The Market Intelligence System:
turning external change into repeatable decisions

How should an organisation turn external change into repeatable decisions?

Market intelligence is not a report, dashboard or research department. It is an organisational system that detects consequential change, tests what it means, routes evidence to a named decision and learns from the result.

Most organisations do not suffer from a simple shortage of information. They suffer from weak conversion: signals are collected without a decision in view; internal data describes yesterday while external evidence lacks context; research arrives after commitment; forecasts hide assumptions; and insights have no owner. Generative AI can multiply this failure by producing faster, more persuasive synthesis whose provenance or validity is uncertain.

The effective response is a closed intelligence-to-decision loop:

Frame the decision → define what must be known → sense broadly → validate and triangulate → interpret commercially → decide → observe outcomes → update the system.

This report recommends that leaders build the system around a small portfolio of recurring decisions, not an unlimited stream of topics. Each intelligence product should identify the decision, owner, deadline, assumptions at risk, evidence strength, plausible alternatives and next action. AI should extend coverage and synthesis, while accountable people retain authority over framing, validation and consequential recommendations.

1. What market intelligence is—and is not

For this programme, market intelligence means the coordinated capability to understand customers, demand, competitors, ecosystems and external change well enough to improve a business decision.

It is broader than market research. Research is an important method for answering a defined question; intelligence connects multiple forms of evidence to decisions over time. It is also different from business intelligence, which usually emphasises internal performance data. A capable market intelligence system combines both:

• Outside-in evidence: customer behaviour, primary research, competitor moves, regulation, capital flows, channel change, supply conditions, technology adoption and weak signals.
• Inside-out evidence: sales patterns, win/loss evidence, pricing realisation, churn, service demand, product usage, pipeline quality and frontline observations.
• Interpretation: causal hypotheses, uncertainty, alternative explanations and economic implications.
• Decision activation: a named forum, owner, trigger and action.
• Learning: comparison of assumptions and forecasts with observed results.

The ICC/ESOMAR Code characterises research and data analytics as objective, fact-based work intended to illuminate attitudes, needs and behaviours, and it emphasises privacy, accountability, transparency and human oversight.¹ ISO 20252:2026 sets vocabulary and service requirements across market, opinion and social research, including insights and data analytics.² These are useful foundations for evidence quality, but an enterprise intelligence system must additionally ensure that evidence changes a decision.

2. Begin with the decisions

An intelligence system should start with a decision inventory. Leaders identify recurring, high-value or high-uncertainty choices—for example:

• which markets, categories or customer situations deserve investment;
• which demand signals justify a product, proposition or capacity change;
• when a competitor move requires response and when it does not;
• which customer segment should receive differentiated service or pricing;
• which assumptions must hold before entering a geography or channel;
• when a trend has moved from interesting to decision-relevant.

Each decision receives a Priority Intelligence Requirement (PIR). A useful PIR is specific enough to change an action: “What evidence would cause us to accelerate, delay or reject entry into market X?” is stronger than “What is happening in market X?”

Decision brief

Every material intelligence assignment should contain:

This discipline prevents an attractive research output from becoming an answer without a question.

3. The intelligence-to-decision architecture

3.1 Frame

Translate an executive concern into a decision, hypotheses and disconfirming questions. Separate what is known, assumed and unknown. State the cost of being wrong and whether the decision is reversible; this determines the required depth and speed.

3.2 Sense

Build a deliberately diverse evidence field. A global organisation needs comparable macro and category sources, but local context matters: customer language, channel structure, regulation and competitive behaviour can invalidate a global average. Hong Kong organisations can incorporate official recurring series such as retail sales, external trade, business receipts, the Quarterly Business Tendency Survey and SME business-situation surveys from the Census and Statistics Department.³

Signals should be tagged by subject, geography, sector, source type, date, direction, maturity and affected assumption. Horizon scanning is valuable only when it looks beyond familiar timeframes, sources and organisational culture. The UK Government Office for Science defines it as systematic collection of emerging trends and weak signals to identify threats, risks and opportunities.⁴

3.3 Validate and triangulate

No single source should carry a consequential conclusion when independent corroboration is reasonably available. Validation should examine:

• provenance and incentives of the source;
• definition, sample, coverage and reference period;
• whether the evidence measures stated preference, observed behaviour or an outcome;
• comparability across markets and periods;
• missing populations or survivorship effects;
• plausible alternative explanations;
• recency relative to the decision;
• whether AI transformed, inferred or generated any part of the evidence.

Triangulation is not majority voting. Several sources may repeat the same original claim or share the same bias. Analysts should trace claims to origin and ask whether different methods independently converge.

