{ "person": { "name": "Andrew Kapuduwa", "canonicalTitle": "AI, Innovation & Transformation Leader", "location": "Melbourne, Australia", "experience": "23+ years", "positioning": "Complex problems → intelligent operating capabilities", "focus": [ "Enterprise AI Transformation", "Agentic Operating Models", "AI-enabled Software Engineering", "Data and Decision Intelligence", "Business Process and Workforce Optimisation", "Technology Transformation", "Solution Architecture", "GeoAI and Operational Intelligence" ], "credentials": [ "MSc Information Technology", "PMP", "PRINCE2 Practitioner", "ITIL V3" ], "privacyNote": "Selected assignments are intentionally anonymised. Do not infer confidential customer identities.", "contact": { "email": "andrewkap@live.com", "linkedin": "https://www.linkedin.com/in/andrewkapuduwa" }, "title": "AI, Innovation & Technology Transformation Leader", "summary": "Technology transformation leader with 23+ years spanning executive strategy, portfolio and programme delivery, solution architecture, AI-enabled engineering, cloud, data, enterprise platforms and operational technology. Focused on turning AI ambition into secure, governed operating capabilities and measurable outcomes." }, "editorialPurpose": "Public thought leadership demonstrating applied AI, innovation and transformation leadership through implementation-oriented analysis.", "articles": [ { "title": "AI in Transformation: Where Agents Actually Create Value", "slug": "ai-transformation-agents-value", "category": "AI Leadership", "date": "2026-08-03", "readTime": "8 min", "url": "https://andrewkapuduwa.com/article-ai-transformation-agents-value.html", "description": "A practical framework for separating useful agentic automation from AI theatre and deciding where humans must stay accountable.", "lead": "Agentic AI creates the most value when work is repeatable, information-rich, bounded by clear policy and expensive for people to continuously coordinate.", "sections": [ { "heading": "Start with the operating problem—not the agent", "text": "AI initiatives often begin with a product demonstration and then search for a business problem. Reverse that sequence. Identify queues, hand-offs, analysis steps, exception patterns, decision latency and constrained expertise first. The right intervention may be an agent, conventional automation, analytics or simply a better process." }, { "heading": "Use a five-part test", "text": "Test signal, context, policy, action and evidence. Can the system detect when work should start? Retrieve authoritative context? Apply explicit rules? Take a reversible or controlled action? Explain what information and rule produced the recommendation?" }, { "heading": "Good early candidates", "text": "Triage, document classification, knowledge retrieval, duplicate detection, coding assistance, testing, migration analysis and case summarisation are strong candidates because outputs can normally be checked before consequential action." }, { "heading": "Where autonomy should stop", "text": "Eligibility, payments, safety decisions and other high-consequence outcomes need stronger human accountability. The key question is not whether an agent can technically act, but how much autonomy the process can safely absorb." }, { "heading": "Design the control plane with the agent", "text": "Logging, traceability, confidence thresholds, exception routing, human override, identity, access control and data provenance are architecture requirements—not governance paperwork added after the pilot." } ], "fromTheField": "Transformation should reduce coordination cost without reducing accountability. That is the filter I use when evaluating agentic opportunities.", "image": "assets/articles/ai-transformation-agents-value.png" }, { "title": "Why Transformation Estimates Fail Before Delivery Starts", "slug": "why-transformation-estimates-fail", "category": "Transformation", "date": "2026-08-05", "readTime": "8 min", "url": "https://andrewkapuduwa.com/article-why-transformation-estimates-fail.html", "description": "Why reuse assumptions, incomplete elaboration, data-model debt and process redesign can make early transformation estimates unreliable.", "lead": "The estimate is often blamed when a programme costs more than expected. In many cases the real failure happened earlier: the organisation estimated a problem it did not yet understand.", "sections": [ { "heading": "The reuse assumption", "text": "Funding models frequently assume existing platforms, workflows, APIs or data structures can be reused. Later elaboration can reveal obsolete processes, architectural debt, security gaps, tight coupling or so much required change that reuse becomes a rebuild with extra steps." }, { "heading": "Estimates need confidence levels", "text": "A concept estimate and a delivery-ready estimate are not the same artefact. Treating both as equally certain creates false precision. Associate planning ranges with explicit confidence and name the assumptions that can move them." }, { "heading": "Process improvement changes the technology requirement", "text": "Transformation should not faithfully reproduce a broken process. But once the business process is redesigned, the system being estimated may be materially different from the funding case. That is legitimate evolution and should be governed explicitly." }, { "heading": "Data-model debt multiplies cost", "text": "Weak data structures make workflow, reporting, security, integration