Insights
We publish what we have learned, with the evidence behind it
Our articles cover three topics: AI governance, systems integration and sector notes. Each article states its main claim and the evidence for it.
Topics
UNIT 01
AI governance
AI governance means showing, with evidence, that a deployed AI system still does what it was built to do, months after launch. Singapore's IMDA framework, the EU's rules and ISO/IEC 42001 all set the same requirement. Each system needs a named owner, a log of its decisions and a way to roll it back that is tested before it is needed. These articles are written by an ISO/IEC 42001 Lead Auditor. They are for operators who must defend a deployment to a regulator, a board or an auditor.
9 articles published
Systems integration
Most AI failures are integration failures, even when the model takes the blame. Common causes include data that differs from one system to the next, a protocol chosen without comparing the alternatives, and a pilot that never received the engineering a production system needs. These articles argue that you should first make your machines and systems agree on their data. Only then should you add AI on top. An AI layer built on inconsistent data inherits the errors, and often nobody notices.
17 articles published
Sector notes
The right first step depends on the sector. Consider a port terminal automating a yard crane, a food plant defending its HACCP (Hazard Analysis and Critical Control Points) records, and an SME (small or medium-sized enterprise) choosing its first AI use case. Each needs to know where automation pays back and what evidence supports that claim, and the answer is different for each. These articles cover maritime and port operations, manufacturing and food safety, logistics, and small and mid-sized businesses. Each one is based on a named case.
14 articles published
Robotics
A robot's certificate covers the robot, not the site it works in. These articles cover what the evidence shows about humanoid and mobile robots in production, the safety standards a site must design to, and where robot automation pays back in a small factory. They are for operators deciding whether a robot suits a task, and how to test that before buying a fleet.
5 articles published
Assurance
Assurance is evidence that a management system, or an AI system inside it, works as stated and keeps working after the audit. These articles cover ISO/IEC 42001 certification, the deadlines of the European Union (EU) AI Act, AI vendor contracts, and why assurance costs less when it is designed in before go-live. They are written by an ISO/IEC 42001 Lead Auditor for organisations preparing for certification or an audit.
4 articles published
Recommended reading for today
UNIT 02We recommend starting with these three articles. The selection changes every day.
Sector notesFive questions to answer before you buy an AI tool
If you buy an AI tool before checking whether the business can use it well, you waste the money. You also create operational debt that the next project has to work through.
AI governanceShadow AI is already inside your organisation
Many employees already use personal AI tools for work, whether or not a policy allows it. A ban leaves the risk in place and hides the usage from you.
AssuranceAI assurance costs less if you build it in before go-live
Assurance is evidence that a system does what it should and continues to do so after handover. Adding assurance to a system already in production costs materially more.
All articles
UNIT 0349 of 49 articles
- Assurance
Why ISO/IEC 42001 certification matters more as AI spreads
ISO/IEC 42001 certification is the only auditable, internationally recognised way to show that an organisation controls its AI. We explain what it proves and how it relates to the EU AI Act.
AssuranceThe EU AI Act after the 2026 delay: what applies now and what comes next
The 2026 Digital Omnibus delayed the EU AI Act's high-risk deadlines, but most duties that affect operators already apply. We set out every date and the Act's reach outside the EU.
- Robotics
Humanoid robots at work, 2024 to 2026: what the evidence shows
Between 2024 and 2026, humanoid robots began their first paid work, in thousands of units. We set out what was deployed, who reported the numbers and what remains unproven.
Sector notesWhat grant assessors look for in an application
Two of our named projects were grant-funded: the Raymang Eggs robotics line under an Enterprise Development Grant, and the PestBusters programme, the first project under the Lead Enterprise Development Scheme. The technology did not win either grant. Both applications succeeded because they set out a defined problem, a scoped project and an outcome that could be measured.
Systems integrationHandover is the last deliverable, and most contracts leave it undefined
It is easy to write 'You own the system' into a contract. In the thirteenth month, when the integrator has left and something needs changing, that clause comes down to a short list of specific items. Each item was either handed over or it was not.
AI governanceBuild data retention rules into the system's data layer
A retention policy in a document depends on people following it. A retention rule built into the database schema is enforced by the system. The bunkering system at PSA Marine purges a face record six months after the account is deactivated. The purge is part of the system, so it needs no procedure and nobody has to remember to run it.
Sector notesReport automation results in hours instead of percentages
A ratio reported without a baseline cannot be checked. We restated a robotics line's outcome from an efficiency percentage to the time the same work now takes. The new figure could be checked, and it was also larger.
AI governanceWhere your data goes belongs in the scope document
To a public-sector procurement reader, an integrator with engineers in five jurisdictions raises a cross-border data question first. The answer belongs in the scope document, in three sentences.
Systems integrationWhy AI must sit outside a machine's safety-rated control loop
A machine's safety function must respond within a certified, bounded time. A model's inference time is neither certified nor bounded. Most problems described as AI integration challenges in plants come from this mismatch, which teams often misdiagnose as a data or model problem.
