Plain-language AI terms.

Checking that an invoice, the original purchase order, and the goods-receipt note all agree in quantity and price before an invoice is approved for payment. Done by hand, it's the single biggest source of AP delay; done well, it's mostly automatic.

Money a business owes to its suppliers for goods or services already received but not yet paid for. The AP process — receiving, checking, and approving invoices — is where most manual invoice-matching work happens.

Money owed to a business by its customers for goods or services already delivered but not yet paid for. AR aging (how overdue an invoice is) is a core input into cash flow forecasting.

An AI system that doesn't just answer a question but takes multi-step action toward a goal — checking a database, calling an API, updating a record — with limited human intervention at each step. Distinct from a chatbot, which only responds.

The work of making sure an AI system actually does what its operators intend, especially as the system becomes more capable or autonomous — a research area, not a solved problem, and a big part of why AI agents still need human oversight on consequential actions.

When a model's outputs systematically favor or disadvantage certain groups or outcomes, usually because the data it learned from reflected that same imbalance. Catching and correcting it is a standard part of responsible model development, not an edge case.

An AI assistant embedded directly inside a piece of software to help a person do their existing job faster — drafting an email inside your inbox, suggesting code inside your editor — rather than being a separate destination you have to go visit.

Using a model to automatically flag data points that don't fit the normal pattern — an unusual sensor reading, an off-pattern transaction, a sudden drop in a metric — so a person can look at the handful of things that actually need attention instead of everything.

A defined way for two pieces of software to talk to each other — how your ERP, your website, and a payment processor exchange data without a person copying it by hand. Most 'integrations' are, underneath, one system calling another system's API.

Software that stores and moves candidates through a hiring pipeline. Most ATS tools are good at storage and bad at proactively telling a hiring manager when a requisition has quietly stalled.

Software that performs tasks — recognizing images, understanding language, making predictions — that would normally require human intelligence. Everything else in this glossary is a piece of that broader umbrella.

Turning raw clock-in/clock-out and leave data into something a manager can act on — patterns like recurring lateness, unplanned absences clustering around certain shifts, or a team quietly running under-staffed.

The part of a transformer model that decides which other words or tokens in the input matter most for understanding a given word — how a model knows 'it' refers back to 'the invoice' three sentences earlier, for example.

Projecting how much cash a business will have on hand over a future period, based on expected inflows (payments received) and outflows (payments due). Accurate forecasting depends on invoice and payment data actually being up to date.

Prompting or training a model to work through a problem in visible intermediate steps before giving a final answer, rather than jumping straight to a conclusion — often improves accuracy on multi-step problems.

Software that holds a conversation with a person through text or voice, ranging from simple scripted decision trees to full LLM-powered assistants that can handle open-ended questions.

Running software on servers managed by a provider (AWS, Google Cloud, Netlify, and similar) instead of physical machines a company owns and maintains itself. It's the default for almost any modern web application, including this one.

AI that interprets images or video — reading a scanned invoice, counting items on a shelf from a photo, spotting a defect on a production line. Often paired with OCR (optical character recognition) when the goal is extracting text from an image.

AI software sold as a set of adjustable building blocks (templates, low-code rules, connectors) rather than a finished product. Fast to start with, but closing the gap between the generic template and your actual process is usually left to your own team.

The maximum amount of text (measured in tokens) a language model can 'see' and consider at once — its prompt, any retrieved documents, and the conversation history all have to fit inside it.

A neural network architecture especially good at recognizing patterns in images — edges, textures, shapes — by scanning small regions at a time. The traditional backbone of most computer-vision systems.

Building a model and data pipeline around a company's actual data and actual process from the start, instead of adapting a generic platform to fit afterward. Slower to start, but avoids the 'almost right' ceiling of configured platforms.

The often-manual work of tagging raw data with the correct answer — marking which emails are spam, drawing boxes around objects in photos — so it can be used as supervised-learning training data.

The automated sequence that moves data from where it's created to where it's used — pulling sensor readings into a model, or invoice data into a dashboard — cleaning and transforming it along the way. A forecasting model is only as reliable as the pipeline feeding it.