and migration harder. Fixing the model can raise the immediate estimate while reducing long-term complexity and rework." }, { "heading": "Re-estimation is not automatically failure", "text": "Classify estimate movement into original estimation error, incorrect assumptions, newly discovered scope and conscious solution improvement. That creates a much better executive conversation than simply declaring the programme over budget." } ], "fromTheField": "The executive question should be: what did we know then, what do we know now, and what changed between those points?", "image": "assets/articles/why-transformation-estimates-fail.png" }, { "title": "Designing AI Agents for Dynamics 365 Case Management", "slug": "ai-agents-dynamics-365-case-management", "category": "AI + Business Process", "date": "2026-08-07", "readTime": "9 min", "url": "https://andrewkapuduwa.com/article-ai-agents-dynamics-365-case-management.html", "description": "How AI agents can support triage, workload allocation, duplicate detection, assessment preparation and case summarisation in Dynamics 365.", "lead": "Case-management environments contain large amounts of coordination work around a smaller number of consequential decisions. That makes them strong candidates for AI augmentation.", "sections": [ { "heading": "Start with triage", "text": "Applications may contain structured fields, documents, free text, addresses and multiple people. An agent can classify the case, identify missing information, summarise evidence, flag possible duplicates and recommend a queue without making the final eligibility decision." }, { "heading": "Make workload allocation intelligent", "text": "Instead of only round-robin or manual allocation, an agent can consider caseload, complexity, officer skill, service targets, priority, absence and specialisation. Supervisors retain override authority." }, { "heading": "Detect relationships, not only exact duplicates", "text": "Families, addresses, phone numbers, bank accounts and dependants create relationship patterns simple rules can miss. AI-assisted matching should return evidence explaining why records appear related rather than a black-box flag." }, { "heading": "Prepare the assessment, do not hide the judgement", "text": "An agent can assemble household composition, identity status, address eligibility, prior applications, documents, missing information and applicable rules. The assessor spends less time searching and more time judging." }, { "heading": "Keep Dynamics as the system of record", "text": "Agents should operate through controlled APIs and approved actions. Case status, customers, assessments, decisions, tasks, queues and audit history remain in the authoritative platform." } ], "fromTheField": "Automate the coordination surrounding human judgement before attempting to automate the judgement itself.", "image": "assets/articles/ai-agents-dynamics-365-case-management.png" }, { "title": "When Should an AI Agent Be Allowed to Make a Decision?", "slug": "when-should-ai-agent-make-decisions", "category": "Responsible AI", "date": "2026-08-09", "readTime": "7 min", "url": "https://andrewkapuduwa.com/article-when-should-ai-agent-make-decisions.html", "description": "A practical authority model for deciding how much autonomy an AI agent should receive.", "lead": "Once an AI agent can retrieve enterprise data, reason about a situation and call tools, technical capability becomes an authority-design problem.", "sections": [ { "heading": "Autonomy is a spectrum", "text": "Separate observe, analyse, recommend, prepare, execute with approval and execute autonomously. These are different risk profiles and should not be grouped under one label such as automation." }, { "heading": "Test consequence and reversibility", "text": "A wrong internal queue recommendation is recoverable. A wrong payment or safety decision may not be. Autonomy should rise as consequence falls and reversibility increases." }, { "heading": "Test rule clarity and evidence quality", "text": "Clearly defined policy and trusted source data make bounded autonomy more viable. Ambiguous rules, weak data and hidden assumptions argue for stronger human review." }, { "heading": "Demand explainability", "text": "For consequential processes, reconstruct what information was used, which rule or model influenced the decision, what confidence existed and who or what authorised the action." }, { "heading": "Earn trust progressively", "text": "Start with recommendations, then human-approved execution, then limited autonomous action and exception-based supervision. Each increase in authority should be evidence-based." } ], "fromTheField": "The target is not maximum autonomy. It is maximum useful autonomy proportionate to the consequence of being wrong.", "image": "assets/articles/when-should-ai-agent-make-decisions.png" }, { "title": "Why Legacy Modernisation Is Really a Business-Rules Problem", "slug": "legacy-modernisation-business-rules", "category": "AI Engineering", "date": "2026-08-11", "readTime": "8 min", "url": "https://andrewkapuduwa.com/article-legacy-modernisation-business-rules.html", "description": "Why successful legacy modernisation requires discovering the institutional knowledge encoded in old systems, not merely translating code.", "lead": "The hardest legacy systems are difficult not only because the technology is old, but because decades of business decisions became encoded into software faster than