AI came up on 337 earnings calls last quarter. Most leadership teams still can't draw the stack.

In plain EnglishSoftware that improves at a task by being shown examples, instead of being given rules by a programmer. Nobody writes "an invoice usually has a total in the bottom right" — the system infers it from ten thousand invoices.

How it actually worksYou supply labelled examples, the system makes a guess, measures how wrong it was, and adjusts itself slightly. Repeat several million times. The output is not a program you can read; it is a set of numbers that happen to produce good answers. This is why "explainability" is a genuine engineering problem rather than a documentation problem.

In plain EnglishMachine learning arranged in many stacked layers, loosely inspired by neurons. "Deep" simply refers to the number of layers — the depth is what lets it learn subtle patterns nobody thought to describe.

How it actually worksEach layer transforms the data slightly and passes it on. Early layers in an image system detect edges; middle layers detect shapes; later layers detect "this is a cracked weld." No human specified those stages. They emerged from training, which is both the power and the governance headache.

In plain EnglishOne very large model trained on a very broad body of data, then adapted to many specific jobs. The opposite of the old approach, where you trained a separate small model for each task.

How it actually worksTraining happens once, at enormous cost, by a handful of organisations worldwide. Everyone else adapts what exists — through prompting, retrieval, fine-tuning, or a thin application layer. This is the single most important structural fact about the current market: the expensive part has already been paid for by someone else.

In plain EnglishA foundation model specialised in language. It predicts what text should come next, which — at sufficient scale — turns out to look a great deal like writing, summarising, translating and reasoning.

How it actually worksIt has no database of facts and performs no lookup. It generates a plausible continuation, one fragment at a time, from statistical structure learned during training. Everything else — accuracy, citation, tool use — is engineering built around that core behaviour, not a property of the model itself.

In plain EnglishDeliberately smaller models, trained or tuned for one domain — legal, clinical, industrial maintenance, a single language. Less general knowledge, far lower cost, and able to run on your own hardware.

How it actually worksYou trade breadth for economics and control. A model that will never need to discuss Renaissance art can be a fraction of the size while outperforming a general model on your contracts, because it has seen far more of them.

In plain EnglishTraining is teaching the model — once, at vast expense. Inference is using it — every single time, forever. Confusing the two is the most common costing error in enterprise AI.

How it actually worksTraining is a capital event measured in months and hundreds of millions. Inference is an operating expense measured per request. Your AI bill is almost entirely inference, and unlike seat-based software it scales with usage, not headcount — which is exactly why it surprises finance teams.

In plain EnglishA token is a fragment of text — very roughly three-quarters of a word. The context window is how much the model can hold in mind at once: your prompt, your document, and its own answer, all counted together.

How it actually worksYou are billed per token in and per token out. When a conversation or a document exceeds the window, the earliest material silently drops away — the model does not announce that it has forgotten, it simply answers as though the missing part never existed.

In plain EnglishThe internal numbers a model adjusts during training. "70 billion parameters" is a size claim — roughly analogous to engine capacity, and about as reliable a guide to actual performance.

How it actually worksMore parameters generally means more capability and always means more cost to run. But training data quality, tuning method and how the model is deployed matter at least as much. Since 2025 the frontier has been moving toward smaller models that are better trained rather than simply larger ones.

In plain EnglishTaking an existing model and training it further on your own examples, so it adopts your tone, format or specialist judgement.

How it actually worksYou supply a curated set of input–output pairs and run a much smaller training job. It shifts behaviour and style reliably. It is a poor and expensive way to teach the model facts, because every time your facts change you must do it again.

In plain EnglishBefore answering, the system searches your own documents and hands the relevant passages to the model. The model then answers using that material rather than from memory.

How it actually worksYour documents are indexed and stored. A question triggers a search, the best-matching passages are pasted invisibly into the prompt, and the model composes an answer grounded in them — usually with a citation back to the source page. Update a document and the answer updates immediately, with no retraining.

In plain EnglishThe craft of telling the model precisely what you want and giving it the right material to work from. It is the only item on this entire page that you personally can get better at, without buying anything.

How it actually worksOutput quality is governed far more by the instruction and the supplied context than by which model was chosen. "Context engineering" is the 2026 term for the discipline this matured into — deciding deliberately what goes into the context window, in what order, and just as importantly what is left out of it.

In plain EnglishWhen a model states something false with complete confidence. Not a bug that will be patched — a direct consequence of a system built to produce plausible continuations rather than to look things up.

How it actually worksThe model has no internal sense of certainty about facts. Fluency and accuracy are produced by the same mechanism, which is why a wrong answer reads exactly as convincingly as a right one. Grounding, citation and human review reduce the rate substantially; nothing available today eliminates it.

In plain EnglishMachines recognising things — images, speech, defects, faces, signatures, sounds. The oldest capability on this list and, for industrial businesses, still the one with the clearest return.