A type of machine learning that uses neural networks with many layers to learn increasingly abstract patterns — edges, then shapes, then whole objects in an image, for example. It's the approach behind most of today's most capable AI, including LLMs and image generators.

Predicting how much of a product customers will want to buy, at a specific location, over a specific period — the basis for deciding what to reorder and when, before a shelf actually goes empty.

A generative model that creates an image (or other content) by starting from random noise and gradually refining it into a coherent result. The technique behind most modern AI image generators.

A live, continuously updated virtual model of a real physical system — a factory line, a piece of equipment — used to simulate scenarios or catch problems before they happen in the real thing.

Running an AI model directly on a local device — a camera, a sensor, a phone — instead of sending data to the cloud for processing. Faster response and no dependence on an internet connection, at the cost of less computing power to work with.

The rate at which employees leave a company over a given period, whether by choice or not. Tracked alongside attendance and performance data, attrition patterns can point to a specific team or shift worth a closer look.

The central system of record a company uses to run finance, inventory, procurement, and often HR — SAP, NetSuite, Oracle, and similar. Most finance automation projects live or die on how well they integrate with the existing ERP.

Techniques and tools that make a model's decision understandable to a person — why this loan application was flagged, why this part was predicted to fail — instead of the model being a pure 'black box'. Increasingly important wherever an AI decision needs to be justified to a customer or a regulator.

Selecting and shaping the specific pieces of data (features) a model is trained on — turning a raw timestamp into 'day of week', for instance — often the difference between a mediocre model and an accurate one.

Giving a model a handful of examples of the task you want, right inside the prompt, instead of fine-tuning it — a lightweight way to steer behavior without retraining anything.

Taking an existing, already-trained model and continuing its training on a smaller, specific dataset — your support tickets, your invoices — so it performs better on your exact use case without training a model from scratch.

A large, general-purpose model — usually an LLM — trained on broad data, meant to be adapted (via fine-tuning, prompting, or RAG) to many different specific tasks rather than built for just one.

A setup where two neural networks compete — one generates fake content, the other tries to spot the fakes — and both improve until the generated output is convincing. An older but still-used approach to image and synthetic-data generation, mostly superseded by diffusion models for images.

AI that creates new content — text, images, audio, code — rather than just classifying or predicting from existing data. What most people mean today when they casually say 'AI'.

Specialized processors — Graphics Processing Units and Tensor Processing Units — that run the huge number of parallel calculations AI training and inference need, far faster than a general-purpose CPU. Training a large model without them would take impractically long.

The core mathematical process most models use to learn: repeatedly nudging the model's internal values in the direction that reduces its prediction error, a little at a time, until the error is as small as it can get.

Rules and checks built around an AI system to keep its outputs within acceptable bounds — blocking certain topics, catching obviously wrong numbers, requiring a human sign-off above a certain dollar amount.

When a model confidently generates information that's false or made up, rather than admitting uncertainty. It's the main reason AI outputs that touch money, compliance, or customer commitments still need a human review step.

A content management system that stores and delivers content through an API, with no built-in front-end of its own — the website (or app) is built separately and just pulls the content in. It's what lets a blog or glossary get new content without a developer touching the site's code.

A system design where a person reviews or approves an AI's output at a key step, rather than letting it act fully automatically — the standard pattern for anything touching money, compliance, or a customer commitment.

A setting chosen before training starts that controls how a model learns — how fast, how many layers, how much data at once — as opposed to the values the model learns on its own from data.

A way for a website to serve fast, pre-built pages while still refreshing their content automatically on a schedule — daily, hourly, whatever's set — without a developer manually rebuilding and redeploying the whole site every time content changes.

The moment a trained model actually makes a prediction or generates an output on new, real data — as opposed to 'training,' which is when it was learning. When a product 'runs the AI' on your invoice, that's inference.

Turning raw stock and sales data into decisions — what to reorder, how much, and when — instead of just reporting what's currently on the shelf. The difference between a dashboard and something that actually prevents a stockout.

Confirming that an invoice, the payment made against it, and the company's own records all agree. A close cousin of 3-way matching, but focused on the payment side rather than the pre-payment approval side.