the organisation documented them elsewhere.", "sections": [ { "heading": "Old code is organisational memory", "text": "Production systems accumulate regulatory rules, exceptions, workarounds and operational knowledge. Some logic is legitimate, some obsolete and some accidental complexity. The modernisation team has to distinguish them." }, { "heading": "Translation is not transformation", "text": "Converting a legacy language into a modern one can produce a modern-language copy of an old design. The valuable objective is to understand behaviour, decide what still matters and design the target capability deliberately." }, { "heading": "Create a behaviour catalogue", "text": "For meaningful logic, record the rule, source, inputs, outputs, exceptions, owner, current implementation, target decision and test evidence. This becomes the bridge between old code and target architecture." }, { "heading": "Use AI for code archaeology", "text": "Generative AI can accelerate summaries, dependency discovery, probable business-rule extraction, candidate tests, documentation and translation prototypes. Human SMEs and engineers still decide what the behaviour means." }, { "heading": "Characterisation tests become evidence", "text": "Capture representative legacy inputs and outputs before replacement. First understand what the system does; then deliberately preserve, change or remove that behaviour." } ], "fromTheField": "The modernisation question is: what does this system know that the organisation still needs to know?", "image": "assets/articles/legacy-modernisation-business-rules.png" }, { "title": "The Architecture–Delivery Gap Is a Leadership Problem", "slug": "architecture-delivery-gap", "category": "Transformation", "date": "2026-08-13", "readTime": "7 min", "url": "https://andrewkapuduwa.com/article-architecture-delivery-gap.html", "description": "Why architecture and delivery must operate as one decision system during technology transformation.", "lead": "Some of the most expensive transformation problems do not belong entirely to architecture or delivery. They exist between the two.", "sections": [ { "heading": "Architecture decisions are delivery decisions", "text": "A decision to introduce an integration layer, redesign the data model or change identity architecture alters sequencing, environments, testing, operations, dependencies, cost and risk. It cannot remain isolated in an architecture forum." }, { "heading": "Delivery constraints shape architecture", "text": "An ideal target state may assume all legacy systems can change together, data can migrate in one cutover or dependent teams are available. Good architecture needs a transition path that respects reality without becoming trapped by it." }, { "heading": "Programme leaders need architecture literacy", "text": "They do not need to be the solution architect, but they should recognise when a technical choice materially changes scope, risk, sequencing or benefits." }, { "heading": "Architects need delivery literacy", "text": "Diagrams must answer transition questions: what changes first, what coexists, how data moves, what happens when a dependency is late and who supports the intermediate state." }, { "heading": "Translation is the differentiator", "text": "Strong technology leaders explain architectural concerns as business risks, funding constraints as trade-offs, dependencies as sequencing and delays as benefit impacts." } ], "fromTheField": "Strategy, architecture and delivery should be treated as parts of the same decision system, not as sequential departments.", "image": "assets/articles/architecture-delivery-gap.png" }, { "title": "Using Geospatial Eligibility in Emergency Assistance Systems", "slug": "geospatial-eligibility-emergency-assistance", "category": "GeoAI + Emergency", "date": "2026-08-15", "readTime": "8 min", "url": "https://andrewkapuduwa.com/article-geospatial-eligibility-emergency-assistance.html", "description": "How polygon-based geospatial eligibility can improve emergency assistance assessment, auditability and automation.", "lead": "When emergency assistance depends on whether an applicant was inside an impacted area, geography should become a machine-queryable business rule rather than a map someone visually inspects.", "sections": [ { "heading": "Treat the impact area as event data", "text": "Store authoritative polygons, source identifiers and effective dates against the emergency event. This preserves which geography was used for each assessment and supports multiple impact areas per event." }, { "heading": "Turn addresses into spatial evidence", "text": "Geocode the applicant address into a point and evaluate whether that point falls inside the applicable polygon. This is more precise than suburb or postcode eligibility when emergency boundaries cut across administrative areas." }, { "heading": "Integrate emergency feeds safely", "text": "Retrieve incident feeds, identify records with useful geometry, preserve identifiers and convert coordinates into the geometry format the target platform uses. Coordinate order and projection need deterministic validation." }, { "heading": "Make the result explainable", "text": "Store the assessed address, geocode confidence, coordinates, polygon identifier, spatial result and timestamp so an officer or reviewer can reconstruct the basis of the result." }, { "heading": "Handle uncertainty deliberately", "text": "Low-confidence addresses, rural properties and