# Relevance AI

> Relevance AI is a platform for building, deploying, and governing AI agents that carry out business tasks without code. You compose tools (multi-step workflows that call LLMs, APIs, and integrations), attach them to agents with instructions and knowledge, and group agents into Workforces that hand work to each other. It is billed on two meters, Actions for task runs and Vendor Credits for the underlying model spend, and is sold as a free tier, two self-serve plans, and a sales-led Enterprise plan.

- Category: GTM Engineering (https://saastracker.org/categories/gtm-engineering)
- Website: https://relevanceai.com
- Starting price: Free (200 Actions per month); Pro from $19 per month billed annually
- Free plan: 200 Actions per month, a one-time grant of 1,000 Vendor Credits, one build user, one project
- Free trial: No time-limited trial; the free plan serves that purpose
- Founded: 2020, HQ: Sydney, Australia, Ownership: Independent, venture-backed (OnSearch Pty Ltd trading as Relevance AI)
- Profile last reviewed: 2026-08-23
- Canonical profile: https://saastracker.org/products/relevance-ai

## Overview

Most go-to-market automation tools are workflow engines that occasionally call a language model. Relevance AI inverts that: the primary object is an agent, a named worker with a role, a set of tools it may use, knowledge it can read, and a policy for when it should escalate to a human. The workflow builder is still there, and it is the part most teams use first, but the pitch is that a research task or an inbound triage task becomes a persistent role rather than a scheduled job.

The company calls the result an AI Workforce, and the vocabulary is deliberate. Agents have job titles in the interface (BDR Agent, Research Agent, Support Agent). Workforce is the feature that lets a manager agent delegate to specialists. The platform pieces underneath are the ones a team would otherwise assemble from separate vendors: an agent builder, an orchestration layer, a job queue, an LLM router that picks the cheapest model that passes, a tracing view, an evals framework, an MCP gateway that governs which tools an agent may reach, and a shared Context Layer holding tone of voice and business knowledge so it is not restated in every agent.

Relevance AI is an Australian company, founded in Sydney in 2020 and originally built around vector search and unstructured data before pivoting to agents in 2023. It raised a $24 million Series B led by Bessemer Venture Partners in May 2025, bringing total funding to roughly $37 million. Since then the product has moved noticeably upmarket. As of August 2026 the public pricing page shows only the Enterprise plan with a Talk to sales button; the Free, Pro, and Team plans still exist and are still documented, but you now have to find them inside the docs or the signup flow rather than on the marketing site. A small business can still buy it, but the vendor has stopped merchandising to them.

The second thing to understand before buying is the meter. Since September 2025 the platform charges Actions (one per tool run, including runs that fail) and Vendor Credits (raw model spend, passed through at wholesale with no markup and able to be bypassed entirely by bringing your own API keys). This is more honest than a single opaque credit, and it is also the single most common source of billing surprise, because a chatty multi-agent workflow burns Actions faster than most teams estimate and retries are billable.

## How it works

1. You start in a project and build a Tool, which is a chained workflow of steps: LLM prompts, API calls, code, data lookups, and conditional branches. Tools can be run directly, exposed as an API endpoint, embedded, or handed to an agent as a capability. Most teams' first useful build is a research or enrichment tool that takes a company domain and returns a structured summary.

2. An Agent is created by writing instructions (its role and constraints), selecting a model, attaching tools, and attaching Knowledge, which is the platform's retrieval layer over uploaded files, synced Google Drive, Notion, SharePoint, or Confluence content, and structured tables. Approval gates let you require human sign-off before specific actions, with thresholds you define, such as any refund over a dollar amount.

3. Agents are put to work through triggers rather than manual runs: inbound email, a Slack or Microsoft Teams message, a calendar event starting or ending, a schedule, a webhook, a form, or a CRM record change. Enterprise plans add triggers on Salesforce, Snowflake, and Zendesk. A Workforce groups several agents so a manager agent can delegate subtasks and collect results.

