AI Playbook

The domestic market is small, so the first instinct is to ask whether this travels. Expect to be benchmarked against a regional peer before you finish the summary.

• Who else in the region already runs this, and what did it cost them to get to production?
• Does this deployment travel to at least two neighbouring markets, or is it Singapore-only?
• Which of PDPA, the IMDA framework or MAS FEAT applies here, and who confirmed that?
• What is the reference customer we can actually name in a board paper?

Named out loud here: PDPC (PDPA) · IMDA Model AI Governance Framework · AI Verify · MAS FEAT principles

A visible failure in front of a customer costs more than a year of delay. Objections often arrive as polite questions and are sharper than they sound.

• Which departments have already agreed to this, and when were they consulted?
• What happens to the people whose work this changes?
• What is the measured defect rate from the first narrow deployment?
• How does this handle personal data under APPI, and who signed that off?

Named out loud here: APPI · METI/MIC AI Guidelines for Business · sector ministry guidance

Speed is read as competence and delay as weakness. Expect a data-residency question early and a competitor comparison immediately after.

• Why is this not in production already?
• Can the whole thing run in-country, including the model?
• Which models are permissible here, and has the filing been done?
• What is a competitor shipping in the next six weeks that makes this irrelevant?

Named out loud here: PIPL · Data Security Law · CAC generative AI measures and algorithm filing

The room is comfortable arguing numbers in public. Unit economics at national scale will be tested harder than the technology.

• What does this cost per transaction at ten million transactions, not ten thousand?
• How many languages does it work in, and how well in each?
• What breaks first when volume goes up by two orders of magnitude?
• Which DPDP obligations attach to the data this touches?

Named out loud here: DPDP Act 2023 · sector regulators such as RBI and IRDAI

Courteous conversation, non-negotiable operational question. Assume mobile-first and assume the senior engineering bench is thin.

• Who operates this locally once the launch team goes home?
• Does it work properly on a mid-range phone over a patchy connection?
• Where does the data physically sit?
• What is the staffing plan that does not assume scarce senior ML engineers?

Named out loud here: UU PDP (Personal Data Protection Law) · Kominfo electronic system rules · OJK for financial services

Moving second is treated as losing, but a privacy incident is a front-page event with personal liability attached.

• What personal data touches this, and whose name is on the filing?
• What is the launch date, and what are you cutting to hit it?
• Who is the accountable executive by name?
• How fast can a competitor copy this once it is visible?

Named out loud here: PIPA · AI Framework Act · Financial Services Commission for financial services

Jargon is treated as evasion. The workforce conversation will happen — the only question is whether it happens in the boardroom or the media.

• How would you explain this to the person whose job it changes, in words they would accept?
• Can you say what this does in plain English, without the word "leverage"?
• What is the workforce impact statement, and was it written before or after the build?
• What would a customer see if they asked how a decision was made?

Named out loud here: Privacy Act / OAIC · Voluntary AI Safety Standard · APRA CPS 230 for financial services

Model risk and explainability. A regulator can ask for the reasoning behind a decision years after it was made.

• If a customer disputes a decision this touched, can you reconstruct why — two years later?
• Does the human decision-maker stay accountable, or does the model decide?
• Where is the audit trail, and how long is it retained?
• What is the false positive rate, and who absorbs the cost of those?

Where value has landed: Document-heavy back office: underwriting, claims, KYC refresh, advertising and marketing compliance review. High volume, clearly bounded, human stays in the decision.

The physical world does not tolerate a hallucination, and the data you need often sits behind an OT boundary that exists for good reasons.

• Does this touch anything that moves? If not, why is the value still real?
• Which side of the OT boundary does the data live on?
• What is the measured defect rate, not the demo accuracy?
• What happens to a warranty claim if the system was wrong?

Where value has landed: Engineering knowledge retrieval, standards and regulation lookup, maintenance and warranty analysis, dealer and aftersales experience. The wins are in the documents, not the machines.

A citizen-facing error becomes a parliamentary question, and procurement moves slower than the technology.

• What is the data classification, and has this cleared it — or are we assuming it will?
• Is this officer-facing or citizen-facing, and why that order?
• What does the failure look like on the front page of a newspaper?
• How does this get procured, and how long does that actually take?

Where value has landed: Officer productivity first — case notes, policy search, contact-centre assist. Citizen-facing triage second. In education, educator support rather than autonomous grading.

Volume is enormous and margin per interaction is thin, so unit cost decides whether this ships at all.

• At real traffic, what does one interaction cost?
• Who owns what the model generates, and has legal confirmed it?
• What is the deflection rate, and what happens to the calls it fails?
• How does this behave when a public figure or a minor is involved?

Where value has landed: Contact-centre deflection and agent assist, network operations triage, content production and localisation, subscriber personalisation.

Patient safety and clinical liability. Tolerance for a confident error is close to zero, and liability attaches to a named clinician.

• Is this a medical device in any market we operate in, and who made that determination?
• Does it influence a diagnosis or treatment decision, even indirectly?
• Which clinician signed off on the scope?
• What is the escalation path when it is wrong and nobody notices for a month?

Where value has landed: Clinical documentation and coding, prior authorisation and claims, trial document workflows, patient administration. Removing paperwork, not making clinical calls.

Brand voice and customer trust. One screenshot of the model saying something wrong travels further than any launch campaign.

• What does this say on its worst day, and how quickly would we find out?
• Who approves the brand voice, and can they veto a release?
• What cohort did this run against before full rollout?
• How do we substantiate any claim the model makes about a product?

Where value has landed: Product discovery and search, catalogue and content generation, customer support assist, merchandising. Generated content at scale is usually where measurable lift shows first.

The work being automated is often the billable work, so the business model question arrives before the technical one.

• If this makes delivery faster, what happens to what we bill?
• Does client data leave our tenancy at any point?
• Who owns the output — us, the client, or the vendor?
• What do we tell a client who asks whether AI touched their deliverable?

Where value has landed: Engineering productivity, pre-sales and proposal generation, contract and document review, internal knowledge retrieval.