One agent per workflow. Twelve workflows.

An AI agent is software that completes a defined piece of work by reading data, reasoning about it, and taking actions in your systems. Meridian's agents are each scoped to one workflow: resolving an HR case, screening a candidate, packaging audit evidence, testing a payment for duplicates. Each reads only the data it is permitted to, logs every step, and hands consequential decisions to a named person.

A chatbot answers. An agent finishes the work. Meridian's agents open and close cases, book interviews, prepare reconciliations, and produce evidence packages inside your systems of record. They are also narrower by design: each does one job, against defined data, under an approval policy, with a measurable outcome the Registry reports.

Each agent has a policy naming which actions are consequential: moving money, changing pay, rejecting a candidate, sending a document externally. When the agent reaches one, the action is held at the Gateway and routed to the approver you named, in Slack, Teams, email, or the workspace. The agent cannot proceed until a person approves, edits, or rejects. Every decision is logged.

The ones you choose. Meridian is model-agnostic and supports frontier models from leading providers, including Anthropic and OpenAI models. HR and finance domain reasoning is routed to Meridian's own tuned models. You can pin a provider and region per workspace and change it later; evaluation sets re-run so you see any behavior change first.

Weeks, not quarters. Design partners connected their first systems in week one and ran a live pilot of the first agent in weeks 3 to 5. Most were in production on two or three agents within a quarter. Connectors, approval templates, and evaluation sets are prebuilt; the time goes to tuning scope against your real data.