4. Operationally, the Tasks view shows every run with its approvals, escalations, and errors; tracing shows the full step-by-step execution including which model was called and what each step cost; and the Evals framework (Enterprise) scores agent versions against scenarios with an LLM judge so a regression is caught before it ships. Analytics reports Action consumption, Vendor Credit spend, and concurrency against your plan's limits.

## Best for

Revenue and operations teams that want AI agents doing repeatable research, enrichment, outreach, and triage work under supervision, and that have someone willing to treat agent building as an ongoing job rather than a one-week setup. Strongest where the work is judgment-heavy and unstructured, which is exactly where a rules-based automation tool stalls.

## Not the right fit for

- Teams that want a deterministic automation tool; if the task is if-this-then-that with no reasoning involved, an agent adds cost, latency, and a failure mode you did not previously have.
- Buyers who need a published price for the plan they will actually run on; anything above the Team plan is quote-only and the marketing site no longer lists self-serve pricing at all.
- Agencies running many client environments, since the workspace and credential model makes per-client isolation awkward and duplicating projects multiplies the seat and plan cost.
- Organizations without an owner for the output; agents that write to a CRM or send email need someone reviewing quality weekly, and unowned agents degrade quietly.
- Teams whose core need is a customer service inbox or a full CRM; Relevance AI acts on those systems but does not replace them.

## Features

### Agents

The primary object: a named worker with instructions, tools, memory, and escalation rules.

- **Instruction-based agent builder**: An agent is defined by a written role, constraints, and success criteria rather than a flowchart, which is why non-engineers can author one and why prompt discipline matters more than diagram skill.
- **Tool attachment**: Any tool you have built, cloned from the marketplace, or connected through an integration can be granted to an agent individually, so capability is scoped per agent rather than per workspace.
- **Knowledge and retrieval**: Uploaded files, structured tables, and auto-syncing connectors for Google Drive, Notion, SharePoint, and Confluence give agents grounded context instead of relying on model training data.
- **Approval gates**: High-stakes actions pause for human sign-off at thresholds you set, for example any refund above a dollar limit, which is the mechanism that makes autonomous write access tolerable.
- **Escalation rules**: Agents can be told what they are not allowed to decide, routing the case to a person with the context attached rather than guessing and being wrong at scale.
- **Model selection and LLM router**: Choose a specific model per agent from OpenAI, Anthropic, and Google, or let the router pick the cheapest model that still passes your evaluation bar for that task.

### Tools and workflow building

The no-code chain builder that most teams actually spend their time in.

- **Multi-step chains**: Steps for LLM prompts, API requests, JavaScript or Python, data transforms, and branching conditions, composed visually and testable step by step with live output.
- **Bulk runs over tables**: Run a tool across a list of rows, which is the workhorse pattern for enrichment: a thousand domains in, a structured column set out, with per-row failures visible.
- **Tools as API endpoints**: Every tool can be exposed as an HTTP endpoint or an embeddable component, so a workflow built here can be called from your own app or from another automation platform.
- **Marketplace templates**: Pre-built agents, tools, and Workforces can be cloned into your project as a starting point, which shortens the gap between signup and something that runs.
- **Structured output enforcement**: Steps can be constrained to return a defined schema, which is the difference between an agent that populates a CRM field correctly and one that writes a paragraph into it.

### Workforce and orchestration

Multiple agents coordinating, with a queue underneath so failures do not vanish.

- **Multi-agent teams**: A manager agent decomposes a request and delegates to specialist agents, collecting their results, which suits long research tasks better than one agent with twenty tools.
- **Job queue and retries**: Runs are queued rather than fired and forgotten, so a failed step retries instead of disappearing, and concurrency limits are visible per project.
- **Trigger library**: Email, Slack, Microsoft Teams, Google Calendar and Teams Calendar meeting start and end, schedules, webhooks, and forms, with Salesforce, Snowflake, and Zendesk triggers reserved for Enterprise.
- **Pause and resume controls**: Individual triggers, whole agents, or an entire Workforce can be paused in bulk without deleting configuration, which matters the first time an agent misbehaves in production.
- **Context Layer**: Tone of voice, business definitions, and shared knowledge are defined once at the project level and inherited by every agent, instead of being copy-pasted into each set of instructions.