How it actually worksA model trained on many labelled examples learns to classify new ones. It does not understand what it sees; it matches learned patterns extremely well. Accuracy on a narrow, well-defined task now routinely exceeds a tired human on the third shift.

In plain EnglishProducing new content — text, images, code, audio, video, designs — rather than classifying existing content. The capability that made AI a boardroom subject in 2023.

How it actually worksThe model generates one fragment at a time, each conditioned on everything before it. Quality depends far more on the instruction and the material you supply than on which model you chose, which is why prompt discipline outperforms model-shopping in almost every organisation.

In plain EnglishModels that work a problem through in steps before answering, rather than responding immediately. They spend more time — and more money — on hard questions, and get substantially better answers on them.

How it actually worksThe model generates an internal chain of intermediate steps, checks itself, and can revise before producing a final answer. This is why responses arrive more slowly and cost more per query. On multi-step analysis, engineering calculation and code, the improvement over standard generation is not marginal.

In plain EnglishAn AI agent is the thing itself — a single system given a goal and the means to pursue it. Agentic is the property: how much initiative it is permitted. The two words are used interchangeably in the market and should not be. Systems that take actions on their own initiative — planning a sequence, using tools, calling other software, adapting when something fails — rather than only answering a question you asked.

How it actually worksA reasoning model is given a goal, a set of tools it may call, and permission to loop: plan, act, observe the result, re-plan. Autonomy is a dial, not a switch. Nearly every deployment that works in practice keeps a human approval gate on anything that spends money, contacts a customer or changes a record.

In plain EnglishSeveral specialised agents working together on one process — one gathers data, one drafts, one checks, one executes — each narrow and therefore more reliable than a single agent attempting everything.

How it actually worksAn orchestrator hands work between agents with defined roles and handover rules. It mirrors how you would structure a competent team, which is both the appeal and the difficulty: errors compound across handovers, and diagnosing a failure across four agents is materially harder than across one.

In plain EnglishAI that senses and acts in the real world — robots, drones, autonomous handling equipment, machines that adapt rather than repeat a fixed program.

How it actually worksIt combines perception (cameras and sensors), a model that decides, and actuators that move — increasingly trained in simulation before touching a real machine. Gartner describes it as AI models combined with IoT sensors, robotics and automation systems to enable real-time sensing, analysis and execution.

In plain EnglishOne model that handles several kinds of input together — text, images, audio, video, diagrams — instead of a separate system for each.

How it actually worksDifferent input types are converted into a shared internal representation, so the model can reason across them at once: read a drawing and the specification beside it, or watch a video and answer a question about what happened at 4:12.

In plain EnglishAn AI helper sitting inside software your people already use — drafting, summarising, searching, suggesting. It advises; the human still acts.

How it actually worksA generative model is wired into the application's own context — your document, your inbox, your ticket — so its suggestions are relevant without anyone pasting anything anywhere. Adoption is easy precisely because nothing new has to be learned.

In plain EnglishAI deployed against your own institutional knowledge — contracts, standards, drawings, policies, project history — under your own access controls.

How it actually worksAlmost always retrieval over your document estate, with permissions honoured so a user only ever sees answers drawn from material they were already entitled to read. The hard part is never the model. It is the state of your documents and the accuracy of your permissions.

In plain EnglishA defined multi-step process handled end to end — intake, extract, validate, route, draft, escalate — with humans holding the exceptions rather than the whole queue.

How it actually worksTraditional automation handles the deterministic steps; AI handles the judgement ones — reading an unstructured email, deciding whether an invoice matches a purchase order, classifying an unclear complaint. The combination outperforms either alone.

In plain EnglishUsing sensor data to predict that a machine is heading for failure, and intervening before it stops — rather than on a fixed calendar or after the breakdown.

How it actually worksVibration, temperature, current and acoustic data are monitored continuously against learned normal behaviour. Deviations that historically preceded failure raise a flag with lead time attached. It needs sensors, historical failure data and someone empowered to act on the alert — the third is the one projects usually miss.

In plain EnglishCameras checking quality, completeness or safety compliance in real time, at line speed, on every unit rather than on a sample.

How it actually worksA perception model trained on your defects runs on a small computer at the line — no cloud, no latency, no data leaving the plant. Modern systems learn new defect types from a modest number of examples rather than thousands.

In plain EnglishCustomer-facing systems that hold an actual conversation — across chat, voice and messaging — and increasingly resolve the issue rather than routing it.

How it actually worksA language model grounded in your policies and product data, connected to the systems that can actually do something: check an order, issue a credit, book a slot. The connection to those systems, not the conversation quality, is what separates a resolution from a well-worded apology.

In plain EnglishA live virtual copy of a physical thing — a machine, a production line, a building, a supply chain — fed by real sensor data, used to test a change before you commit to it in the real world.

How it actually worksSensors stream the asset's actual state into a model that mirrors it. You then run the proposed change against the twin: a new schedule, a different feed rate, a rerouted supply. What the twin tells you is only as good as the sensor coverage feeding it, which is why this sits on top of your IoT investment rather than replacing it.

