Relevance AI provides tools for building AI agents and operational workflows.
Relevance AI
Explore features, practical uses and pricing below.
Relevance AI is a platform for building agents that carry out business processes using connected tools and company knowledge. An agent can interpret a request and choose operations; a Workforce connects several specialists on a visual canvas. Teams can reuse tools and knowledge across these configurations rather than rebuilding every process from scratch.
Invent is the platform’s conversational builder. Describe the desired outcome, systems, exceptions and quality requirements, and it can create or revise agents, tools, triggers and Workforces. Its settings distinguish session approvals from project access: accepting a proposed operation does not give Invent permission to use an unavailable integration.
This suits operations and go-to-market teams that know how their recurring work should run. A precise task such as preparing a customer account brief gives the builder clearer requirements than an open-ended instruction to manage sales. Teams still need to decide which evidence is authoritative and who handles an incomplete request.
Tools supply concrete operations such as reading a connected document or updating a record. Knowledge sources organize information through tables, uploaded documents and supported integrations. Preparing this material matters because retrieval cannot repair an obsolete policy or contradictory account record.
The memory documentation distinguishes task-level metadata from persistent memory across runs. Metadata can track fields during a process and support later filtering. In a Workforce, sharing that metadata depends on continuing the same task rather than starting a new one. Treat this as an explicit configuration choice when designing handoffs.
A Workforce can use AI-directed handoffs, mandatory next steps or conditional routing. For example, one specialist can gather account evidence, another draft a brief and another review required sections. A fixed transition is useful when a review must happen every time.
Approval Mode is configured per tool. Autopilot allows that operation to run without a prompt, while human-in-the-loop mode requires explicit approval. This lets a read operation and a CRM update have different review rules within the same process.
Start with an account identifier and approved source list. Configure a retrieval tool to collect the relevant notes, define a brief with goals, unresolved questions and source references, and make the reviewer flag missing evidence. Keep the tool that writes the brief into the CRM behind approval while you compare sample runs with the original records.
The task queue distinguishes in-progress, upcoming, failed, processed and cancelled jobs. Inspect failures as well as final prose. A completed task can still contain a poor interpretation of a source, so review success against the expected business result.
Several agents can pass along the same mistaken assumption. Define stop conditions and a route for conflicting records. Keep connected credentials and permissions aligned with the intended operations, and verify that tools behave correctly when required data is absent.
Version history records draft saves and published states. Changes and restored versions land in draft; the live asset continues until you publish. This supports reviewed iteration without assuming that a rollback immediately changes production.
The public pricing page emphasizes an enterprise offer with custom actions and vendor credits. Confirm your account’s allowances, builder usage and required controls with the vendor. Invent’s documentation specifically notes substantial credit consumption. Use the official website and introduction to scope access before deployment.