The vocabulary behind the writing
Short and full definitions of the terms used across my essays on AI strategy, operating models and evaluation. In an article, the first mention of a term links here.
A
Abstention
A deliberate option for an AI decision system to decline to decide and return uncertain or high-consequence cases to a person, rather than forcing every input into yes or no.
Full definition →Agentic AI
AI systems that pursue a goal by planning, using tools and taking multi-step actions with some autonomy, rather than only answering a single prompt.
Full definition →AGI
Artificial general intelligence: a hypothetical AI that matches human ability across a wide range of tasks. Definitions vary, which is why the practical question is which capability threshold changed a decision.
Full definition →AI adoption
How widely an organization's people use AI tools — licences, active users, use cases. It is rising faster than enterprise value, because adoption is not the same as changing how the work gets done.
Full definition →AI data agent
An AI system that answers business questions by querying an organization's data on someone's behalf. Beyond its answers, the pattern of questions it receives reveals what people need to know.
Full definition →AI incident reporting
Recording anomalous AI behaviour as evidence at the moment it is observed, before root cause is known, keeping what was seen separate from the later explanation of why.
Full definition →AI pilot
A limited trial of an AI use case to test whether it works in practice. A pilot proves one team can succeed; capability starts when others can reproduce the result.
Full definition →AI procurement
Buying AI products and services. A team cannot buy 'reasoning'; it buys a configured system with a route, harness and failure behavior, which is what the evaluation should test.
Full definition →AI transformation
Changing how a company works — its workflows, decisions and operating model — around what AI makes possible, rather than adding AI tools to existing processes.
Full definition →AI-first
A slogan for companies that want AI at the center of how they work. Without changes to workflows and operating model it tends to produce licences and pilots, not transformation.
Full definition →Akl-ı selim
Turkish for sound judgment or common sense. Cahit Arf used it to argue that expertise should help us exercise our own understanding rather than become a conclusion we repeat because an authority said it.
Full definition →API
Application programming interface: a defined way for one piece of software to ask another to do something. Having APIs does not make a business agent-ready; the actions must be described precisely.
Full definition →Attention recovery
The real benefit of persistent AI delegation: not faster execution, but more responsibility safely leaving your active attention so you stop holding a project together in your head.
Full definition →Automation rate
The share of work an AI system completes without a person doing it. A useful headline number that says nothing on its own about the human attention left behind.
Full definition →
B
Becoming the loop
When a human in an AI workflow is so consumed by supplying context, checking and repairing output that they effectively perform the work through the machine, rather than exercising judgment at key moments.
Full definition →Benchmark
A standardised test used to compare AI models. Useful as evidence about someone else's system, but only trustworthy for you when it describes the exact serving route, harness and conditions you will run.
Full definition →Binding constraint
The main reason an AI idea was rejected at a given time, recorded so the decision can be reopened only when that specific constraint actually changes.
Full definition →Blast radius
How far the damage of a failure can spread: which data, systems, people and money an AI action can affect if it goes wrong. Smaller radius means safer delegation.
Full definition →Boundary design
Designing where an AI system's authority starts and stops: which actions it may take, with what data and permissions, and where it must hand back to a person or recover.
Full definition →
C
Cahit Arf
Turkish mathematician (1910–1997) who published a 1959 lecture asking whether machines can think. He is known for the Arf invariant and Hasse–Arf theorem, and his adaptation threshold remains useful.
Full definition →Calibration
How well a model's stated confidence matches reality: if it says 90% on a hundred comparable cases, roughly ninety should be correct. Once probabilities drive actions, it becomes an operating metric.
Full definition →Callable surface
The set of actions an organization exposes so AI agents can invoke them directly, each described precisely enough to be agent-ready: meaning, inputs, permissions, side effects and recovery.
Full definition →Capability discovery
The first phase of generative AI, when attention went to finding out what models could do. The next phase, boundary design, is about where they should act and with what authority.
Full definition →Capability is not authority
A design principle for AI agents: being technically able to perform an action does not mean the agent should be permitted to perform it. Delegation needs gradients of authority.
Full definition →Capability readiness
How ready an organization's actions are to be used by AI agents: whether each action is a real business capability, exposed cleanly, with clear permissions, side effects and recovery.
Full definition →Channel preservation
Escalating or declining an automated interaction early, on purpose, to protect the customer's willingness to use the automated channel again — treating escalation as protection, not failure.
Full definition →Chatbot
A conversational interface, usually for customer service or support, that answers questions in natural language. Judged on one conversation it can look cheap; judged across channels it may not be.
Full definition →Cloud computer
A remote machine an AI agent can use to run tools and applications for long tasks, so work continues without the user's own device. It removes one constraint and exposes others, like permissions and oversight.
Full definition →Computer use
An AI agent operating software the way a person would — looking at screens, clicking and typing — which gives it an execution surface when cleaner integrations or APIs are unavailable.
Full definition →Computing Machinery and Intelligence
Alan Turing's 1950 paper that asked whether machines can think and proposed testing it through an imitation game, now known as the Turing test.
