Fifty AI and technology spend terms in plain English, each with a question worth asking.
Searches terms, definitions and questions.
The software around a model that turns it into a working agent: instructions, tools, memory, permissions, the execution loop and its checks. Two agents on the same model can differ hugely in cost because of it.
Ask Are we comparing harnesses, or just the models underneath?
A system that uses a model to pursue a goal over several steps. It chooses what to do next, calls tools and repeats until done. Some workflows fix most steps and let the model decide only a few.
Ask Who sets the step and spend limits, and where?
A single access point all AI traffic passes through, handling routing, logging, spend limits and policy. Usually the most reliable source of usage and attribution data.
Ask Does all our AI traffic pass through one, or only some?
The policies, roles and controls that make AI use secure, compliant, effective and affordable. Financial controls belong inside it, alongside legal, security and quality, not on a separate track.
Ask Who can approve new AI spend, and where is it recorded?
Showing a measurable improvement in business outcomes that can be credibly attributed to AI. Spend visibility answers “what did it cost?”. Value realisation answers “was it worth it?”.
Ask What baseline are we measuring against?
Sending many requests to be processed together, with results returned later, usually at a substantial discount. Suits anything not needed in real time, such as overnight classification.
Ask Which of our workloads need an instant answer?
A way to describe how token use multiplies as designs grow: requests (n) × model calls per request (k) × layers of agents (a), written T(n·k·a). From the Tokenomics Foundation.
Ask What are n, k and a for this design?
Reusing work instead of paying for it again. It comes in four kinds. Prompt caching discounts a repeated prompt opening. Semantic caching reuses an answer to a similar question. Application caching reuses identical results. KV caching works inside the serving engine.
Ask Which kind is proposed, and has the simple one been done first?
Time or throughput freed up without any reduction in spend. It only becomes money saved if the time is redeployed to something measurable or a cost is removed. See three kinds of saving.
Ask What will the freed time be used for?
A reduction in spend that shows in budgets or actual results: a contract cancelled, licences reduced, a cost removed. The kind of saving finance can bank with confidence.
Ask Which ledger line moves, and when?
An automatic stop that halts a process when a threshold is hit: a spend cap, a failure count or a maximum number of agent steps. A basic protection against runaway agents.
Ask Does one exist before this goes live?
A contractual promise to spend a set amount over a period, usually in return for a discount. AI usage may have its own commitment or draw down an existing cloud one.
Ask Does AI usage count against our cloud commitment, and should it?
The most material a model can consider in one request, measured in tokens. Filling it is paid for on every call and slows responses, and very long context can bury details.
Ask How much context does each request actually send?
Spend that would reasonably have happened but no longer needs to, such as a planned purchase not made. Legitimate, but easy to overstate, and not a saving.
Ask Where is the documented baseline?
Total cost (tokens, tools, infrastructure, review, rework) divided by the tasks that met the agreed standard. Set against what each outcome is worth, it gives the unit economics.
Ask What does one completed task cost, all in?
Usage drawn from a pool of credits, with each action consuming credits at rates the vendor sets. Credits obscure the real unit price.
Ask What does a credit buy, does it pool across users, and when does it expire?
AI sold inside software subscriptions: per-user add-ons, AI tiers or AI bundled into a licence. Often a significant and poorly visible AI spend, because it sits in the software estate rather than an API bill.
Ask Which of our SaaS renewals now carry an AI uplift?
Energy used per successfully completed task, including infrastructure overhead. More useful than energy per token, because retries use energy too. The resulting carbon depends on where and when the work runs.
Ask Could this workload run in a lower-carbon region or hour?
Repeatable tests that check whether a model or agent meets the required standard on a specific task. Without them, “the cheaper model is fine” is an opinion.
Ask What evidence shows the cheaper option still clears the bar?
When one request spawns several parallel or follow-on tasks, each with its own model calls. One user request can trigger many model calls, and each one is billed.
Ask How many calls sit behind one user request?
Further training an existing model on your own examples so it performs better on a specific task. Costs money upfront, and may need repeating when the base model changes.
Ask Have better prompting and retrieval been tried first?
Applying FinOps, the shared practice of managing technology spend across engineering, finance and the business, to AI. Most disciplines carry over. What’s new is variable consumption, credits, embedded AI and quality-dependent value.
Ask Who owns AI spend end to end?
The FinOps Open Cost and Usage Specification: a common format for billing data, so costs from different providers can be combined. It is being extended to AI, but coverage is still incomplete.
Ask Which of our AI providers supply FOCUS-format data?
An informal label for the most capable models available at any moment, usually large general-purpose “foundation” models. They cost the most per token, and many routine tasks don’t need them.
Ask Has a smaller model been tested on this task?
Applying energy, carbon and water considerations to technology decisions, alongside cost and value. The usage data that drives AI cost analysis can drive carbon analysis too.
Ask Are we measuring footprint from the same data as cost?
Any control on what goes into an AI system, what comes out or what it may do. Spend caps, access rights and data-handling rules are guardrails as well as content filters.
Ask Which of our guardrails are financial?
Output that sounds plausible but is wrong or unsupported. Errors caught cost review time and errors missed carry business risk. Both raise the real cost per outcome.
Ask What is the error rate, and who catches the errors?
A design where a person reviews or approves outputs before they take effect. Often the right control, but it costs time. If every output needs checking, the business case may not stand.
Ask Is review effort included in the cost?
Running a trained model to get an answer. Training is largely the provider’s one-off cost. Inference is the recurring cost you pay, and it grows with adoption.
