An AI proposal may promise a fine-tuned assistant, grounded responses, high accuracy, and human oversight. Each term describes something potentially useful, but the proposal remains difficult to assess until the team knows what those words mean in that application. Fine-tuning may be unnecessary for the information problem, and high accuracy may refer to a test that does not resemble the intended work.
Business leaders do not need to become model researchers to ask useful questions. They need enough shared vocabulary to distinguish capabilities, evidence, and responsibilities. The definitions below focus on distinctions that affect scope, procurement, evaluation, and operation. They can help a team turn a broad proposal into requirements that different specialists interpret consistently.
AI, Machine Learning, and the Complete System
Artificial intelligence is the broad field of systems performing tasks associated with capabilities such as perception, language, planning, and problem-solving. Machine learning is an approach within that field in which models are fitted using data. Deep learning uses neural networks with multiple layers. Generative AI produces content such as text, images, audio, or code.
These terms describe overlapping categories. A language model may use deep learning and generate text; a forecasting model may use machine learning without generating prose. Asking for an AI platform does not specify which capability the business needs. A requirement such as estimating weekly demand gives the technical discussion a more useful starting point.
A model is the computational component that maps inputs to outputs using its structure and parameters. An AI system includes that model and the application, data connections, permissions, interfaces, and operational processes around it. A supplier’s model benchmark is therefore evidence about only part of the proposed service. The building-blocks guide follows these components through a working example.
Training, Inference, and Adaptation
Training fits a model’s parameters using data and an objective. Inference applies the trained model to an input. A company using a hosted model may pay mainly for inference while the provider carries out initial training. Running the application does not necessarily update the model.
Fine-tuning continues training an existing model for a more specific purpose. Prompt engineering changes the instructions, examples, or context supplied to a model during use. These are different interventions: one changes learned parameters, while the other changes the input. Before paying for fine-tuning, establish the specific behaviour it is intended to improve and how that improvement will be evaluated.
A feature is an input representation used by a model, such as customer tenure or recent order frequency. A label is a target associated with an example in supervised learning. Labels might record confirmed fraud or an agreed document category. Not every learning method requires manually labelled examples, and an organisation adopting a pretrained service may need evaluation examples rather than a new training dataset.
Definitions affect data preparation. If churn means no purchase for six months, a model trained against that label predicts a different outcome from one trained to predict a cancelled subscription. The business owner needs to approve that meaning before the team treats the label as an established fact.
Tokens, Context, Embeddings, and Retrieval
A token is a unit processed by a language model, often a word or part of a word. Tokens do not correspond to a fixed number of characters across languages and content. The context window limits how much tokenised material a model can consider in one interaction. A large limit does not guarantee that every detail in a long document will be used correctly.
An embedding is a numerical vector representing an item in a space useful for tasks such as similarity search. A document search application may compare embeddings to find passages related to a question. Similarity is not the same as applicability: a passage about a similar product may be unsuitable for the exact product being discussed.
Retrieval-augmented generation, or RAG, retrieves information and supplies it to a model as context for generation. It can connect an application to current or private sources without retraining the model on every document change. The retrieval process, source quality, and generated answer all need evaluation.
Grounding usually means connecting an output to supplied evidence or another reference. Ask what evidence is used and how support is checked. A source link does not prove that the linked passage supports every claim in an answer. These distinctions help establish whether the proposed solution needs better source management, better retrieval, different instructions, or model adaptation.
Accuracy, Precision, Recall, and Thresholds
For a classifier, accuracy is the share of classifications that are correct. Precision measures how many flagged cases are actually positive. Recall measures how many actual positive cases were detected. Google’s classification metrics guide defines these measures and explains why their usefulness depends on the task and error costs.
Consider a hypothetical invoice review system. Out of 1,000 invoices, 100 contain a defined error. It flags 80 invoices, including 60 with errors and 20 without. It misses 40 erroneous invoices and correctly leaves 880 error-free invoices unflagged.
| Measure | Calculation | Result |
|---|---|---|
| Accuracy | 940 correct classifications out of 1,000 | 94% |
| Precision | 60 erroneous invoices among 80 flagged | 75% |
| Recall | 60 detected errors among 100 actual errors | 60% |
The same system can therefore have high accuracy while missing a substantial proportion of the cases the review process exists to find. A false positive is an error-free invoice flagged for review; a false negative is an erroneous invoice missed by the system. Their costs depend on review effort, payment exposure, and who is affected.
A threshold converts a score into an action or classification. Changing it can alter the balance of missed cases and unnecessary reviews. That is a business decision informed by evidence, rather than a purely technical setting to leave unexplained.
Confidence, Calibration, and Evaluation Data
A confidence score needs a defined interpretation. It may help rank outputs without being a reliable probability of correctness. Calibration concerns how predicted probabilities correspond to observed frequencies. Among comparable cases assigned a probability near 0.8, the event should occur roughly 80% of the time for a well-calibrated model. Google’s machine learning glossary provides a fuller definition. That property needs checking in the relevant setting.
Training data fits the model. Validation data supports development choices such as model selection or settings. Test data provides a separate assessment after those choices. Repeatedly adjusting a model to improve its test result weakens the test’s independence.
Data leakage occurs when development or evaluation uses information that would not legitimately be available for the intended prediction. A churn forecast should not use an account closure field populated after departure. Overfitting describes learning behaviour that performs well on development examples but generalises poorly. Both can make a reported result look more useful than it will be in practice.
Drift, Hallucination, and Explainability
Data drift refers to changes in the input distribution. Concept drift refers to changes in the relationship between inputs and the target. Neither term should be used as a catch-all for every production problem. A broken data transformation or changed business rule may require a different response from retraining.
Hallucination, also called confabulation, describes generated content that is false or unsupported despite appearing plausible. NIST’s Generative AI Profile treats confabulation as a relevant risk. For procurement, the useful question is how unsupported output will be detected and handled in the proposed workflow.
Explainability concerns making a system’s behaviour or outputs understandable to an intended audience. A feature-attribution chart, a customer-facing reason, and a record of application actions serve different purposes. An attribution is not automatically a causal explanation. A generated narrative is not automatically a faithful account of how an output was produced.
Agents, Oversight, and Production Readiness
An AI agent generally uses a model within a process that can select tools or actions over multiple steps towards an objective. The term varies between products. Ask which tools are available, which permissions apply, how the process stops, and which actions require approval. An assistant that drafts an email has different authority from one permitted to send it.
Human oversight means a defined human role in supervising or reviewing the system. Its effectiveness depends on what the reviewer can see, how much time they have, and whether they can change or stop an outcome. Naming a reviewer does not establish that those conditions exist.
Production-ready should refer to an agreed set of conditions: useful performance in the intended workflow, security and privacy controls, operational ownership, monitoring, support, and recovery arrangements. Those conditions differ by use case. A document-tagging assistant needs different evidence from a system influencing access to essential services.
For a supplier discussion, record the agreed meaning beside the requirement. If the proposal promises grounded answers, specify the source collection and how support will be checked. If it promises human oversight, identify the decisions reviewers can change. This gives procurement, delivery, and operations the same reference when they assess whether the promised capability has actually been delivered.
The vocabulary becomes useful when it produces concrete questions in a pilot evaluation: what changes, what is measured, which evidence supports the result, and who can act when the system is wrong. Shared definitions make those decisions easier to discuss and record.