Full definition →Confidence score
A number a model or classifier reports to indicate how sure it is. It is only useful for decisions if it is calibrated, and it never decides on its own what action is permitted.
Full definition →Containment rate
The share of customer interactions that stay inside automation without moving to a human agent. A common chatbot KPI that can look healthy while the channel itself weakens.
Full definition →Context debt
The growing difficulty of knowing what a persistent AI agent currently believes, where that belief came from, and how to correct it, as the agent accumulates memory over time.
Full definition →Control centre
In Cahit Arf's 1959 model of thinking, the component that takes the question from preliminary memory and retrieves and applies material from memory, including deciding to consult outside sources.
Full definition →Cost per accepted outcome
The total cost of getting one AI-assisted result that is actually accepted and used, counting model spend, retries, human review, exceptions and recovery — a truer measure than cost per call.
Full definition →Cost-migration ledger
A small record attached to a meaningful AI change that lists, for each cost dimension, the before, after, delta and owner, so shifted costs stay visible across budgets.
Full definition →Cost-migration test
A check that traces every apparent AI saving across budget owners to see whether the cost was removed or merely moved into review, exceptions, recovery, governance or adoption.
Full definition →
D
Dartmouth proposal
The 1955 proposal for a summer research project at Dartmouth that named the field 'artificial intelligence' and set out the aim of making machines simulate aspects of human intelligence.
Full definition →Decision behavior
How a particular AI system tends to decide across cases — where it leans, abstains or errs. Once embedded in operational thresholds, it becomes a dependency that makes switching vendors harder.
Full definition →Decision debt
The maintenance burden created when an AI system's thresholds and decision boundaries are set once and never revisited, even as data, models, costs and risks change.
Full definition →Decision layer
A dedicated component between AI judgment and organizational action that decides whether the work should continue, separate from the part of the system that does the work.
Full definition →Decisions API
OpenAI's announced decision layer: developers pose a question with a fixed set of possible answers plus text or image context, and the model returns a choice, for classification, routing or an agent's next action.
Full definition →Dependable outcome
A result from an AI-assisted workflow that is actually usable without further repair — the right denominator for measuring the human attention and cost AI automation really consumes.
Full definition →
E
Escalation
Passing a case from an automated system to a person (or a higher level) because it cannot be resolved safely or well. Handled early, it is protection, not failure.
Full definition →Exception handling
The human and system work needed for cases an automated process cannot complete normally. It is a major place where AI savings quietly reappear as cost.
Full definition →
F
Fallback
A backup path used when the first choice fails, is unavailable or is not confident — such as a different model, a simpler rule or a person. A good fallback is designed and tested, not an afterthought.
Full definition →False positive and false negative
The two ways a detector can be wrong: flagging something that is fine (false positive) or missing something that is not (false negative). Their costs differ and must be set per workflow.
Full definition →Fine-tuning
Further training a pre-trained model on specific data so its weights change to suit a task or style. It is one of several ways a system can learn, alongside memory, retrieval and workflow changes.
Full definition →
H
Hallucination
When an AI model produces fluent, confident output that is false or unsupported by its sources. What matters is which errors are unacceptable in a given workflow and how they are detected.
Full definition →Harness
The software around a model that manages context, tool use, sub-agents and the execution environment. It shapes real-world capability, so it must be evaluated with the model, not ignored.
Full definition →Human in the loop
A design where a person reviews, approves or corrects an AI system's work at defined points. In practice it can mean a person with real judgment, or a person who has become the loop.
Full definition →
I
IBIB
A protocol for measuring enterprise AI systems by serving route rather than model identifier. Its authors audited 18 benchmarks and found all scored advertised model names, not the full route.
Full definition →Inference
Running a trained AI model to produce an output for a request, as opposed to training it. Inference cost is the model-side spend that is easy to see but only one part of total AI cost.
Full definition →Intibak kabiliyeti
Turkish for the ability to adapt. In Cahit Arf's 1959 lecture on thinking machines, it is the threshold that matters: not how many problems a machine solves, but whether it can adapt to new ones.
Full definition →
L
Latency
The time an AI system takes to respond. It is one of the constraints, with quality, privacy, cost and recovery, that decides which configured system is best for a workload.
Full definition →Learning loop
A repeating cycle where people try work, see results and improve, which is how expertise builds. AI that compresses tasks can remove the loops that used to train junior workers.
Full definition →LLM
Large language model: an AI model trained on very large amounts of text to predict and generate language, which can be used to answer, summarise, write, reason about and act on text.
Full definition →
M
Model Context Protocol
An open standard, usually called MCP, for connecting AI models and agents to external tools and data sources in a consistent way, so they can discover and call an organization's capabilities.
Full definition →Model routing
Directing each AI request to a particular model or route based on cost, quality, latency or availability, so the model that answers can differ from request to request.
Full definition →
O
Observation vs. explanation
The split between recording what happened in an AI incident (observation) and the current theory of why (explanation), so early facts are preserved while explanations are revised.
