3–6 months, with a named business owner who is accountable for the result (not only “IT will explore AI”)?

A good goal names the channel, the user, and the outcome. A weak goal is “become an AI company” or “use ChatGPT more.”

Example. Good: “Answer product and pricing questions on WhatsApp within 30 seconds after hours, then pass hot leads to sales.” Weak: “Deploy AI across the business.”

Tools and data we already use

Can we list the main systems and tools this AI goal would touch — including unofficial use of ChatGPT, Copilot, or similar tools by staff today?

You do not need a full architecture diagram. A simple inventory is enough: CRM, WhatsApp, website, Excel, vendor bots, and any personal AI accounts used for work.

Example. Example list: WhatsApp Business, HubSpot or Google Sheets leads, product catalogue PDF, staff using ChatGPT for draft replies.

The customer or staff journey

Can a non-technical manager explain, in a few steps, what happens from the moment a customer (or staff member) starts, through the AI response, to when a human takes over if needed?

If only the vendor or one developer can explain the journey, the business is not ready to own the outcome.

Example. Example: Customer taps Instagram ad → opens WhatsApp → AI answers FAQs → captures budget → sales gets a summary → human closes the deal.

Known data problems

For the records this AI goal depends on (customers, products, prices, tickets, inventory, etc.), do we already know whether data is often incomplete, duplicated, conflicting, or out of date?

Ask the people who live in the spreadsheets and CRM. If they say “we never trust that field,” treat that as a finding — you do not need a technical audit yet to answer Yes / Somewhat / Not yet.

Example. Common issues: two records for the same customer, missing mobile numbers, prices only correct in Excel, product names that do not match the website.

The people or cases AI will actually serve

Are we confident the data is good enough for the customers, branches, products, or languages this AI will serve — not only for a clean sample from one team, one outlet, or one product line?

AI trained or grounded on HQ data often fails for East Malaysia, a second brand, or Bahasa Malaysia enquiries. Think about who will use it on day one.

Example. If the bot will serve all Malaysia retail enquiries, but clean data only exists for KL flagship SKUs, answer Not yet or Somewhat.

Approved content the AI may use

Do we have up-to-date, approved materials the AI should answer from — such as FAQs, price lists, product sheets, policies, or service descriptions — and do we know the major topics that are missing or outdated?

Approved means someone in the business signed off. Random SharePoint folders and old brochures do not count unless you choose them on purpose.

Example. Ready: current fee schedule + FAQ signed off by sales. Not ready: “everything is in WhatsApp history and someone’s laptop.”

Official knowledge vs live chat vs stored history

Do business owners understand three different things: (1) official knowledge the AI is allowed to use, (2) what happens in the live conversation with a customer, and (3) what gets saved afterwards (leads, transcripts, logs)?

Mixing these up causes problems: old prices in “knowledge,” promises made only in chat, or personal data kept longer than intended.

Example. Official: product FAQ PDF. Live: customer asks “berapa harga?” Saved: name, number, and transcript in CRM or a sheet for follow-up.

What the AI must not do alone

Have we written a clear list of topics or decisions the AI must not handle without a human — for example special discounts, legal or medical claims, complaints, refunds, or anything that could commit the company?

This protects brand and customers. Without it, teams either over-block the AI or let it invent promises.

Example. “AI may share published price ranges. AI must escalate: custom quotes, complaints, PDPA requests, and anything not in the approved FAQ.”

Same meaning across teams

Do sales, operations, finance, and IT agree — in writing — on the meaning of the main terms this AI will use (for example “customer”, “active”, “lead”, “open order”), and is there a named owner for each definition?

If two teams argue about the dashboard today, the AI will argue with your customers tomorrow.

Example. “Active customer” = purchased in last 12 months (Marketing) vs has an open contract (Finance). Pick one definition for this AI goal.

Status codes and messy notes

Do we understand the status codes, product categories, and free-text notes people use day to day well enough that an AI would not misread them (for example “pending” meaning three different things, or remarks only one staff member understands)?

Free-text fields and tribal abbreviations are a common reason AI answers look confident but wrong.

Example. CRM status “OK” used for paid, delivered, and “do not call.” Product code “STD” means different packages in two brands.

Where personal data sits in this journey

Can we point to where personal data appears for this AI goal — for example names, phone numbers, IC numbers, chat transcripts, CRM fields, spreadsheets, or vendor AI tools — including copies people keep informally?

You are not writing a full PDPA file yet. You are checking whether the business can see the footprint before AI multiplies it.

Example. Customer WhatsApp number → AI transcript → Google Sheet → sales phone → occasional paste into ChatGPT. Map each hop.

What is allowed under PDPA and company rules

Do we have clear rules (even a one-page policy) on what personal data may be used for this AI, what may be sent to third-party tools, how long chats or logs are kept, and who approves exceptions?

If the answer is “we will figure it out after go-live,” answer Not yet. PDPA issues are harder to unwind once data is in prompts and vendor systems.

Example. Rule example: “No IC numbers in AI prompts. Chat logs kept 90 days. Only approved vendor accounts. Export to personal Gmail forbidden.”

Top blockers written down

Have we listed the top data problems that would block or embarrass this AI project if we started tomorrow, with a rough sense of effort and a named owner for each item?

A backlog of five real blockers beats a vague “data quality initiative.” Focus on what this AI goal needs, not the whole enterprise.

Example. 1) Deduplicate customer mobiles (Owner: Ops, 2 weeks). 2) Publish one price list (Owner: Sales). 3) Fix product names vs website (Owner: Marketing).

Hidden cleanup knowledge

When staff already clean or “fix” data in Excel, WhatsApp forwards, or personal scripts before work can proceed, is that know-how written down so the AI project does not depend on one person’s memory?

If the business only works because someone cleans a file every Monday, the AI will inherit that Monday ritual — or fail without it.

Example. “Every week Ah Chong removes test accounts and fixes postcodes before the report.” That step must be documented or automated before AI scale.

Which systems feed this goal

Can we name the main systems that supply data or conversations for this AI goal (CRM, ERP, website, WhatsApp, forms, spreadsheets), and do we know which ones break, lag, or disagree most often?

This is about trust under pressure: when the website price and the ERP price differ, which one should the AI use?

Example. WhatsApp ↔ catalogue PDF ↔ ERP stock. Stock is often wrong on weekends. Decide the source of truth before go-live.

When humans take over

Do front-line staff know exactly when they must take over from the AI, and do they receive enough context (what the customer asked, what the AI said, key details captured) to continue without starting from zero?

Handoff without context destroys trust. Customers hate repeating themselves after talking to a bot.

Example. Escalate if customer asks for a discount, is angry, or AI confidence is low — and open the chat with the full transcript for the agent.