QA Category: The AI Fundamentals Q&A

This is the question that matters most, and the one where I’ve seen the most money wasted. The good news: the failure patterns are well known and avoidable. The starting point is never the technology. Start from pain, not from AI. List your top operational problems with a quantifiable cost: false-reject rates on inspection lines, unplanned downtime on critical equipment, hours spent manually investigating supply chain exceptions, slow deviation handling. AI is a means; if a problem can be solved with better rules or better process, solve it that way and save your AI budget for problems that genuinely need…

Large Language Models are the most visible face of AI today, and the pressure to “do something with ChatGPT” reaches every boardroom. In our industry they have real, valuable uses – provided you understand one fundamental characteristic of how they work. An LLM is a deep learning model trained on enormous amounts of text to do one thing: predict plausible continuations of language. From this single capability emerge remarkable skills, summarizing, translating, drafting, answering questions, writing code. But the mechanism matters: the model generates what is statistically plausible, not what is verified true. When plausibility and truth diverge, you get…

Having spent years working on predictive maintenance with smart sensors, I can tell you where the real value is – and where the marketing promises quietly exceed what physics and data allow. The idea is compelling: instead of maintaining machines on a fixed schedule (often too early, wasting parts and downtime) or after failure (too late, causing unplanned stops), you monitor their actual condition and intervene exactly when needed. What AI does well here rests on a physical truth: most mechanical degradation announces itself. Bearings developing faults change their vibration signature weeks before failure; motors under stress show thermal and…

“It’s a black box – Quality will never accept it”. I hear this in almost every pharma conversation about AI. It’s a legitimate concern, but the answer is more nuanced, and more actionable – than the objection suggests. The starting point: regulated industries have always validated systems whose internals are complex. What validation actually requires is not that every internal mechanism be humanly readable, but that the system’s behavior be specified, tested, documented, and controlled. The question shifts from “can I read the logic?” to “can I demonstrate, with evidence, that the system performs its intended use reliably – and…

A model that performed brilliantly at validation can quietly become unreliable a year later without anyone changing a single line of code. This phenomenon has a name – drift – and understanding it is essential for anyone operating AI in production. First, what a model actually is: the output of the training process. Concretely, it is a file containing an architecture (the structure of the computation) and parameters (millions of numbers tuned during training). Given an input – an image, a sequence of events, a sensor window – it produces an output: a classification, a score, a prediction. The model…

Visual inspection is where AI meets pharma most concretely: cameras, vials, blisters, labels moving at hundreds of pieces per minute. But when I ask people how they think the machine recognizes a crack, the answers reveal a fundamental misunderstanding worth clearing up. A digital image, for a computer, is just a grid of numbers – pixel intensities. Traditional machine vision works by having engineers define explicit operations on those numbers: thresholds, edge filters, geometric measurements. “If the dark region in this zone exceeds N pixels, reject.” This works beautifully for stable, well-defined tasks – dimensional checks, presence/absence, code reading –…

Companies in pharma traceability sit on enormous datasets – billions of serialized events, years of aggregation hierarchies, complete audit trails. So the question comes naturally: “We have all this data, surely we can do AI?” The honest answer is: it depends on what your data actually contains – and, above all, on what question you are asking. AI needs data for one reason: models learn by example. To learn what a “suspicious shipment pattern” looks like, a model needs to see many shipments – normal and anomalous – with enough context to tell them apart. Three properties matter far more…

These three terms are used interchangeably in vendor presentations, and that is a problem: they refer to nested but very different technologies, with very different costs, data requirements, and risk profiles. Think of them as concentric circles: Machine Learning (ML) is the broad discipline: algorithms that learn patterns from data. This includes venerable, battle-tested techniques – regression, decision trees, gradient boosting – that power most industrial applications today. Predicting reject rates from process parameters, classifying deviation reports, forecasting spare-parts demand: classical ML, often trained on tabular data, frequently the most robust and explainable choice. Deep Learning (DL) is a subset…

When I explain that a machine learning model is not programmed but trained, I often see skepticism – until I explain what this means for testing, validation, and change management. Then the skepticism turns into a long list of very good questions. In traditional software development, engineers write explicit logic: inputs go in, deterministic rules apply, outputs come out. If something is wrong, you find the faulty line of code and fix it. The behavior is fully inspectable. Machine learning inverts this. You don’t write the logic – you provide examples (data) and a learning algorithm derives the logic itself,…

Every conversation I have in a pharmaceutical plant starts the same way: someone uses the word “AI” to describe five completely different things – a rules-based alarm, a statistical dashboard, a vision system, a chatbot, and a Hollywood robot. Before deciding whether AI can help your operations, we need to agree on what we’re talking about. At its core, Artificial Intelligence is software that performs tasks which, when done by humans, require perception, judgment, or learning – recognizing a defect on a vial, predicting when a motor will fail, extracting meaning from a maintenance report. The key distinction from traditional…