What is Relevance AI?
Relevance AI is a platform for building and operating AI agents without starting from a general-purpose orchestration framework. Its core objects are Tools, Agents, Knowledge, and Workforces. A Tool is a reusable workflow or action. An Agent combines instructions, models, knowledge, and tools. A Workforce coordinates multiple agents with different roles. Invent lets a builder describe changes in natural language, while visual and code-capable steps handle more explicit implementation.
The platform is aimed at business automation: researching accounts, updating a CRM, triaging support, preparing content, qualifying leads, sending messages, enriching records, or coordinating work across connected applications. Integrations and generic API calls reduce initial engineering, but the product does not make a business process correct. Every external read and write still needs authorization, validation, failure handling, and an accountable owner.
Marketplace templates can accelerate a prototype. Treat them as examples, not audited applications. Inspect every instruction, credential, trigger, tool, destination, knowledge source, model, field mapping, default recipient, and write action. Clone into a non-production project, replace secrets, and test with synthetic data before connecting a live inbox, CRM, data warehouse, messaging channel, or customer system.
Designing an agent that can be trusted
Start with a narrow task and a measurable contract. Define accepted inputs, required evidence, permitted tools, fields that can be read, fields that can be changed, prohibited actions, completion criteria, maximum steps, timeouts, retries, duplicate prevention, escalation, and rollback. Use the least-privileged service account available and separate development, staging, and production credentials.
High-risk tools should not be available merely because an agent might need them. Put approvals before sending external email, publishing, changing CRM ownership, deleting records, issuing credits, creating contracts, or executing code. Validate structured output against a schema and re-check policy outside the model before an action. Idempotency keys and deduplication are essential when timeouts can make the agent retry an operation that already succeeded.
Knowledge can ground answers in uploaded or connected information, but retrieval is not proof. Measure whether the correct source was retrieved, whether the answer is supported, and whether access controls prevent one user or agent from seeing another project's data. Defend against prompt injection inside emails, webpages, tickets, attachments, and documents by treating retrieved text as untrusted data rather than instructions.
Workforces add coordination, not certainty. Multiple agents can repeat a mistake, pass malicious context, create loops, and make a failure harder to trace. Limit delegation depth and total tool calls; maintain a complete event trail; test handoff, disagreement, and escalation; and assign one human owner for the final outcome.
Actions, Vendor Credits, and current plans
Relevance AI's current billing separates Actions from Vendor Credits. One Tool execution is one Action, even if the Tool has several internal workflow steps. The billing guide explicitly says a failed Tool still counts. Vendor Credits cover model and vendor usage without a stated markup; paid plans can bring their own LLM API keys to bypass that credit pool, though the external provider then bills and governs the traffic directly.
The official pricing documentation checked August 9, 2026 lists Free at $0 with 200 Actions per month and a one-time $2 or 1,000 Vendor Credit allocation. Pro is $19 per month on annual billing or $29 month-to-month, with 30,000 Actions and $240 Vendor Credits yearly, or 2,500 Actions and $20 monthly, plus two build users, scheduling, chat, escalations, and BYO LLM.
Team is $234 per month at the annual rate or $349 monthly. It lists 84,000 Actions and $840 Vendor Credits per year, or 7,000 Actions and $70 monthly, five builders, 45 end users, five shared projects, calling and meeting agents, A/B testing, analytics, and priority support. Enterprise is custom and adds controls such as SSO, RBAC, audit logs, evaluations, work-hour controls, and broader user/project capacity.
Current billing docs price top-ups at $80 per 1,000 Actions and $20 per 10,000 Vendor Credits. Included Actions reset at renewal; purchased Action top-ups roll forward, while Vendor Credits roll over indefinitely as long as the subscription remains active. Older marketing pages may show legacy amounts, so verify the live checkout, contract, included credits, renewal, concurrency, and rollover before purchase.
Cost models should include scheduled runs, triggers, agent-to-agent calls, failures, retries, test traffic, evaluations, model tokens, third-party API fees, human review, and incident recovery. Add hard spend ceilings and alerts before enabling a recurring or externally triggered agent.
Privacy and security
Relevance AI's security documentation states SOC 2 Type II compliance, GDPR alignment, TLS 1.2+ in transit, AES-256 at rest, and selectable regions. It says customers own uploaded data and that customer data is not used for model training or service improvement unless a specific partnership agreement applies. It also says Tool inputs and outputs are not logged by Relevance AI, while steps may send data to the selected model or integration vendor.
Agent conversations are stored inside the selected project region and can be deleted; Knowledge is stored until the customer deletes it. The security page describes OpenAI and Anthropic API paths as not logged or trained by those vendors under the relevant arrangements. Confirm the exact model, region, API account, retention, subprocessors, DPA, support access, backups, deletion propagation, and incident terms rather than generalizing one row to every provider or integration.
Enterprise buyers can request security documentation and assessments under NDA. Obtain the current SOC report and scope, penetration evidence, subprocessor list, architecture, business continuity, key management, vulnerability process, breach notice, audit-log coverage, and contractual data restrictions. The general privacy policy also covers website, account, marketing, analytics, and support personal information, which is a different data category from project content.
Verdict
Relevance AI is a capable option for teams that want to move from an agent prototype to scheduled, connected business workflows without building every orchestration layer. The platform's Actions and Vendor Credits make two major cost drivers visible, and its security documentation is stronger than a generic “enterprise-ready” badge.
The deciding factor is governance, not how quickly a demo can send an email. Pilot one bounded process with synthetic or low-risk data. Measure task success, unsupported claims, wrong tool selection, duplicate actions, escalations, human correction, cost per accepted outcome, p95 duration, vendor failures, and recovery. Expand permissions only after evidence shows the agent is safer and more economical than the controlled alternative.
