
Computer vision for quality inspection has matured to the point where the model itself is rarely the difficult part of the project. The difficult parts are everything around the model: lighting, calibration, the operator workflow, the defect taxonomy that the production line agrees with, and the operating discipline to keep the system trustworthy through a year of normal manufacturing variation.
Pattern one: the defect taxonomy is the project
Every line we have worked with has two defect taxonomies — the one the quality system uses and the one operators use when grading manually. They are never the same, and a model trained against either one alone will be wrong by the other. The first deliverable on any deployment is a reconciled taxonomy that both operators and the quality function will stand behind. Without it, the model is technically capable and operationally useless.
Pattern two: installation conditions consume most of the accuracy budget
A model trained in conditions that precisely match the production line will perform. A model trained in slightly different conditions — different lighting, different angle, different line speed — will not, and the failure mode will resemble a model problem when it is actually an installation problem. The discipline that separates effective deployments is the willingness to spend a week on lighting and fixturing before a single image is labelled.
Pattern three: trust is built one shift at a time
- Operators need to see what the model is flagging and why — not simply receive a verdict
- False positives need a fast and transparent remediation path; the operator who sees three in a shift without resolution will stop trusting the system
- Every model recall after a defect escape needs a clear post-mortem that the production crew sees, not just the quality function
- The model owner needs to be present on the line regularly — a deployment that stops being visited is a deployment that quietly drifts
The vision technology is no longer the constraint. The constraint is the operating discipline to install it correctly and maintain the trust of the people working alongside it. The deployments that hold their performance treat that discipline as seriously as the model selection itself.
More insights

Where enterprise AI creates durable business value
Most enterprise AI programmes fail not because of the technology but because of how organisations approach them. The firms capturing lasting value share three disciplines that have nothing to do with model performance.

The network operations centre is a strategic asset waiting to be unlocked
Telecom operators have invested heavily in network infrastructure. The returns on that investment are increasingly determined not by the infrastructure itself but by the intelligence layer running on top of it.

AI strategy in high-growth markets: lessons from the Gulf
Governments and enterprises across the Gulf are committing significant capital to AI transformation. The organisations seeing results are those that treated AI as a strategic management question before treating it as a technology question.

Rethinking enterprise security: from access control to risk intelligence
The enterprise security function is evolving from a compliance obligation into a source of operational intelligence. The transition requires more than new technology — it requires a different conception of what security is for.

The data architecture decisions that determine transformation outcomes
Most organisations approach data architecture as an IT infrastructure question. The firms that extract the most business value from their data treat it as a strategic design question — one that belongs in the boardroom, not just the server room.

AI in financial services: separating the durable from the hype
Financial institutions have been among the earliest and most enthusiastic adopters of AI. They have also produced some of the clearest illustrations of where AI programmes go wrong and what it takes to make them last.

Building a customer intelligence capability that actually changes decisions
Retailers collect more customer data than ever. Most of it sits in systems that don't talk to each other, analysed by teams that don't talk to merchants. The result is insight that arrives too late to matter.

How to tell if your organisation is actually ready for enterprise AI
AI readiness is not a technology question. It is a diagnostic across strategy, data, governance, talent, and operating model. Most organisations that think they are ready are not — and the signals are detectable before any money is spent.

AI in healthcare: the organisational challenges matter more than the technical ones
Healthcare AI projects fail at a rate that should give every executive pause. The failure mode is rarely the algorithm — it is the operating environment the algorithm is dropped into without adequate preparation.

The logistics transformation most companies get wrong
Route optimisation is frequently the first AI application logistics companies pursue. It is also the one most likely to deliver disappointing results — not because the technology doesn't work, but because the problem definition was wrong from the start.

Energy transition strategy: the grid challenges that determine the pace
The energy transition story is usually told as a generation story. The harder problem — and the one operators are investing the most in — is grid flexibility: matching variable supply and variable demand in real time across infrastructure built for neither.

From smart city pilot to permanent capability: what changes at scale
Smart city pilot projects are familiar. Smart city permanent operating capabilities are considerably rarer. The difference between cities that make the transition and those that do not comes down to a small number of decisions made at the start.

Government digital transformation: the implementation gap
Government digital programmes report high levels of service digitisation. The metric that matters — end-to-end completion rate — tells a different story. Most programmes have not yet made the measurement straightforward enough to be honest about it.

Insurance claims transformation: speed, accuracy, and the fraud trade-off
Claims automation has a sound business case. It also has a set of trade-offs that are routinely underweighted in the evaluation process. A candid note for insurance executives assessing the move.

ESG performance reporting: from compliance exercise to strategic data asset
ESG reporting is moving from voluntary disclosure to assurance-grade obligation. Most enterprises are not ready, and the reason is rarely policy — it is the data layer underneath the report.

Data governance under AI regulation: the four decisions that matter most
Data governance was a back-office concern until AI made every production model a regulated artefact. The four decisions that determine whether an enterprise's governance approach holds up under supervisory scrutiny.

MLOps in regulated industries: the discipline gap in enterprise AI delivery
MLOps reference architectures are well documented. None of them describe what changes when the model is regulated and every promotion to production needs documented approval from a named accountable person.

Five questions every board should be asking about their organisation's AI programme
Boards have moved past asking whether the enterprise should use AI. The question is now what the enterprise is exposed to by using it, and whether management can demonstrate the answer is known and managed.

AI-augmented defence: how the threat landscape is changing for enterprise security
AI in the security operations centre tends to fall into one of two patterns: the kind that converts existing alert noise into actionable signal, and the kind that adds more noise. The distinction is detectable before you sign the contract.