AI governanceUse SAE's levels of driving automation to define your machine's decision rights
Since 2014, SAE International's J3016 standard has given the automotive industry a shared six-level scale for how much a vehicle decides on its own. Industrial equipment has no equivalent. You can close that planning gap now, without waiting for one to be published.
Systems integrationUnifying your data can remove the access controls that protected it
Connecting scattered data is the first step. Once that unified data feeds or trains a model, the permissions that kept finance data separate from HR data do not carry over automatically.
Systems integrationMost AI projects fail because the data is not ready
Gartner estimates that 85 percent of enterprise AI projects fail to deliver real business value. Buying a capable model before the underlying data is connected makes the route to value longer.
Systems integrationCompanies fail to scale AI because of the systems around it
McKinsey's November 2025 survey found that 78 percent of companies use AI regularly, but only 33 percent are scaling it across the organisation. MIT NANDA's much-cited figure, that around 5 percent of pilots reach lasting financial gain, points the same way. Its method has been contested, so we treat it only as a sign of direction.
Sector notesSingapore SMEs tripled their AI use, but the gap with larger firms still widened
Data from the Infocomm Media Development Authority (IMDA) shows that 95.1 percent of small businesses already use at least one digital tool, and that AI use tripled between 2023 and 2024. The main constraint is choosing the wrong first use case.
RoboticsRobotics return on investment in SME factories: where it is real and where it is overstated
A collaborative robot (cobot) performs the four-thousandth repetition exactly like the first. That consistency is the whole case for buying one. The case fails once the task changes more often than the cycle repeats.
- Assurance
AI assurance costs less if you build it in before go-live
Assurance is evidence that a system does what it should and continues to do so after handover. Adding assurance to a system already in production costs materially more.
AI governanceNow that agentic AI is sold commercially, buy it on traceability
Once an agent is billed for completing tasks, the buyer's question changes. A benchmark score matters less than whether the agent finished each task and can prove that it did.
Sector notesThe Scope 3 figure your customer must report is already in your OT data
From the 2026 financial year, Singapore's largest listed companies must disclose Scope 3 emissions. If a supplier offers no activity-based data from its operational technology (OT), the customer has to report an industry-average estimate for that supplier instead.
Systems integrationSome problems need a lookup table instead of an AI model
The question has shifted from whether AI can do something to whether it can run reliably, at scale and within budget. Cost now rises with adoption, whatever value the system creates.
Sector notesSingapore SMEs have seven funding instruments but still lack a structured approach
Singapore's policies are designed to lower the cost of AI adoption for SMEs. The main constraint is a lack of alignment between what leaders intend, how processes are defined and how consistently the work is carried out.
Sector notesA five-year autonomy roadmap: what to buy first, what to buy third and what to delay
Every capability covered in our articles (retrofit-first integration, AMR safety standards, edge compute and digital twins) has to be introduced in some order. Most autonomy programmes that fail bought the right technology in the wrong sequence.
AssuranceWhat your AI vendor contract is probably missing
Most AI agreements reuse SaaS contract templates. Those templates assume software that does what it is told, performs consistently and does not learn from your data. An AI system meets none of these assumptions.
- AI governance
Shadow AI is already inside your organisation
Many employees already use personal AI tools for work, whether or not a policy allows it. A ban leaves the risk in place and hides the usage from you.
Sector notesYour first AI use case decides whether the programme survives
MIT's research states clearly that most pilot failures come from poor use case selection. The first two or three use cases carry far more weight than later ones, and the usual instinct for choosing them is the opposite of what works.
- Sector notes
Five questions to answer before you buy an AI tool
If you buy an AI tool before checking whether the business can use it well, you waste the money. You also create operational debt that the next project has to work through.
Sector notesEmbedded hardware AI faces far less competition than software AI
Every enterprise platform now has an AI layer, and every SaaS product has a copilot. Intelligence built directly into equipment and physical infrastructure is a category that has not yet been reinvented in the same way.
Systems integrationWhy your WMS and your ERP probably disagree
A warehouse management system (WMS) rarely closes the gap it was bought to close. The gap usually sits at the handoff into the enterprise resource planning (ERP) system, where someone still reconciles the two systems by hand. Most of the real cost of a warehouse project lies in that reconciliation step.
Sector notesAutonomous ships now have a global code, and Singapore's Tuas port was already planned for it
The International Maritime Organization (IMO) adopted its first global code for autonomous ships in May 2026. The code is voluntary from July that year and mandatory by 2032. PSA's Tuas mega-port is automating in phases toward 65 million TEU (twenty-foot equivalent units) a year. It has followed the same planning horizon without waiting for the regulation.
Systems integrationLocal inference versus API: where the break-even point lies
Running models on your own infrastructure is the right choice under certain conditions. The break-even point depends on sustained token volume, regulatory constraints and whether you have machine learning operations (MLOps) capability in-house.
Systems integrationWhy machine vision inspection pays back on some lines and stalls on others
The return depends mainly on whether the defect you want a camera to catch has a stable visual signature. The sophistication of the model matters far less.
