14 terms defined in plain English. Each entry explains what the concept means, why it matters to your business, and what to ask a vendor who throws the term around. No hype, no prerequisites.
A five-rung scale (L0–L4) that describes how systematically an organization measures, governs, and scales AI. L0 is aware but unmeasured. L4 means AI drives owner/operator targets with audited deltas. Most pre-engagement companies land between L1 and L2.
The pre-engagement metric a project's success is measured against. Without a baseline, there is no way to know if AI moved the needle. A baseline is a specific number on your income statement, agreed to in writing before any work starts.
Contract language that ties payment to a measurable, baseline-anchored outcome. It specifies: the metric (one number on your P&L), the baseline (where it sits today), and the measurement window (a date by which you measure). If a vendor won't sign one, that's informative.
The verified change in a metric between baseline and measurement date. "Audited" means signed off by a neutral party — your finance team, an external auditor, or a mutually agreed calculation method. An unaudited delta is a marketing claim.
Large Language Model. A type of AI trained on text data that can generate, summarize, translate, and reason about language. GPT-4, Claude, and Gemini are all LLMs. "Using an LLM" is not an AI strategy — it's a tool choice. What matters is what you measure after deploying it.
An AI system that can take actions on your behalf — browsing the web, writing code, sending emails, querying databases — not just answer questions. Agents introduce audit trail requirements: if an agent makes a business decision, you need to know why and be able to reproduce it.
The instruction or query you give an LLM. A well-engineered prompt is a business asset — it encodes your judgment, your data, and your constraints. Prompts should be versioned, tested, and owned by your organization, not individual employees.
When an LLM generates plausible-sounding but factually incorrect output. The rate of hallucination is a spec-level requirement, not an "that's just how AI works" excuse. Any vendor who can't quote their hallucination rate on a task similar to yours has not measured it.
A repeatable business process where AI handles one or more steps. Examples: quarterly close automation, vendor negotiation copilot, invoice matching. A workflow is only AI-ready when you have a baseline metric, a measurement window, and an audit log.
A set of test cases — inputs plus expected outputs — used to measure how well an AI system performs on your specific task. A vendor without an eval suite on your data does not know if their system works for your use case. "We use GPT-4" is not an eval.
The process of evaluating an AI vendor before signing a contract. At minimum: a measurement clause negotiation, a POC on your actual data, three references in your sector (talk to two), and a SOC 2 Type II report dated within 12 months.
A tamper-evident record of every decision an AI agent makes: what input it received, what it did, when, and why. If you can't export the audit log, you don't own the AI system — you rent it. Required for regulated industries; good practice everywhere.
The elapsed time between starting an AI engagement and receiving the first measurable outcome. Shorter is not always better — a 21-day TTV with a soft measurement is worse than a 90-day TTV with a signed clause. Measure both: TTV and quality of measurement.
Financial Planning & Analysis. The function inside finance responsible for budgeting, forecasting, and variance analysis. FP&A is one of the highest-leverage AI targets because the workflows are data-rich, repetitive, and have a clear P&L connection. Also one of the most risk-averse teams — expect a long measurement clause negotiation.