AI terms, minus the jargon

Agentic AI is when an AI tool doesn't just answer but acts — running jobs on a schedule, using your other software, completing multi-step work on its own.

Bank feeds are the live connection between your bank and your bookkeeping software — transactions flow in daily without manual entry.

A benchmark is a standardized test that measures what an AI model can do — accuracy, refusal rates, hallucination rates — so models can be compared.

Categorization is when the software decides what a transaction was for — office supplies, meals, travel — and files it into your books automatically.

ControlNet is a technique that puts guardrails on image generation — feeding the model a pose, depth map, or edge outline so the output follows a structure you chose.

A dark pattern is interface design that tricks you into something you didn't mean to do — burying the opt-out, making cancellation feel impossible, hiding the fee until it's charged.

Data residency is the country where your information physically sits on a server — and it decides which government's laws can reach it.

Data retention is how long a company keeps what you typed after you've closed the app — days, years, or forever, depending on their terms.

An early termination fee is what a company charges you for leaving a contract before it ends — sometimes up to half the remaining value.

Encryption scrambles your data so anyone who intercepts it can't read it — the question that matters is who holds the key.

GDPR and CCPA are privacy laws — Europe's and California's — that give people the right to know what data a company holds on them and have it deleted.

Generative credits are the monthly meter on AI features in design tools — every image generation, fill, or re-roll spends from a fixed allowance that resets each month.

The GPU is the processor that runs AI models locally, and VRAM is the memory on that card — together they decide what you can run and how fast.

A hallucination is when AI makes something up and states it with total confidence — a fake statistic, a wrong price, a court case that doesn't exist.

HIPAA is the US health-privacy law that governs how patient information is stored, shared, and processed — if your work touches healthcare, your tools must comply.

Indemnification is contract language that says who pays if using the product gets you sued — for AI tools, that usually means disputes over who owns the output.

Inference is the act of running a trained AI model — every time you send a prompt, something somewhere is doing the inference.

An LLM — large language model — is the text engine inside chatbots like ChatGPT and Claude: it predicts what words should come next, which is how it writes, summarizes, and answers.

A LoRA is a small add-on file that teaches an image model a specific style or subject — train one on your products or brand look, and every generation follows it.

A model license is the rulebook for what you may do with an AI model's output — and for local models, whether commercial use is allowed at all.

Model training is when an AI company uses what you type — and sometimes what you upload — to improve its product, so your words literally become part of the thing you're using.

Multi-entity means one software setup handling books for several legal entities — separate companies, one consolidated picture.

A multimodal model handles more than text — images, audio, or screenshots alongside words — in the same conversation.

A node graph is an interface where every step of a process is a visible box you wire together — the opposite of a single button that hides the machinery.

OCR — optical character recognition — is the tech that reads text out of photos and PDFs: it's how receipt-scanning turns a crumpled fuel-stop receipt into a transaction.

Open weights means the model's trained parameters are downloadable, so anyone can run it on their own machine — no account, no permission, no per-use bill.

Opt-in means the company waits for your yes before using your data; opt-out means it starts using your data immediately and waits for your no.

PII — personally identifiable information — is any data that identifies a specific person: name, email, phone number, customer record, tax ID.

RAG — retrieval-augmented generation — is when an AI tool looks up real documents first and writes its answer from them, instead of answering purely from memory.

Reconciliation is checking that what your books say matches what your bank says — every transaction accounted for, nothing missing or duplicated.

Seat pricing charges per person using the tool — and it's rarely the whole cost, because usage often bills separately on top.

An SLA — service level agreement — is a contractual promise about uptime and support: if the vendor misses it, you're owed compensation, not an apology.

SOC 2 is an independent audit of a company's security practices — a passing report means a third party checked, not that your data is guaranteed safe.

A token is a chunk of text — roughly three-quarters of a word — and it's the unit AI companies meter and bill by, especially through APIs.

Upscaling is enlarging an image with AI filling in plausible detail — turning a small generation into something usable in print or full-screen.

Vendor lock-in is what happens when leaving a tool costs more than staying — in money, migration time, or data you can't take with you.

C2PA content credentials are an industry standard that embeds an image's origin history — a tamper-evident receipt showing how it was made and edited.

Zero data retention is a contract-level promise that the company keeps nothing you send once it has processed your request.