Ask How will inference volume grow as adoption grows?
Input is everything sent to the model: instructions, documents, history and tool results. Output is what it writes. Output costs more per token, but agents re-send history every step, so input often dominates.
Ask What share of our bill is input?
When something becomes cheaper, people tend to use more of it, so total spend rises. Falling token prices do not guarantee a smaller bill.
Ask Does our forecast assume usage grows as prices fall?
A model trained on vast amounts of text that reads and writes language. It is the engine inside most chat assistants and agents. Model choice is often a major lever on quality, speed and price.
Ask Why this model for this task?
An open standard for connecting AI applications to tools and data, so one connector works across many applications. It makes connecting easy, including to sensitive systems.
Ask What can each connector read and change?
Sending each request to the most suitable model by rules on quality, cost and speed, so simple queries go to a small model and hard ones to a frontier model. A significant way to reduce cost.
Ask What evidence shows the cheaper route still meets the standard?
A model whose trained parameters are published, so you can run it yourself. Not the same as open source: training data may be withheld, and the licence may restrict use.
Ask What does it cost to host, and what does the licence allow?
An attack where content an AI reads, such as a web page, email or document, carries hidden instructions to leak data or take unauthorised actions. A leading security risk for agents that read and act.
Ask What can this agent do if it is tricked?
Reserving fixed model capacity for a period at a fixed price, instead of paying per token. Like a cloud reservation: cheaper if fully used, wasted if not.
Ask What utilisation are we forecasting, and on what evidence?
The minimum acceptable standard for a workload (accuracy, safety, speed or task success), agreed before any optimisation. A saving only counts if the result still clears it.
Ask Who signed off the floor?
Running a model at lower numerical precision so it needs less memory and compute. Can reduce hosting cost, sometimes at the expense of quality.
Ask Have evals confirmed quality still holds?
Finding relevant documents and inserting them into the prompt so the model answers from them. Usually cheaper than fine-tuning, but over-retrieval inflates input tokens, and the stored content needs governing.
Ask How much text is retrieved per question, and is it all needed?
Tokens some models generate internally to work through a problem before answering. Often hidden, but typically billed as output. A short answer can carry a large hidden bill.
Ask How is reasoning effort set, and is it reported?
Running a step again after a failure or a poor result, either automatically or by a user rephrasing. Retries are paid in full, so a low success rate makes a cheap model expensive.
Ask What is the first-time success rate?
Net benefit divided by cost over an agreed period, measured against a baseline of what would have happened without AI. Agree the baseline before deployment; reconstructing it later is weak evidence.
Ask Was the baseline agreed before go-live?
AI used without approval or visibility: expensed personal subscriptions, free tools given company data, unsanctioned add-ons. A cost leak and a security risk, found with familiar SaaS-discovery and expense techniques.
Ask How many AI tools are being expensed today?
Showback reports costs to the team responsible. Chargeback bills them. Both need each cost attributed to a team, application or agent, which for AI usually means a gateway or tagging from day one.
Ask Can we attribute every AI cost to an owner?
Standing instructions that define how a model behaves in an application, sent with every request. A long one is a recurring cost. A stable one can be discounted through caching.
Ask How long is it, and does it change between calls?
The unit a model reads and writes, often a word or part of one. Most APIs bill per million tokens. Models split text differently, so the same text costs different amounts on different models.
Ask Are we comparing prices on the same text?
The economics of AI end to end: producing tokens, consuming them, and the value that results. The Tokenomics Foundation, a Linux Foundation body formed in 2026, is developing shared frameworks and terms.
Ask Is this decision about production, consumption or value?
The full cost of an AI capability: models, platforms and licences, infrastructure, integration, operations, people, governance and rework. Comparing API prices alone is a common error in AI business cases.
Ask What isn’t in this number?
Sorting AI benefits into distinct categories (revenue, capacity, cost avoidance, cashable savings, quality, risk) so each is measured properly. Overlapping categories double-count, and inflated business cases usually do.
Ask Could any benefit here be counted twice?
No terms match . Try a shorter word, or clear the search and browse A to Z.
No single team usually sees all of it. Put all six on one page before deciding anything.
Per-token bills from model providers
Usually seen by: Engineering, FinOps
Model services billed through a hyperscaler, often drawing down a commitment; reserved capacity
Usually seen by: FinOps, cloud team
Per-user AI add-ons, AI tiers, credits bundled into subscriptions
Usually seen by: ITAM/SAM, procurement
GPUs, hosting, operations staff, open-weight licence terms
Usually seen by: Infrastructure, ITAM
Human review, rework, integration, governance effort
Usually seen by: Business owners (rarely tagged as AI)
The footprint of AI workloads
Usually seen by: Sustainability, GreenOps
An AI question is often also a licensing, cloud-commitment or sustainability question. Answering it usually needs these teams working together.
An AI tool saves ten analysts two hours a week each.
Report them separately. They carry very different weight with a CFO. Figures are illustrative.
Several frameworks here come from the Tokenomics Foundation, including Big-T notation, value classification and the production, consumption and value model. FOCUS is a FinOps Foundation specification.
All definitions are paraphrased and simplified. Check current provider documentation before making commercial decisions.
Synyega is an independent technology-spend consultancy. We help organisations make defensible decisions across software, cloud and AI spend.
We don’t resell, we don’t take vendor commission, and we’re never paid on consumption. That is why we can be straight about what these terms mean for your spend.