The technical term.
What it means in your company.

Connecting an AI capability to the data, software and steps of existing work. The useful result arrives in the right tool, with a status that is easy to understand. Integrating CRM and ERP.

A system that uses a model to choose steps or tools based on a request and on what it observes. Its autonomy depends on the actions it is actually allowed to take, not only on its prompt. Agent, assistant or workflow.

A sequence of operations defined in advance: receive, check, prepare, pass on. Some steps may call a model; not all of them need one. Choose the right level of control.

An approach where engineering works close to the users to understand their tools, build a solution and support it as it goes live. The term does not refer to a Saphan certification or a vendor partnership. Understand the approach.

A model trained to process or produce language. It is one component of the application, distinct from its documents, its accounts and the software it accesses. Choose the levers of customization.

Using an already trained model to produce a result from an input. Running a model on your machine does not mean training it or giving it a lasting memory.

Processing run on a workstation or infrastructure that is named precisely. A locally installed interface may nevertheless call an external service: you have to look at the whole chain. Installing a local AI.

An architecture that combines several places or modes of processing, for example an internal corpus and an external model for certain tasks. The data sent and the routing rules remain to be defined. Local, cloud or hybrid.

Searching a corpus for useful passages, then using those passages as context for the answer. RAG does not necessarily change the model’s weights and does not automatically apply document access rights. Sources and access.

A numerical representation used, among other things, to bring together texts with a similar meaning. Closeness in vector space proves neither that a passage is true nor that a user has the right to see it.

Changing parameters based on examples, to aim for a targeted behavior or format. Whether it makes sense is weighed against a better prompt, against RAG and against the tools available. Choose before you train.

An adaptation technique that uses additional parameters that are small compared with the full model. Reducing the training effort does not, on its own, show better quality for your line of business. What the adaptation must change.

An exchange contract with a piece of software: available operations, expected data, authentication and responses. An AI connection uses this contract without removing the software’s restrictions.

A protocol that lets an AI application use tools, resources and instructions exposed by a server. It replaces neither permissions nor control over actions. MCP in the enterprise.

A person steps in at a defined point in the flow. It is useful if the person sees the proposal, its sources and the differences before the action concerned. Roles and permissions.

Comparing a system against explicit criteria on representative cases, then running it again after a change. A result applies to this test scope, not to every possible use. Prepare a replacement.

Latency describes a wait; throughput describes a quantity processed per unit of time. For a user, the time to a usable result and the amount of correction also count. Sizing against a task.

Information that makes it possible to understand the state of a flow: steps, errors, delays and observed results. Useful logs link the actions without needlessly copying documents and secrets.

The prepared ability to swap a component along with its interfaces, configurations and exportable data. A compatible API does not prove that the business result will be identical. Check a switchover.

Going back to a software version and restoring data are two different operations. Reverting a configuration does not necessarily undo a migration or an action that has already run. Prepare a consistent recovery.