ambiguous locations should route to manual review. Automation works best when clear cases flow through and uncertainty is surfaced rather than hidden." } ], "fromTheField": "The transformation is not putting a map inside CRM. It is turning location into a first-class business rule.", "image": "assets/articles/geospatial-eligibility-emergency-assistance.png" }, { "title": "Four Conversations a Technology Transformation Leader Must Hold at Once", "slug": "four-conversations-transformation-leader", "category": "Leadership", "date": "2026-08-17", "readTime": "6 min", "url": "https://andrewkapuduwa.com/article-four-conversations-transformation-leader.html", "description": "Why transformation leaders must continuously connect executive, business, architecture and delivery conversations.", "lead": "Transformation fails surprisingly often because everybody is discussing the same programme in different languages and nobody is accountable for preserving meaning between them.", "sections": [ { "heading": "The executive conversation", "text": "This is about outcomes, investment, risk, timing and organisational choices. It asks why the transformation matters and what value justifies the disruption." }, { "heading": "The business conversation", "text": "This is about how work should operate: customers, policies, decisions, exceptions, roles, measures and where friction currently exists." }, { "heading": "The architecture conversation", "text": "This translates the operating model into data, applications, AI, integration, identity, security and non-functional capability." }, { "heading": "The delivery conversation", "text": "This converts capability into sequencing, teams, dependencies, releases, tests, adoption and operational readiness." }, { "heading": "Leadership is preserving intent across boundaries", "text": "The job is not to personally perform every discipline. It is to recognise when a decision in one conversation changes assumptions in the others and force the connection before the cost appears later." } ], "fromTheField": "I see transformation leadership as translation at scale: preserving business intent as it moves through architecture and delivery into operational reality.", "image": "assets/articles/four-conversations-transformation-leader.png" }, { "title": "What Executives Should Ask Before Funding an AI Programme", "slug": "executives-before-funding-ai-programme", "category": "AI Leadership", "date": "2026-08-19", "readTime": "7 min", "url": "https://andrewkapuduwa.com/article-executives-before-funding-ai-programme.html", "description": "The questions executives should ask before funding an enterprise AI programme or agentic transformation.", "lead": "A compelling AI demonstration is not yet an investment case. Executives need clarity on the business constraint, data, authority, risk, operating cost and path to measurable value.", "sections": [ { "heading": "What business constraint are we removing?", "text": "Define the economic or operational friction first: decision latency, constrained expertise, repetitive analysis, customer wait time, engineering throughput, coordination cost or another measurable problem." }, { "heading": "Why AI rather than conventional automation?", "text": "If a deterministic rule or workflow can solve the problem more cheaply and reliably, use it. AI earns its place where context, ambiguity, language, pattern recognition or adaptive reasoning materially matters." }, { "heading": "What information will the AI use?", "text": "Identify authoritative sources, data quality, privacy constraints, context freshness and access controls. Weak data plus a powerful model can produce convincing failure." }, { "heading": "What authority will it receive?", "text": "Define observe, recommend, prepare and execute permissions separately. Ask what happens when it is wrong and whether the action can be reversed." }, { "heading": "How will value and reliability be measured?", "text": "Agree outcome measures, quality thresholds, exception rates, operating cost and adoption measures before the pilot. A successful demo is not the same as a successful operating capability." } ], "fromTheField": "The investment case should fund an operating capability, not just access to a model.", "image": "assets/articles/executives-before-funding-ai-programme.png" }, { "title": "AI for Software Teams: Don’t Just Add Copilot—Redesign the SDLC", "slug": "ai-first-software-engineering-teams", "category": "AI Engineering", "date": "2026-08-21", "readTime": "9 min", "url": "https://andrewkapuduwa.com/article-ai-first-software-engineering-teams.html", "description": "How software teams can redesign requirements, development, testing, review and DevOps around AI rather than simply adding coding assistants.", "lead": "The first phase of AI in engineering was personal productivity. The larger opportunity is to redesign the software-development lifecycle itself around humans, agents and deterministic controls.", "sections": [ { "heading": "The SDLC becomes a collaboration system", "text": "Requirements, architecture context, code generation, review, testing, documentation and operations can all be augmented. The value comes from improving flow across the lifecycle rather than maximising output at one developer workstation." }, { "heading": "Context becomes engineering infrastructure", "text": "Agents need architecture principles, coding standards, API contracts, data rules, test expectations, repository