3.4 Interpret

Interpretation converts evidence into business meaning. A concise intelligence assessment should answer:

1. What changed? Evidence and reference period.
2. Why might it be changing? Tested causal hypotheses, including alternatives.
3. Why does it matter? Customer, competitive and economic consequence.
4. What remains uncertain? Confidence and missing evidence.
5. What should we do? Options, trade-offs, trigger and owner.

Facts, analytical judgments and recommendations must be labelled separately. Confidence should reflect evidence strength and agreement, not rhetorical certainty.

3.5 Decide and activate

Intelligence needs a destination. Establish recurring decision forums aligned to cadence: weekly signal triage, monthly commercial review, quarterly portfolio review and event-driven escalation for discontinuities. Each forum should close with a recorded decision, owner, assumptions and review trigger.

3.6 Learn

Maintain an assumption and decision ledger. Record what the organisation believed, what evidence supported it, the decision taken, expected indicators and the later outcome. This makes intelligence auditable and improves calibration. A forecast that was wrong for a well-specified reason can be more valuable than a vague forecast that appeared right.

4. Evidence portfolio: match method to question

An efficient system does not apply heavyweight research to every issue. It uses an evidence ladder:

• Continuous signals: official statistics, transactional and behavioural data, competitor and regulatory monitoring.
• Rapid diagnosis: expert interviews, frontline inquiry, search and social patterns, targeted desk research.
• Focused validation: customer interviews, concept tests, win/loss work, pricing or choice experiments.
• High-consequence confirmation: representative research, pilots, controlled tests, due diligence and independent challenge.

The required standard rises with consequence, irreversibility and uncertainty. Speed matters, but premature precision creates false confidence. ISO’s research standard is particularly relevant where work must be comparable across countries: ISO notes that common service requirements increase confidence in consistency across regions.⁵

5. AI-native, decision-led

An AI-native market intelligence system redesigns the workflow around machine-scale sensing and human accountability. It does not merely add a chatbot to an old research process.

Appropriate AI roles

• monitor and classify high-volume permitted sources;
• extract entities, themes, claims and changes;
• translate and compare multilingual material;
• retrieve relevant prior evidence and decisions;
• generate competing hypotheses and disconfirming questions;
• identify contradictions, gaps and stale assumptions;
• draft traceable summaries for analyst review.

Human-reserved responsibilities

• set the decision and acceptable risk;
• determine lawful and ethical use of data;
• judge source fitness, context and causality;
• commission primary evidence where observation is insufficient;
• challenge attractive but weak narratives;
• approve consequential conclusions and recommendations.

NIST’s AI Risk Management Framework organises risk work around Govern, Map, Measure and Manage and is intended to support trustworthy use throughout the AI lifecycle.⁶ Its Generative AI Profile highlights risks including confabulation, data privacy, information integrity and overreliance, and proposes actions matched to organisational priorities.⁷ For market intelligence, this implies source-level traceability, retrieval boundaries, evaluation against known cases, human review, access controls and incident handling.

The Hong Kong Privacy Commissioner’s AI Model Personal Data Protection Framework provides local guidance for governance, risk assessment, implementation and ongoing management where AI systems process personal data.⁸ Compliance requires case-specific legal review; the business principle is simpler: minimise personal data, state purpose, control access and retain a human line of accountability.

6. Global design, APAC and Hong Kong application

APAC is not one market. An effective system preserves comparability without erasing local structure.

Use a two-layer model

• Common core: shared definitions, taxonomy, evidence ratings, decision templates and economic measures.
• Local layer: language, cultural context, customer journey, channel economics, regulation, competitor set and data availability.

Avoid four regional errors

1. Global-average substitution: treating aggregated regional evidence as a local forecast.
2. Language flattening: translating words without interpreting context, idiom or category meaning.
3. Platform visibility bias: assuming accessible digital signals represent the whole population or buying system.
4. Market-boundary error: defining opportunity by administrative geography when customers, supply and competition operate across borders.

Hong Kong can function as a market, operating base, financial and professional-services hub, and gateway into wider networks; the relevant unit of analysis depends on the decision. Official statistics offer a dependable baseline, while customer and ecosystem evidence explains mechanisms that aggregate series cannot.³

7. Operating model and accountability

A lean system can work if responsibilities are explicit:

Centralisation creates consistency but can lose proximity; decentralisation creates relevance but can fragment methods. A hub-and-network model is usually stronger: a small central team owns standards and shared infrastructure, while embedded market and functional contributors own context and activation.