A structured map of entities — people, products, companies — and how they relate to each other, which AI systems can query for precise facts. A complement to an LLM's more fuzzy, language-based understanding.

An AI model trained on huge amounts of text to predict and generate language — the technology behind tools like chatbots and document summarizers. Powerful at understanding and producing text, but it doesn't inherently know your business's specific data unless it's connected to it.

A way of building software that learns patterns from data instead of following rules a person wrote by hand. It's the umbrella term that covers most of what people mean when they say 'AI' in a business context, from demand forecasting to fraud detection.

Monitoring progress against the key checkpoints in a project timeline — foundation complete, permits approved, inspection passed — and flagging early when one is at risk of slipping, rather than finding out at the deadline.

The practices and tooling for reliably deploying, monitoring, and updating machine-learning models in production — the ML equivalent of DevOps. Covers things like tracking model versions, watching for model drift, and automating retraining.

When a model's real-world accuracy quietly degrades over time because the world it's predicting has changed since it was trained — customer behavior shifts, a new product line appears — and the training data no longer matches reality.

Multiple AI agents, each handling a different sub-task, working together — and sometimes checking each other's work — to complete something more complex than any one of them could reliably do alone.

A model that can understand or generate more than one type of content at once — text and images together, for example — instead of being limited to just text or just images.

Automatically identifying and categorizing specific things in text — company names, dates, dollar amounts, people — as a structured first step before doing something with that information.

The branch of AI focused on understanding and generating human language — reading an invoice email, classifying a support ticket, summarizing a contract. Most 'AI reads your documents for you' features are NLP under the hood.

A machine learning model loosely modeled on how neurons connect in a brain — layers of simple mathematical units that, combined, can learn very complex patterns. It's the underlying architecture behind most modern AI, including LLMs and image recognition.

Tools that let someone build a workflow or app by connecting pre-built blocks instead of writing code. Fast for a first version or an internal fix; the maintenance and reliability cost catches up once several of them are quietly running part of the business.

The process of getting a new hire set up and productive — paperwork, equipment, introductions, initial training. Onboarding that relies on someone remembering every step is where new hires fall through the cracks in a fast-growing company.

Extracting text from an image or scanned document — a photo of a receipt, a PDF that's really just a picture — so it becomes searchable, editable data instead of a static image.

When a model learns the training data too specifically — including its noise and quirks — and performs worse on new, real-world data as a result. A core risk any model-training project has to guard against.

The system that records a transaction at checkout — in-store or online. POS data is usually the primary input for demand forecasting and inventory intelligence, since it's the closest thing to real-time sales truth.

Using historical data and statistical or machine-learning models to forecast what's likely to happen next — demand next month, a customer's likelihood to churn — rather than just reporting what already happened.

The process of sourcing and purchasing the materials, equipment, and subcontracted work a construction or engineering project needs. Procurement delays are one of the most common causes of a project milestone slipping.

Writing and structuring the instructions given to an AI model to get a reliable, useful output — closer to writing a precise spec than 'talking' to the model. Well-engineered prompts are one reason two products built on the same underlying model can behave very differently.

A document a buyer sends a supplier authorizing a specific purchase — quantity, price, terms — before the goods ship. It's one of the three documents checked in 3-way matching.

A system that predicts what a specific person is likely to want next — a product, a video, a candidate — based on patterns in their behavior and similar users' behavior.

An older neural network architecture designed to handle sequences (like text or time-series data) by keeping a running 'memory' of what came before. Mostly replaced by transformers for language tasks, but still used in some time-series and signal-processing work.

Training a model by letting it take actions and rewarding the outcomes you want — closer to trial-and-error than labeled examples. Common in robotics and game-playing AI, and increasingly used to fine-tune how chat models behave.

An internal request to fill an open role. A requisition can sit open and 'stuck' for weeks without anyone noticing unless something is actively watching for stalled hiring pipelines.

An organization's practices for building and deploying AI fairly, transparently, and safely — covering bias testing, explainability, privacy, and human oversight — rather than treating those as afterthoughts.

A technique where a model looks up relevant information from your own documents or database before generating an answer, instead of relying only on what it learned during training. It's how AI tools answer questions about your specific company data accurately.