structure and business terminology. Teams that make this context machine-readable will gain more from AI than teams with undocumented systems." }, { "heading": "Combine LLM reasoning with deterministic evidence", "text": "Static analysis, unit tests, integration tests, vulnerability scanning and policy checks remain essential. AI can reason about intent and context, while deterministic controls provide evidence that specific engineering conditions are satisfied." }, { "heading": "Change the measures", "text": "Lines of AI-generated code are not a useful productivity metric. Measure lead time to production value, review time, rework, escaped defects, test quality, onboarding speed and the size of software estate a team can safely manage." }, { "heading": "Judgement becomes more valuable", "text": "As routine generation gets cheaper, engineering judgement shifts toward architecture, problem framing, validation, security and deciding whether a generated solution is appropriate." } ], "fromTheField": "The competitive advantage is not having the best autocomplete. It is turning AI into a governed engineering capability across the whole lifecycle.", "image": "assets/articles/ai-first-software-engineering-teams.png" }, { "title": "Can LLMs Write Business Requirements and User Stories? Yes—but That Is the Wrong Question", "slug": "llm-business-requirements-user-stories", "category": "Business Analysis", "date": "2026-08-23", "readTime": "8 min", "url": "https://andrewkapuduwa.com/article-llm-business-requirements-user-stories.html", "description": "How LLMs can improve business analysis by interrogating evidence, surfacing gaps and preserving traceability instead of simply generating user stories.", "lead": "An LLM can produce a grammatically perfect user story in seconds. That does not mean it has discovered the right requirement.", "sections": [ { "heading": "Start with evidence, not prompting", "text": "Feed approved source material such as process maps, policies, workshop notes, data definitions, current-system behaviour and known pain points. Ask the model to separate explicit requirements, inferred requirements, assumptions and unresolved questions." }, { "heading": "Use AI to challenge ambiguity", "text": "If a workshop says urgent work goes to the least busy officer, the useful AI output is a set of questions: what defines urgent, how workload is measured, whether complexity matters, what overrides exist and what happens when capacity is exhausted." }, { "heading": "Do not let user stories replace analysis", "text": "Complex capabilities may need business rules, decision tables, process models, state transitions, data definitions, integration requirements, non-functional requirements and user stories. AI should help maintain consistency across them." }, { "heading": "Preserve traceability", "text": "A strong AI-assisted requirement should retain its source, assumptions, related process, business rule and candidate acceptance evidence. As generation gets cheaper, traceability becomes more important." }, { "heading": "Keep human validation", "text": "AI can accelerate discovery, challenge, structure and verification. Business SMEs and accountable analysts still validate that the artefacts represent the actual need." } ], "fromTheField": "The mature pattern is not “AI writes our stories.” It is “AI helps us interrogate evidence, find gaps and convert validated knowledge into delivery artefacts.”", "image": "assets/articles/llm-business-requirements-user-stories.png" }, { "title": "AI-Augmented Business and Functional Analysis: The Analyst Becomes an Orchestrator of Evidence", "slug": "ai-business-functional-analysis", "category": "Business Analysis", "date": "2026-08-25", "readTime": "7 min", "url": "https://andrewkapuduwa.com/article-ai-business-functional-analysis.html", "description": "How AI can connect process, rules, data, requirements, architecture and tests while freeing analysts to focus on problem framing and value.", "lead": "Business analysis contains a large amount of information processing. AI can reduce that administration, but the highest-value outcome is giving analysts more time to analyse.", "sections": [ { "heading": "Build a connected analysis model", "text": "Connect objectives, stakeholders, processes, rules, requirements, data entities, systems, interfaces, risks, decisions and tests. AI can then answer cross-cutting questions that are difficult to maintain manually." }, { "heading": "Interrogate functional requirements systematically", "text": "For an API requirement, AI can ask about authentication, contracts, error handling, retries, volume, privacy, logging, performance and operational support. The analyst resolves the questions rather than discovering them during testing." }, { "heading": "Combine workshops with operational evidence", "text": "Process histories, workflow timestamps and logs show what actually happens. AI can compare operational patterns with what stakeholders say happens and surface where the two diverge." }, { "heading": "Use AI for change impact", "text": "When a business rule changes, identify which process, requirement, interface, story and test may be affected. This is where a connected knowledge model becomes more valuable than isolated documents." }, { "heading": "Move analysts toward strategic work", "text": "Reduce time spent transcribing, formatting and maintaining repetitive artefacts. Increase time spent on problem