### Governance, evaluation, and observability

The controls that separate a demo from something allowed near production data.

- **Evals framework**: Scenario generation plus an LLM judge scores each agent version against a benchmark, so version three failing where version two passed is caught before release. Enterprise only.
- **Agent performance observability**: Continuous LLM-judge evaluation of live production runs at a sampling rate you set, which extends evals from test sets to real traffic.
- **Full run tracing**: Every step of every run is inspectable with inputs, outputs, model used, and cost, which is the only practical way to debug why an agent produced a strange answer.
- **MCP gateway**: Tool access across integrations is governed centrally through Model Context Protocol rather than granted ad hoc per agent, giving one place to revoke a connection.
- **Cost and concurrency analytics**: Dashboards show Action consumption, Vendor Credit spend by project and by user, queued runs, and live slot usage against plan limits.
- **OpenTelemetry export**: Audit logs and execution traces stream to customer-owned destinations, so agent activity lands in the same observability stack as the rest of the estate. Enterprise only.
- **SSO, RBAC, and audit logs**: Role-based access control with SAML single sign-on and audit trails, plus PII masking and a no-training-on-your-data commitment, all gated to Enterprise.

## Use cases

- **Two-person sales team at a B2B software company**: Inbound demo requests arrive with nothing but a work email, and whoever picks them up spends fifteen minutes researching the company before the call, or skips it and runs a generic call. Outcome: A research agent triggered by the form fires on submission, enriches the domain, reads recent company news and the careers page, and posts a five-line brief plus three suggested questions into Slack before the confirmation email lands. Prep time drops to reading a Slack message.
- **RevOps lead at a 60-person company cleaning a CRM**: Thirty thousand accounts have inconsistent industry, headcount, and ICP-fit values entered by five people over four years, and no rules-based enrichment vendor gets the segment definitions right because they are specific to the business. Outcome: A tool runs in bulk over an exported table, applies the company's own written ICP definition through an LLM step with enforced structured output, flags low-confidence rows for human review rather than guessing, and writes back through the HubSpot integration. The judgment calls stay with a person; the eighty percent that are obvious do not.
- **Customer success manager covering 200 accounts alone**: Renewal risk shows up in support tickets and meeting notes, but nobody reads across both systems until the account is already quiet for a month. Outcome: A post-call agent triggered at meeting end summarizes the call, extracts commitments, checks recent Zendesk tickets for the same account, and escalates to a human when it detects a churn signal, with the approval gate ensuring no customer-facing message goes out unreviewed.
- **Founder running outbound without a BDR**: Sequences get written once, sent to a poorly researched list, and produce nothing, but hiring a BDR is not affordable at current revenue. Outcome: An outbound prospector agent researches each account against a written qualification bar, drafts a first-touch message grounded in something specific it found, and queues it for one-click approval. The founder reviews forty drafts in twenty minutes instead of writing four in an hour, and the Actions meter makes the per-message cost explicit enough to compare against a contractor.

## Pricing

Two meters plus a plan fee. Actions are consumed one per tool run, including runs that fail or retry. Vendor Credits cover the underlying LLM and third-party model spend, passed through at wholesale with no markup, and can be bypassed entirely on paid plans by connecting your own OpenAI or Anthropic API keys. Plans set the included Actions, included Vendor Credits, number of build users, and number of projects. This structure took effect on 8 September 2025 and replaced the previous single-credit model; the Business plan was discontinued at the same time.