A formal request sent to suppliers or contractors asking them to bid a price for specific work or materials, usually as a step before submitting a tender. Tracking every open RFQ and its deadline is core to not missing a bid window.

Software that automates repetitive digital tasks by mimicking exactly what a person would click and type — copying data between two systems that don't talk to each other, for example. Rule-based, not AI on its own, though it's increasingly paired with AI for the steps that need judgment.

Extra inventory kept on hand as a buffer against demand spikes or supply delays. Too little risks a stockout; too much ties up cash in unsold inventory — accurate forecasting is what lets a business carry less safety stock without more risk.

The percentage of received inventory that actually sells within a given period. A low sell-through rate on a SKU is an early signal to stop reordering it before it becomes dead stock.

Search that matches by meaning rather than exact keywords — finding 'invoice delay' results for a search of 'late payment issue' — typically powered by vector embeddings under the hood.

Using AI to determine whether a piece of text — a review, a support ticket, a survey response — expresses a positive, negative, or neutral tone, so a team can spot trouble automatically instead of reading everything by hand.

A unique code identifying one specific product variant — size, color, and all — at one specific location. Demand forecasting works at the SKU-and-location level because the same product can sell completely differently store to store.

Converting spoken audio into text — the technology behind voice assistants, meeting transcription, and voice-driven data entry. Also called speech-to-text (STT).

The moment a product a customer wants to buy isn't on the shelf. Stockouts rarely show up as a clean number anywhere — they mostly show up later, as a customer who quietly bought from someone else.

Training a model on examples that already have the correct answer labeled — this invoice was 'approved', this email was 'spam' — so it learns to predict that label on new, unlabeled examples. The most common way business AI models get built.

Artificially generated data that mimics the statistical patterns of real data, used to train or test a model when real data is scarce, sensitive, or expensive to label.

Keeping tabs on every construction/engineering bid a company has submitted or is preparing — deadlines, required documents, and who owns the next step — so nothing closes without the right paperwork in place.

The number of days between a role opening and an offer being accepted. One of the clearest signals that a requisition has gone quiet is time-to-hire drifting well past a team's normal average.

Splitting text into the small chunks (tokens) — often pieces of words, not whole words — that a language model actually processes. It's why a model's 'context window' is measured in tokens, not words or characters.

The examples a model learns from before it's put to use. The old rule still holds: a model trained on messy, unrepresentative data will make messy, unrepresentative predictions, no matter how sophisticated the model itself is.

The neural network design, introduced in 2017, that made modern large language models possible — its key trick, 'attention', lets a model weigh how relevant every other word in a sentence is to the one it's currently processing. Almost every well-known LLM today is a transformer.

A classic thought experiment, proposed in 1950 by Alan Turing, for judging whether a machine can convincingly imitate human conversation. More a historical benchmark and cultural reference point today than something modern AI products are actually evaluated against.

Training a model on data with no labeled 'correct answer', so it finds structure on its own — grouping similar customers together, for instance, without being told in advance what the groups should be.

Uptime is the percentage of time a system is actually available and working. An SLA (Service Level Agreement) is the contractual promise about that number — and what happens if it isn't met.

The difference between what was expected (a purchase order, a budget, a forecast) and what actually happened (an invoice, an actual spend). Flagging variances with the specific reason — not just 'mismatch' — is what separates useful automation from a dumb rules engine.

A database built specifically to store embeddings and quickly find the ones most similar to a given query — the piece of infrastructure that makes 'search by meaning, not just keyword' and RAG possible at scale.

Converting a piece of text, an image, or other data into a list of numbers (a vector) that captures its meaning, positioned so similar things end up as nearby vectors. It's the mechanism behind semantic search and retrieval-augmented generation.

A way for one system to automatically notify another the instant something happens — a new order, a form submission, a published blog post — instead of the second system having to repeatedly ask 'anything new yet?' It's what makes automatic rebuilds and live notifications possible.

Asking a model to perform a task it was never specifically trained or shown examples for, relying entirely on what it already learned in general training. A key reason modern LLMs feel flexible out of the box.