framing, operating-model design, stakeholder negotiation and value." } ], "fromTheField": "AI should eliminate analysis administration, not analytical thinking.", "image": "assets/articles/ai-business-functional-analysis.png" }, { "title": "Data Before AI: Building Operational Intelligence for Emergency Services", "slug": "data-ai-emergency-services-operational-intelligence", "category": "Emergency + Data", "date": "2026-08-27", "readTime": "8 min", "url": "https://andrewkapuduwa.com/article-data-ai-emergency-services-operational-intelligence.html", "description": "How emergency services can combine historical, real-time, contextual and geospatial data with AI to improve operational decision-making.", "lead": "Emergency services rarely have a complete absence of data. The problem is turning fragmented, fast-moving information into operational intelligence quickly enough to change the outcome.", "sections": [ { "heading": "Begin with operational questions", "text": "Ask where demand will exceed capacity, which locations repeatedly create delays, which stations experience simultaneous demand, where response time deteriorates and which operating conditions explain the pattern." }, { "heading": "Combine historical, real-time and contextual data", "text": "Historical incidents reveal patterns. Current unit status shows capacity. Traffic, weather, major events, road closures and hospital conditions explain why today may differ from the historical average." }, { "heading": "Create a trusted data foundation", "text": "Common identifiers, accurate timestamps, geospatial standards, quality measures, lineage, security and operational ownership are prerequisites. AI can make weak data look more convincing; that increases rather than reduces risk." }, { "heading": "Move from dashboards to decision intelligence", "text": "Descriptive reporting says response time increased. Decision intelligence isolates where, when and under which conditions the increase occurs, then identifies the operational lever that may change it." }, { "heading": "Keep humans in operational command", "text": "Forecasting and recommendations should explain evidence and uncertainty. The technology should expand situational awareness, not hide operational judgement inside a model." } ], "fromTheField": "The real value is not predicting an interesting number. It is finding the operational lever behind the number.", "image": "assets/articles/data-ai-emergency-services-operational-intelligence.png" }, { "title": "Case Study: Designing a Multi-Agency Emergency Services Intelligence Platform", "slug": "case-study-emergency-services-ai-data-platform", "category": "Emergency + AI", "date": "2026-08-29", "readTime": "10 min", "url": "https://andrewkapuduwa.com/article-case-study-emergency-services-ai-data-platform.html", "description": "An illustrative architecture combining CAD, traffic, weather, GIS, analytics and AI decision support across police, fire and ambulance operations.", "lead": "A multi-agency intelligence platform should not replace computer-aided dispatch. It should create an intelligence layer around CAD that continuously explains demand, coverage, context and emerging operational risk.", "sections": [ { "heading": "The reference architecture", "text": "Police, fire and ambulance CAD feeds combine with traffic, weather, GIS, events and sensor data in an operational data platform. Analytics and GeoAI produce forecasts and spatial features; AI decision support turns them into evidence-backed operational recommendations." }, { "heading": "Forecast demand by area and time", "text": "Use incident history, time, event type, severity, weather, major events and current capacity to estimate elevated demand by geography. The objective is resource planning, not predicting specific people or individual criminal behaviour." }, { "heading": "Optimise dynamic coverage", "text": "Every dispatch changes coverage elsewhere. Continuously model unit locations, travel times, forecast demand, specialist capability and cross-boundary support to identify where the network is becoming vulnerable." }, { "heading": "Enrich incidents automatically", "text": "When an incident is created, retrieve previous location history, building or asset information, known hazards, road constraints, live traffic and relevant spatial layers so dispatchers and commanders receive a concise operating picture." }, { "heading": "Learn from outcomes", "text": "Compare predicted travel with actual, recommended resources with dispatched resources, escalation, arrival sequence and downstream delays. Use evidence to improve models, data quality and policy." } ], "fromTheField": "The platform should become a learning system around CAD—not an opaque replacement for it.", "image": "assets/articles/case-study-emergency-services-ai-data-platform.png" }, { "title": "Can an Autonomous AI Agent Improve Computer-Aided Dispatch?", "slug": "autonomous-ai-agent-computer-aided-dispatch", "category": "Emergency + AI", "date": "2026-08-31", "readTime": "8 min", "url": "https://andrewkapuduwa.com/article-autonomous-ai-agent-computer-aided-dispatch.html", "description": "How an AI operational agent can analyse crime, traffic, demand, resources and spatial coverage to support CAD without prematurely replacing dispatch authority.", "lead": "Before building an autonomous dispatcher, build an autonomous analyst that can continuously reason about the operating system surrounding the incident.", "sections": [ { "heading": "Give the agent situational awareness", "text": "Analyse current CAD incidents, unit status, vehicle GPS, traffic, road closures, weather, staffing, hospital conditions, major events and historical patterns. The agent can continuously ask where the system is becoming operationally vulnerable." }, { "heading": "Use crime and traffic analytics responsibly", "text": "For policing, focus on calls-for-service, temporal clusters, repeat locations, collision patterns and geographic workload. Use analytics to forecast operational demand and resource need rather than making person-level predictions about criminality." }, { "heading": "Recommend dispatch options, not just the nearest unit", "text": "A fast unit may be the wrong system-wide choice if dispatching it removes specialist or minimum coverage. An agent can compare alternatives and explain travel time, coverage impact and policy constraints." }, { "heading": "Analyse cascading effects", "text": "The most useful capability may be predicting what happens after the current dispatch. Multiple high-priority incidents can create coverage gaps invisible when each event is optimised in isolation." }, { "heading": "Increase autonomy gradually", "text": "Low-risk tasks such as incident enrichment, ETA updates, duplicate-call detection and repositioning recommendations can mature first. High-consequence dispatch authority should remain human until evidence and governance justify otherwise." } ], "fromTheField": "The most useful AI agent in CAD may be an always-on operational analyst sitting beside the dispatcher.", "image": "assets/articles/autonomous-ai-agent-computer-aided-dispatch.png" }, { "title": "GeoAI: The Missing Layer in Intelligent Emergency Response", "slug": "geoai-emergency-response", "category": "GeoAI + Emergency", "date": "2026-09-01", "readTime": "7 min", "url": "https://andrewkapuduwa.com/article-geoai-emergency-response.html", "description": "How geospatial AI can improve routing, dynamic risk, building intelligence and system-wide resource decisions in emergency response.", "lead": "Almost everything in emergency response happens somewhere. That makes location one of the richest analytical dimensions available to police, fire, ambulance and emergency management.", "sections": [ { "heading": "Go beyond nearest-resource routing", "text": "The closest unit is not always the fastest or the best system-wide choice. GeoAI can incorporate traffic, hazards, road restrictions, predicted congestion, vehicle type and the coverage impact of moving a resource." }, { "heading": "Represent risk as dynamic geography", "text": "Flooding, fire, storms, road closures and crowd conditions change. Dynamic geofences and risk surfaces can help route resources around conditions static road networks cannot represent." }, { "heading": "Connect GIS and building intelligence", "text": "For incidents inside complex facilities, responders may need entry points, floor plans, hydrants, hazards, stairwells, utilities and asset information. GIS, BIM and digital twins can connect the outside network with the building context." }, { "heading": "Explain the spatial recommendation", "text": "Instead of “take route B because AI says so,” provide the evidence: travel-time difference, hazard intersections, congestion and impact on remaining coverage." }, { "heading": "Use location as decision logic", "text": "Geospatial information should not merely decorate command dashboards. It should participate in eligibility, risk, routing, demand forecasting, coverage and resource decisions." } ], "fromTheField": "In emergency operations, geography should be part of the decision engine—not a visual layer added after the decision.", "image": "assets/articles/geoai-emergency-response.png" }, { "title": "Secure AI in Airports: Building a Closed and Governed AI Ecosystem", "slug": "secure-ai-airports-closed-ecosystem", "category": "Aviation + AI", "date": "2026-09-02", "readTime": "9 min", "url": "https://andrewkapuduwa.com/article-secure-ai-airports-closed-ecosystem.html", "description": "How airports can deploy enterprise AI through controlled data gateways, model governance, authority boundaries and security-by-design.", "lead": "Airports can gain significant value from AI, but they are critical operational environments. The architecture must prevent AI from becoming an uncontrolled information or transaction path into sensitive systems.", "sections": [ { "heading": "Create an AI security zone", "text": "Use approved model endpoints, retrieval services, tool gateways, evaluation, logging, identity and policy enforcement. Critical airport systems expose controlled capabilities rather than unrestricted application access." }, { "heading": "Separate knowledge from transaction authority", "text": "An assistant may retrieve procedures, approved asset data or flight information without being allowed to change gate assignments, security access, airfield configuration or baggage routing. Read and write authority should be designed independently." }, { "heading": "Protect against AI-specific threats", "text": "Account for prompt injection, data leakage, malicious content, excessive permissions, model or supply-chain compromise and uncontrolled external calls in addition to traditional cybersecurity controls." }, { "heading": "Closed does not mean disconnected", "text": "Approved external services can still be used through enterprise gateways, encryption, DLP, network