How to use this guide
Pick the question closest to your current challenge. For each term, we cover the meaning, the decision, the common trap, and what it looks like in practice. Examples are illustrative, not client performance claims.
The bigger picture
These terms are not twenty independent badges to add to a presentation. They describe connected decisions: who needs the offer, what must work, who owns the next action, and what evidence justifies investment. Read them as a sequence of questions rather than a vocabulary test.
CHAPTER 01 ◈ UNDERSTAND THE CAPABILITY
Know what you are buying
1. AI, Artificial Intelligence
The broad field of systems designed to perform tasks associated with human intelligence, including prediction and language tasks.
→ THE DECISION
Which business task genuinely benefits from an intelligent system?
! COMMON TRAP
Calling every automation AI and assuming a label proves a benefit.
→ IN PRACTICE
A deterministic eligibility rule may be safer and cheaper than a model when the logic is already known.
2. ML, Machine Learning
Methods that learn patterns from data rather than relying entirely on hand-written rules.
→ THE DECISION
Is there enough reliable data and a testable prediction problem?
! COMMON TRAP
Training a model before checking data quality or whether a simpler rule performs well.
→ IN PRACTICE
For document classification, compare a trained approach with a rules-based baseline before investing.
3. GenAI, Generative Artificial Intelligence
AI systems capable of producing new text, images, code or other content from prompts and context.
→ THE DECISION
Does generated output improve the customer or employee workflow?
! COMMON TRAP
Confusing fluent language with correctness or operational readiness.
→ IN PRACTICE
Draft a support response for staff review, then measure correction time and error rates.
4. LLM, Large Language Model
A model trained on large amounts of text and other data to process and generate language.
→ THE DECISION
Which model quality, latency, privacy and cost trade-offs suit the task?
! COMMON TRAP
Choosing the largest model because the demo sounded most impressive.
→ IN PRACTICE
Compare models on the same real customer questions, not on one carefully scripted example.
5. NLP, Natural Language Processing
Methods for analysing and working with human language, including classification, extraction and generation.
→ THE DECISION
Is the task extraction, classification, search or open-ended reasoning?
! COMMON TRAP
Using a free-form generator for structured classification without testing a simpler method.
→ IN PRACTICE
Extract application-document fields with confidence checks and a clear human exception path.
CHAPTER 02 → DESIGN THE WORKFLOW
Connect models to useful evidence
6. RAG, Retrieval-Augmented Generation
A pattern that retrieves relevant information from external sources and supplies it as context when generating a response.
→ THE DECISION
Should answers rely on approved current documents rather than model memory alone?
! COMMON TRAP
Assuming retrieval automatically makes an answer accurate or complete.
→ IN PRACTICE
A policy assistant retrieves the latest approved lending rules and cites them, with tests for missing or conflicting documents.
7. SLM, Small Language Model
A comparatively smaller language model that may be suitable for narrower tasks or tighter deployment constraints.
→ THE DECISION
Can a smaller, cheaper or more controlled model meet the actual acceptance criteria?
! COMMON TRAP
Assuming small means inaccurate or large means automatically better.
→ IN PRACTICE
Test a narrow extraction task with a smaller model against a large model using the same evaluation set.
8. MCP, Model Context Protocol
An open protocol for connecting AI applications with tools and external data sources through a standard interface.
→ THE DECISION
Which tools should an assistant be permitted to access and act through?
! COMMON TRAP
Confusing a connection with permission to perform every action.
→ IN PRACTICE
Allow a clinic assistant to retrieve appointment availability, but require explicit authorization before changing records.
9. API, Application Programming Interface
A defined way for software systems to communicate and exchange data or invoke capabilities.
→ THE DECISION
Where should AI fit within the current workflow and system boundaries?
! COMMON TRAP
Bolting a model API onto a broken process and expecting the process to improve.
→ IN PRACTICE
Map the existing enquiry-to-decision flow before deciding which application should call which model.
10. OCR, Optical Character Recognition
Technology that converts text from images or scanned documents into machine-readable characters.
→ THE DECISION
Do we need reliable reading of documents before any model reasoning begins?
! COMMON TRAP
Treating a clean demo scan as proof that real-world documents will work.
→ IN PRACTICE
Test blurry documents, different formats and handwriting, then route uncertain fields for review.
CHAPTER 03 ◎ PROTECT THE BUSINESS
Make failures visible and recoverable
11. PII, Personally Identifiable Information
Information that can identify a person directly or indirectly, depending on the context and applicable rules.
→ THE DECISION
What data may safely enter prompts, logs, tools and third-party services?
! COMMON TRAP
Sending sensitive customer records to a model without understanding retention or access rules.
→ IN PRACTICE
Minimise or mask identifiers before testing a document assistant and document access boundaries.
12. RBAC, Role-Based Access Control
A mechanism that grants system permissions based on assigned roles and responsibilities.
→ THE DECISION
Which users and agents may view, edit or approve which information?