AI governanceA third of enterprise applications are forecast to act on their own by 2028. Prepare next year's contracts for it now
Gartner predicts that by 2028 a third of enterprise applications will include agentic AI that can take semi-autonomous action. This is a forecast. The best way to use it is to plan the contract terms for machinery you have not yet bought.
Sector notesFood safety records are still handwritten beside sensors that already have the data
Monitoring critical control points (CCPs) is one of the seven founding principles of HACCP (Hazard Analysis and Critical Control Points). Most of this monitoring is still recorded by hand, even where the equipment already produces the same reading digitally. Audit findings repeatedly point to this gap.
Systems integrationOPC UA and Sparkplug B solve different problems, and choosing by trend is costly later
Both protocols let a modern AI system read data from plant equipment. They were built for different jobs. If an integration is scoped around the wrong one, the cost remains long after the pilot ends.
Systems integrationPredictive maintenance works only when the sensor data feeding it is reliable
The US Department of Energy (DOE) credits a mature predictive maintenance programme with cutting maintenance costs by 25 to 30 percent and downtime by 35 to 45 percent. Those figures assume the instrumentation feeding the analytics is already sound. At most plants, nobody has checked.
Systems integrationThe onboard chip determines what a machine can still do when the network fails
NVIDIA's Jetson Thor, generally available since August 2025, puts more than 1,000 TOPS (trillion operations per second) of inference compute on a module built to sit inside a machine. That compute limits how autonomous a piece of equipment can be without a live connection.
AI governanceFrom human-in-the-loop to AI-on-the-loop: redesigning oversight
In several operational areas, AI now checks work more reliably than human reviewers, especially at high volume and over long periods. Governance regimes still assume that a person is accountable, so the design of oversight has to change.
Sector notesCold chain spot checks miss the temperature excursions that happen between readings
A joint study by the UN Food and Agriculture Organization (FAO) and the International Institute of Refrigeration puts the loss from inadequate refrigeration at 526 million tonnes of food a year, 12 percent of the global total. A scheduled spot check measures only the moment someone looked, and that moment rarely coincides with an excursion.
Systems integrationModel a production line before you automate it. ISO 23247 now defines how to build that model
A digital twin is useful only if people can trust it. Until this year, the international standard for manufacturing digital twins lacked the part that defines how a twin is composed. ISO 23247 added that part in 2026.
Systems integrationMeasure OEE before you approve the automation budget
Overall Equipment Effectiveness (OEE), split into its three components, shows which problem your automation budget would actually fix. If you install it as a dashboard after the equipment is already running, it gives you nothing you can act on.
RoboticsThe lesson of Amazon's millionth robot is the thirteen years it took
Amazon passed one million robots deployed across its fulfilment network in July 2025, a figure now roughly equal to its own headcount. For an operator planning automation, the more useful number is how long the fleet took to build.
Sector notesEnergy submetering is the fastest-payback automation project most plants have not yet done
The US Department of Energy's (DOE) own verification data puts the average payback on a metered energy management programme at 1.7 years. Most of the savings came from operational fixes, without new equipment. A plant that reads only its monthly utility bill cannot achieve any of these savings.
RoboticsHumanoid robot forecasts are credible, but a 2026 purchase order is unlikely to be for one
Goldman Sachs projects the humanoid robot market at $38 billion by 2035, and Morgan Stanley projects $5 trillion by 2050. Industrial mobile robots, the less visible wheeled kind, already ship in the hundreds of thousands, and their case does not depend on either forecast.
Systems integrationRetrofitting IoT sensors onto older plant equipment
Most equipment in service today was specified before anyone planned to fit sensors to it. Replacing it just to get telemetry rarely makes economic sense. A retrofit, however, fails in physical ways that software teams often discover only after they have drilled into the enclosure.
Sector notesWhere AI is safe to deploy
Whether an AI deployment is safe is no longer a yes-or-no question. The answer varies by industry, application and regulatory tier. The organisations that get a return match the project's scope to their data readiness and apply the right level of human oversight.
RoboticsAutonomous mobile robots have a safety standard, but most sites do not design to it
ISO 3691-4 sets verification requirements for driverless industrial trucks, and ANSI/RIA R15.08 does the same for industrial mobile robots. A vendor demonstration shows that a robot can move around a site, but neither standard was written to test only that.
Systems integrationWhy the demo works and the production system does not
The difference lies in the operational architecture around the model: context handling, evaluation, versioning, observability and cost control. A demo needs none of these.
AI governanceBuild AI systems that recognise when they do not know
A softmax classifier cannot output 'unknown'. It must always pick an answer, and it will assign high confidence to inputs it has never seen. In embedded infrastructure, that behaviour is a liability.
AI governanceA working pilot does not prove an AI system is ready for production
AI systems fail silently, so a standard risk matrix ranks their risks wrongly. An unnoticed hallucination in a citizen-facing service can be a greater risk than a visible outage that was noticed and fixed within the hour.
Systems integrationClean the data before you tune the prompts
Prompt tuning soon stops improving results, because it cannot fix the documents a retrieval system is reading. On one project, cleaning the knowledge base reduced the hallucination rate more than three weeks of prompt engineering had.