controls and logging. Sensitive workloads can use private deployments while lower-risk use cases use approved cloud models." }, { "heading": "Govern the innovation portfolio", "text": "Define who can create agents, which data is accessible, which models are approved, how evaluations work, what actions require human approval and how production incidents are handled." } ], "fromTheField": "The opportunity is not an airport full of disconnected AI pilots. It is a reusable, governed AI platform inside a clear trust framework.", "image": "assets/articles/secure-ai-airports-closed-ecosystem.png" }, { "title": "AI for Airport Resource Management: Optimising the Turnaround Rather Than One System at a Time", "slug": "ai-airport-resource-management", "category": "Aviation + AI", "date": "2026-09-03", "readTime": "8 min", "url": "https://andrewkapuduwa.com/article-ai-airport-resource-management.html", "description": "How AI can optimise airport turnaround, gates, stands, ground-support equipment and operational resources as a connected system.", "lead": "An airport is a giant resource-coordination problem. The opportunity for AI is to optimise the turnaround and network of dependencies rather than improve one resource schedule in isolation.", "sections": [ { "heading": "Predict the turnaround, not just the timetable", "text": "Use actual arrival, passenger and baggage load, gate, airbridge, catering, cleaning, crew, weather and historical turnaround behaviour to estimate the probability of achieving target off-block time and identify the limiting constraint." }, { "heading": "Make resource assignment dynamic", "text": "Ground-support equipment allocation can consider location, priority, travel time, charge or fuel, maintenance, operator availability and the knock-on effect on later jobs. The objective becomes minimum total operational delay." }, { "heading": "Continuously reassess gates and stands", "text": "Early and late arrivals, equipment failures and passenger connections change the value of the original plan. AI can model alternative allocations while respecting aircraft compatibility, security zoning and operational constraints." }, { "heading": "Support the operations centre", "text": "Use AI to surface emerging resource conflicts, explain likely downstream impact and present intervention options before delay becomes inevitable." }, { "heading": "Keep domain constraints explicit", "text": "Optimisation without operational rules produces mathematically attractive nonsense. Airport AI needs authoritative constraints, human override and evidence behind every significant recommendation." } ], "fromTheField": "The useful airport-AI question is: what will happen in the next 30–120 minutes, why, and which intervention creates the best overall outcome?", "image": "assets/articles/ai-airport-resource-management.png" }, { "title": "AI, Geospatial Intelligence and BIM: The Next Operating Model for Government Asset Management", "slug": "ai-geospatial-bim-government-asset-management", "category": "Government + GeoAI", "date": "2026-09-04", "readTime": "9 min", "url": "https://andrewkapuduwa.com/article-ai-geospatial-bim-government-asset-management.html", "description": "How government can combine asset registers, GIS, BIM, field evidence, IoT and AI for risk-based maintenance, investment and operational optimisation.", "lead": "Traditional asset management asks what assets an organisation owns. A more intelligent operating model asks what condition they are in, what risk is emerging, where to intervene and what happens if investment is delayed.", "sections": [ { "heading": "Connect lifecycle information", "text": "BIM contains detailed asset and component information from design and construction. GIS provides location and network context. Asset systems hold condition and maintenance history. AI becomes useful when those sources form a connected operational information model." }, { "heading": "Use field evidence as part of the loop", "text": "Mobile inspections and snap-and-report workflows can capture photographs, location, asset identifiers, damage type and notes. Computer vision can identify candidate defects while AI retrieves maintenance history and similar patterns." }, { "heading": "Move from predictive to risk-based maintenance", "text": "Failure probability alone is not enough. Combine condition with consequence, network criticality, population served, service dependency, replacement cost, geography, environmental exposure and budget." }, { "heading": "Use geospatial relationships to understand consequence", "text": "A damaged drainage asset may sit upstream of a road, school or flood-prone area. Spatial analysis reveals network effects that an isolated asset register cannot show." }, { "heading": "Build a practical digital twin", "text": "The useful government digital twin is not necessarily photorealistic. It is a decision model connecting asset register, GIS, BIM, inspection, maintenance, IoT, finance, weather and field reports so executives and operators can ask different questions of the same evidence base." } ], "fromTheField": "The opportunity is not simply AI for predictive maintenance. It is a continuous asset decision system combining data, geography, engineering context and operational evidence.", "image": "assets/articles/ai-geospatial-bim-government-asset-management.png" } ], "confidentiality": "Selected assignments are anonymised. Do not infer client identities." }