Full definition →OpenAI Dots
OpenAI's always-on assistant, described as continuing to work across applications between prompts. It is useful as a test case for persistent delegation: can AI carry a goal, not just complete a task?
Full definition →Operating model
How an organization actually runs: its workflows, decisions, roles, governance and handoffs. AI creates value when it changes the operating model, not when it is bolted onto the old one.
Full definition →
P
Persistent delegation
Handing an AI agent a goal it keeps carrying over time — remembering context, noticing relevant events and acting across applications — rather than completing a single prompted task.
Full definition →Productivity is upstream of value
Individual time savings from AI do not automatically become business value; they do only if the workflow, decisions and economics change downstream to use the freed capacity.
Full definition →Prompt
The instructions and context given to an AI model to produce a response. In persistent, agentic systems, the goal increasingly replaces the single prompt.
Full definition →
R
RAG
Retrieval-augmented generation: a pattern where a model is given relevant documents or data retrieved at question time, so its answers rest on current, specific sources rather than only its training.
Full definition →Reasoning model
A type of AI model built to work through problems in intermediate steps before answering. Like 'foundation model' or 'agent', it names a category, not an explanation of what a system does.
Full definition →Recovery contract
A short written artifact in an AI vendor evaluation that states how failures are detected, contained, reversed and handed back to people — and who is responsible at each step.
Full definition →Recovery distance
The amount of human and system work needed to get from an AI failure back to a safe, correct state — a cost that belongs in the business case, not just the demo.
Full definition →Rejection ledger
A record kept beside the active AI portfolio of serious ideas that did not proceed, with the binding constraint, the evidence date and what would have to change to reopen them.
Full definition →Repeat-channel economics
The effect an automated interaction has on whether the customer will choose automation again — the future-facing half of AI customer service ROI that session metrics miss.
Full definition →Replication
Reproducing a successful result in a new setting by a different team. It is not copying: it means carrying over the capability and learning, so the second attempt costs less in learning.
Full definition →Rescue work
The extra, often unrecorded help a second team needs from the original team to get a transferred capability working. Classifying it shows what the capability really depends on.
Full definition →Return-to-automation rate
The share of customers who choose the automated channel again after an earlier automated interaction — a measure of how today's experience affects tomorrow's automation choice.
Full definition →ROI
Return on investment: the value gained relative to the cost. For AI it is easy to overstate by counting local savings and ignoring supervision, exceptions and recovery.
Full definition →Runtime address
The full description of the system that was actually evaluated — model version, serving route, harness, tools, precision and operating conditions — not just the model's name.
Full definition →
S
Second dataset
The record of questions people ask an AI data agent. Unlike the first dataset, which says what happened, it reveals what people need to know and where decision demand and context gaps sit.
Full definition →Second-team test
A test of whether a different competent team can reproduce a pilot's useful result without inheriting the first team's exceptional conditions or constant expert help.
Full definition →Semantic layer
A shared set of business definitions, metrics, calculations and data relationships that sits between raw data and the people or AI agents that query it, so everyone means the same thing.
Full definition →Serving route
The actual path a request takes to be answered — which model, hardware, precision, routing and fallback rules handled it — as opposed to the model name a vendor advertises.
Full definition →Session economics
The cost and outcome of a single customer contact with an automated channel — what this interaction cost and what it resolved. It is the narrow, current-contact half of AI service ROI.
Full definition →Severity and uncertainty
Two separate axes for rating an AI incident: how serious the effect is, and how well it is understood. Something can be minor but unexplained, or serious and well understood.
Full definition →Sub-agent
A secondary AI agent that a main agent starts to handle part of a task, such as research or a code change, often with its own context and tools. It is part of the harness around the model.
Full definition →Supervision load
The human attention an AI system still consumes — context supply, review, approval, correction and exception handling — measured per dependable, completed unit of work.
Full definition →Support decay
The fall in exceptional help a transferred capability needs from its original team over time — the real measure of whether an AI pilot has become an organizational capability.
Full definition →
T
Threshold policy
The explicit rules that turn a model's probability or score into an action — reject, send to a human, or act autonomously — owned and reviewed separately from the model itself.
Full definition →Token
The small chunk of text, roughly a word or part of a word, that language models read and write. Providers usually price and limit usage in tokens, but token cost is not the cost of a result.
Full definition →Tool call
When an AI model asks to run an external function, such as searching, reading a file or updating a record. A failed tool call can be a very different failure from a wrong answer.
Full definition →Total cost of ownership
The full cost of running an AI system across its life: model and inference spend plus integration, human review, exceptions, recovery, governance and maintenance — not just the price per call.
Full definition →
U
Unhappy path
What happens when an AI system fails, is uncertain or meets a case it was not built for. A buyer should design and test it, because vendor demos show only the happy path.
Full definition →Unit of change
The thing an AI program sets out to change: a tool, a task, a workflow, a decision or the whole operating model. Choosing too small a unit is why adoption rises without value.
Full definition →