- **Free**: $0 per month. 200 Actions per month; 1,000 Vendor Credits granted once at signup, not monthly; One build user and one project. Enough to build and test one agent end to end. It is not enough to run anything on a schedule, because a single bulk run can consume the monthly allowance.
- **Pro**: $19 per month billed annually, $29 month to month. 2,500 Actions per month; About $20 of Vendor Credits per month, or bring your own API keys; Action and Vendor Credit top-ups available. The realistic starting point for a solo operator or a small team running two or three agents at modest volume.
- **Team**: $234 per month billed annually, $349 month to month. 7,000 Actions per month and about $70 of Vendor Credits; Five build users, additional end-user seats, five projects; The highest published price before quote-only territory. The jump from Pro to Team is more than tenfold with under three times the Actions, which is the sharpest edge in the pricing and the point where many teams stall.
- **Enterprise**: Custom annual contract, quote only. Custom Actions and Vendor Credits, unlimited agents, users, projects, and Workforces; Evals, agent performance observability, A/B testing, and OpenTelemetry export; SSO, RBAC, audit logs, PII masking, data residency, 99.9 percent uptime SLA, dedicated account manager; Salesforce, Snowflake, and Zendesk triggers, plus calling and meeting agents. This is the only plan shown on the public pricing page as of August 2026. Everything that makes agents safe to run unattended, evals and observability included, sits behind it.

Add-ons:

- Additional Actions ($80 per 1,000 Actions): Purchased top-ups roll over to the next cycle; Actions included with a plan reset at renewal.
- Additional Vendor Credits ($20 per 10,000 Vendor Credits): Vendor Credits roll over indefinitely for as long as the subscription stays active.

Billing notes:

- An Action is charged per tool run whether or not the run succeeds, so failing agents and retries are billable; this is the most frequently reported billing surprise.
- Vendor Credits carry no markup and can be avoided completely by supplying your own model API keys on any paid plan, which meaningfully changes the economics for heavy users.
- The Pro to Team gap is severe: $19 to $234 per month annually for 2,500 to 7,000 Actions, with the seat and project increase doing much of the work in justifying it.
- Top-up Actions at $80 per 1,000 work out to eight cents per run, so overage is expensive relative to the included allowance on Pro.
- The Business plan was discontinued on 8 September 2025; anyone reading an older comparison that quotes a $599 or $199 Business tier is reading stale information.
- Salesforce, Snowflake, and Zendesk triggers, evals, SSO, and audit logs are Enterprise-gated, so several things a mid-sized company would consider baseline require a sales conversation.

Value assessment: Priced against what it replaces, the Pro plan is cheap: $19 a month plus metered model spend for work that would otherwise be a contractor's afternoon. The problem is the shape of the curve above it. Team at $234 a month gives under three times the Actions for more than ten times the price, and the governance features that make agents safe to leave running (evals, observability, audit logs, SSO) are not purchasable at any published price. A small business gets real value from the free and Pro tiers for supervised, human-in-the-loop work. A company that wants agents operating unattended on customer data is buying Enterprise, and should assume a five-figure annual commitment and a procurement cycle.

## Strengths

- The agent abstraction is genuinely well built: instructions, scoped tools, knowledge, approval gates, and escalation rules are all first-class rather than bolted on.
- Tracing and cost visibility are better than most competitors, showing per-step model calls and per-run cost, which turns debugging from guesswork into reading a log.
- The two-meter pricing separates platform cost from model cost honestly, and bring-your-own-keys removes the markup that most agent platforms quietly charge.
- Approval gates and escalation are the right primitives for letting an agent touch a CRM or an inbox without an incident.
- Broad reach through native integrations plus roughly 2,000 more services via a Pipedream connection, so the tool inventory is rarely the blocker.
- The evals and LLM-router combination lets you run on the cheapest model that still passes a quality bar, which is a real cost lever at volume.
- A marketplace of clonable agents, tools, and Workforces shortens the distance from signup to something that actually runs.