8. Intelligence products

Use a small product portfolio rather than bespoke formats for every request:

• Signal note: one material change, evidence, affected assumption and monitoring trigger.
• Decision brief: alternatives, economics, uncertainty and recommended action.
• Market landscape: customers, value pools, competitors, ecosystem and structural change.
• Scenario and stress test: multiple plausible conditions and option resilience. OECD’s foresight toolkit stresses that scenarios are not predictions and recommends challenging assumptions, constructing alternatives, stress-testing and action planning.⁹
• Assumption review: evidence for and against the beliefs supporting a strategy.
• Executive watchlist: only high-priority signals, owners and thresholds.

Every product should link to source records and the decision ledger.

9. Performance: measure decisions, not output volume

Counting reports, searches or dashboard visits rewards activity. A balanced scorecard should measure:

Use and timeliness

• share of priority decisions supported before commitment;
• median time from material signal to accountable review;
• proportion of intelligence products with a named owner and trigger.

Evidence quality

• traceable material claims;
• independent triangulation for high-consequence conclusions;
• freshness and geographic fitness;
• documented uncertainty and contrary evidence.

Decision impact

• investments accelerated, redesigned or stopped;
• avoidable loss or exposure reduced;
• forecast and confidence calibration;
• commercial outcomes versus documented assumptions.

Learning

• assumptions retired or revised;
• post-decision reviews completed;
• repeated errors and blind spots reduced.

Attribution will rarely be perfect. Use contribution evidence: did intelligence arrive in time, alter the option set, change confidence or define a better trigger?

10. Maturity model

Maturity is decision-specific. A firm can be advanced in customer analytics and weak in competitive foresight. Assessment should therefore examine priority decisions, not award one enterprise-wide label.

11. A practical 90-day launch

Days 1–30: focus

• nominate sponsor and intelligence lead;
• select five to ten consequential recurring decisions;
• define PIRs, current assumptions and review dates;
• inventory existing sources, reports, subscriptions and dashboards;
• agree evidence, privacy and AI-use rules;
• baseline current lead time, usage and decision pain points.

Days 31–60: pilot

• choose two decisions with different cadences;
• create source register, signal taxonomy and decision-brief template;
• connect relevant internal and external evidence;
• run human-reviewed AI assistance on a bounded source set;
• hold signal triage and decision review;
• record decisions, confidence, triggers and expected outcomes.

Days 61–90: institutionalise

• assess whether the pilot changed timing, options or confidence;
• remove low-value collection and automate safe repetitive work;
• formalise hub-and-network roles and escalation routes;
• establish quarterly assumption review and post-decision learning;
• publish an executive watchlist and next-quarter intelligence agenda.

The first success criterion is not a complete platform. It is evidence that one important decision became earlier, clearer or more robust.

Sources

1ICC and ESOMAR, *ICC/ESOMAR International Code on Market, Opinion and Social Research and Data Analytics* (2025 revision), especially the introduction and principles on objective evidence, accountability, transparency, privacy and human oversight. https://community.esomar.org/uploads/public/knowledge-and-standards/codes-and-guidelines/ICCESOMAR-International-Code_English.pdf

2International Organization for Standardization, *ISO 20252:2026 — Market, opinion and social research, including insights and data analytics — Vocabulary and service requirements* (2026). https://www.iso.org/standard/88881.html

3Census and Statistics Department, Hong Kong SAR, *Surveys Conducted by C&SD*, including monthly retail sales and SME business-situation surveys and the Quarterly Business Tendency Survey. https://www.censtatd.gov.hk/en/page_90.html

4UK Government Office for Science, *The Futures Toolkit* (updated 2024), Horizon Scanning section. https://www.gov.uk/government/publications/futures-toolkit-for-policy-makers-and-analysts/the-futures-toolkit-html

5ISO Technical Committee 225, *About ISO 20252*, on quality requirements across the research lifecycle and international comparability. https://committee.iso.org/home/tc225

6National Institute of Standards and Technology, *Artificial Intelligence Risk Management Framework (AI RMF 1.0)* (2023). https://www.nist.gov/itl/ai-risk-management-framework

7National Institute of Standards and Technology, *Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1* (2024). https://doi.org/10.6028/NIST.AI.600-1

8Office of the Privacy Commissioner for Personal Data, Hong Kong, *Artificial Intelligence: Model Personal Data Protection Framework* (2024). https://www.pcpd.org.hk/english/resources_centre/publications/files/ai_protection_framework.pdf

9OECD, *Strategic Foresight Toolkit for Resilient Public Policy* (2025). https://doi.org/10.1787/bcdd9304-en