In plain EnglishThe specialised processors and buildings that make all of this possible. GPUs — originally built for graphics — turn out to be ideally suited to the arithmetic that neural networks require.

How it actually worksTraining and inference both run on large clusters of these chips, which consume substantial power and require serious cooling. Availability, not ambition, has become the binding constraint: capacity is being pre-sold years ahead, and power connection timelines now shape what can be built and when.

In plain EnglishWhere the model runs and where your data goes while it does. Cloud means someone else's data centre. On-premise means your building. Sovereign means inside your country's legal jurisdiction, under its rules.

How it actually worksCloud gives you the largest models with no capital cost and a usage-based bill. On-premise means an appliance or server you own — higher upfront cost, no per-token meter, and nothing crossing your firewall. Smaller domain models have made the on-premise option genuinely viable for the first time, because you no longer need a frontier-scale model to do useful work on your own documents.

In plain EnglishIoT is the network of connected sensors and machines generating data. Edge computing means processing that data where it is created — on the machine, in the plant — rather than shipping it to a data centre first.

How it actually worksA sensor produces a continuous stream; a small model at the edge decides what matters and acts in milliseconds; only exceptions and summaries travel upstream. This is what makes real-time inspection and safety intervention possible at all — a round trip to the cloud is far too slow for a moving production line.

In plain EnglishAn open standard for connecting an AI model to your tools and data. The metaphor that stuck is USB-C for AI: one universal connector, instead of a different custom cable for every system you own.

How it actually worksBefore it, every connection was hand-built — one adapter for your ERP, another for SharePoint, another for your CRM — and all of it rewritten whenever you changed model. With MCP, each of your systems exposes a small "server" that describes what it can do and what it will permit. Any compliant model can then discover and call it. Change the model, keep the integrations.

In plain EnglishClosed means you rent access through an API and never possess the model. Open weights means you can download it and run it on your own machines, permanently.

How it actually worksClosed models generally hold the frontier on raw capability, cost nothing upfront, and leave you dependent on a supplier's pricing, availability and terms. Open-weight models you host yourself: higher setup effort, no per-token meter, complete control of where data sits — and the full compliance and security burden lands on you rather than the vendor.

In plain EnglishThe rules you must follow when you deploy AI. The EU AI Act is the most consequential, phasing obligations in by risk tier. India's Digital Personal Data Protection Act governs the personal data that feeds these systems. Sector regulators — financial, medical, safety — apply their own rules on top.

How it actually worksObligations scale with the risk of the use case, not the sophistication of the technology. A simple model making hiring or credit decisions carries far heavier duties than an advanced one drafting marketing copy. The EU framework also reaches beyond Europe: if your output touches the EU market, it can apply to you from Pune or Delaware.

In plain EnglishGuardrails are the checks placed around a model — what it may see, say and do. Prompt injection is when hidden instructions buried inside content the model reads hijack its behaviour. Tool poisoning is when something it is connected to has been tampered with.

How it actually worksA model cannot reliably tell the difference between instructions from you and text it happens to be reading. So a booby-trapped email, invoice, CV or web page can carry commands. While the system only answers questions, that produces a wrong answer. Once it has permissions to act, the same trick produces an action.

In plain EnglishAn eval is a fixed set of your own real cases, with known-correct answers, that you score every model and every version against. A benchmark is the public equivalent — and it is largely marketing.

How it actually worksYou assemble perhaps a hundred genuine examples from your own work — real tender clauses, real invoices, real support tickets — and record what a good answer looks like for each. Then any vendor, model or upgrade is scored against that same set. Suddenly the comparison is arithmetic rather than opinion.

In plain EnglishStaff using AI tools nobody approved, on personal accounts, usually with company material pasted into them. Sometimes called BYOAI.

How it actually worksSomeone has a deadline, a consumer tool solves it in four minutes, and a customer contract or a draft board paper gets pasted into a system with no logging, no data-processing agreement and no retention control. There is no malice anywhere in that sequence, which is exactly why policy alone does not stop it.

In plain EnglishWhether your documents and records are in a state AI can actually use. Not whether you have data — everyone has data — but whether it is findable, current, machine-readable and correctly permissioned.

How it actually worksThe obstacles are mundane and universal: scanned drawings with no extractable text, three conflicting versions of the same standard with no way to tell which is live, permissions that stopped reflecting the organisation chart four reorganisations ago. None of it is glamorous and all of it is decisive.

In plain EnglishDeciding — deliberately, and in writing — which decisions AI may make alone, which it may only recommend, and which stay entirely human. Then recording what was decided and why.

How it actually worksEach decision type is classified by consequence and reversibility. Low-consequence and easily reversed can be automated. High-consequence or irreversible keeps a named human accountable. The classification is written down, reviewed periodically and audited — which is what turns "we use AI responsibly" into something a regulator or a customer can actually inspect.