! COMMON TRAP
Giving an AI assistant broader access than the employee it serves.
→ IN PRACTICE
An agent helping a doctor must not inherit clinic-admin privileges merely because it uses the same application.
13. DLP, Data Loss Prevention
Controls intended to detect or prevent sensitive information from leaving approved boundaries.
→ THE DECISION
How will we reduce accidental disclosure through prompts, responses or integrations?
! COMMON TRAP
Assuming a policy document will stop data from being pasted into an external tool.
→ IN PRACTICE
Introduce technical checks for sensitive exports and test them with realistic scenarios.
14. HITL, Human in the Loop
A process in which people review, guide or approve selected machine decisions or outputs.
→ THE DECISION
Which decisions need human oversight and what evidence should reviewers receive?
! COMMON TRAP
Adding a meaningless “approve” button to every output without giving reviewers time or context.
→ IN PRACTICE
Require a qualified reviewer to approve high-impact eligibility recommendations with supporting evidence.
15. EVAL, Evaluation
A systematic test of how an AI system performs on representative tasks, often shortened to “evals”.
→ THE DECISION
What evidence must exist before launch and after model or prompt changes?
! COMMON TRAP
Calling a few successful demos a test suite.
→ IN PRACTICE
Build a representative set of cases, expected outcomes and failure categories, then rerun it after changes.
CHAPTER 04 ◆ MEASURE THE INVESTMENT
Check cost, reliability and value
16. TCO, Total Cost of Ownership
The full cost of buying, building, integrating, operating, supporting and governing a system across its life.
→ THE DECISION
Will the economics still work after integration, reviews, monitoring and support?
! COMMON TRAP
Comparing a model’s token price with a full employee salary as though they are interchangeable.
→ IN PRACTICE
Include implementation, evaluation, exception handling, human review and ongoing operations in the model.
17. ROI, Return on Investment
A measure comparing attributable net benefits with the investment required to produce them.
→ THE DECISION
Which real business outcome justifies the AI initiative?
! COMMON TRAP
Claiming all time saved as cash savings without checking whether capacity was redeployed.
→ IN PRACTICE
Measure cycle time, rework, support effort and quality against a pre-launch baseline.
18. SLA, Service Level Agreement
A commitment describing expected service availability or performance and the remedies or responsibilities attached.
→ THE DECISION
What level of reliability does the workflow require?
! COMMON TRAP
Treating an impressive uptime statistic as a guarantee that critical decisions will be correct.
→ IN PRACTICE
Specify availability, response times, support escalation and fallback for a customer-facing assistant.
19. TTFT, Time to First Token
The elapsed time from a model request until the first generated token is produced, under a defined measurement method.
→ THE DECISION
Will the experience feel responsive enough for users?
! COMMON TRAP
Optimising total response time while ignoring the wait before anything appears.
→ IN PRACTICE
Measure how quickly a support assistant starts answering on ordinary and busy days.
20. TPS, Tokens Per Second
A throughput measure describing how quickly tokens are generated or processed, depending on the stated convention.
→ THE DECISION
Can the system meet volume and latency needs without unsustainable cost?
! COMMON TRAP
Choosing a model solely on headline throughput without testing end-to-end workflows.
→ IN PRACTICE
Benchmark real requests including retrieval, tool calls, output review and peak usage.
THE DECISION LAB
A scenario worth discussing
Imagine a growing clinic network considering an AI assistant for appointment and documentation enquiries. The flashy version is an LLM connected to everything. The useful version starts with the business job: reduce routine administrative time without exposing patient information. NLP and OCR may help read documents; RAG may help answer from approved policies; APIs and MCP may connect permitted services. RBAC, PII safeguards, DLP and HITL control risk. Evals check correctness, and TCO and ROI determine whether the deployment is worth running. If the process itself is unclear, a model will not kindly repair the business logic on its own.
Five questions for the next leadership meeting
- What exact customer or operating problem are we solving, and what is the non-AI alternative?
- What data enters the workflow, who may access it, and what happens when the model is unsure?
- Which representative cases will we evaluate before and after launch?
- What is the full operating cost, not just the model API bill?
- Which measured outcome would justify continuing, expanding or stopping this initiative?
The Parikh Advisory point of view
The acronym is not the achievement. A good decision is. Before commissioning another feature, campaign or tool, identify the assumption, the evidence and the trade-off. That is what we mean by Logic Before Code.
Sources and further reading
Terminology and measurement conventions can vary by industry, platform and contract. These references provide a starting point; define your own calculation boundaries and responsibilities before acting.
- Google Cloud, Generative AI glossary
- Google Cloud, Retrieval-augmented generation
- OpenAI, Evaluation best practices
- OpenAI, Practical guide to building AI agents
- Model Context Protocol, Introduction
Illustrative educational material; not legal, financial or technical compliance advice. Validate important decisions against actual evidence, agreements and applicable requirements.