## Limitations

- Cost is hard to forecast. Actions are charged per tool run including failures and retries, and a multi-agent Workforce fans out into far more runs than a first-time buyer estimates.
- The public pricing page shows only Enterprise as of August 2026; Free, Pro, and Team still exist but you have to dig through documentation to find their terms, which is a deliberate move away from self-serve buyers.
- The Pro to Team step is punishing, roughly $19 to $234 per month for under three times the Actions, and there is no longer a Business tier in between since it was retired in September 2025.
- Governance essentials are Enterprise-gated: evals, agent performance observability, audit logs, SSO, RBAC, and Salesforce, Snowflake, and Zendesk triggers are all out of reach of a self-serve buyer.
- The workspace and credential model makes agency use awkward, with agencies reporting that isolating multiple client environments cheaply is not really possible.
- Complexity arrives quickly. Simple builds are genuinely no-code, but custom API handling, nested conditionals, and multi-agent orchestration demand real technical skill and iteration.
- Knowledge storage limits on the lower plans fill up fast with document-heavy agents, and hitting the ceiling is not obvious until retrieval quality drops.
- It is not a system of record and has no customer service inbox of its own, so it always sits alongside a CRM and a helpdesk rather than replacing either.

## Comparisons

- **Relevance AI vs Gumloop**: Gumloop is the closer thing to a visual automation canvas: node-based flows, easier to reason about, faster for a linear pipeline like scrape, enrich, write to a sheet. Relevance AI is stronger when the unit of work is an agent that decides what to do next, delegates to other agents, and asks a human before acting. If your process is a diagram, Gumloop is the calmer purchase. If your process is a job description, Relevance AI fits better, and its tracing and approval controls are more mature.
- **Relevance AI vs Cargo**: Cargo is purpose-built for revenue orchestration: routing, waterfall enrichment, and pushing scored records into a CRM, with GTM data plumbing as the whole point. Relevance AI is a general agent platform that happens to be sold heavily into GTM. A team whose problem is specifically lead routing and enrichment logic will get there faster on Cargo. A team that also wants support triage, meeting agents, and internal research from the same platform will prefer Relevance AI, at the cost of building more of the GTM logic themselves.
- **Relevance AI vs Clay**: Different jobs that get compared constantly. Clay is a spreadsheet with waterfall enrichment underneath and is unmatched for building and enriching a prospect list from dozens of data providers. Relevance AI has no comparable data waterfall and expects you to bring or buy the data. The common pattern is both: Clay assembles and enriches the list, Relevance AI runs the agents that research, draft, and act on it. Buying Relevance AI expecting Clay's data coverage is the most common mismatch in this category.
- **Relevance AI vs Lindy**: Lindy is the more approachable agent builder, especially for email and calendar work, and it is easier to get one useful agent running in an afternoon. Relevance AI goes further on the operational side: multi-agent Workforces, a job queue with retries, tracing with per-step cost, evals, and an MCP gateway for governing tool access. Choose Lindy for a personal or small-team assistant, Relevance AI when agents will run unattended against company data and someone will be asked why one of them was wrong.
- **Relevance AI vs n8n**: n8n is the engineer's answer: self-hostable, source-available, no per-Action meter, and with AI agent nodes that cover a good share of what Relevance AI does, at the price of running the infrastructure yourself. Relevance AI is the managed answer, with agent governance, approvals, evals, and observability built rather than assembled. If you have engineering capacity and want to avoid metered pricing, n8n almost always wins on cost. If you want business users authoring agents with guardrails, n8n will not get you there without significant internal work.
- **Relevance AI vs Zapier**: Zapier remains the widest integration surface in software and now includes AI agent features of its own. It is the right tool when the task is deterministic and the value is in the connector count. Relevance AI is the right tool when the task requires reading unstructured input and making a judgment, and when you need tracing and approval gates around that judgment. Many teams keep Zapier as the plumbing and call a Relevance AI tool endpoint from it for the step that needs reasoning.

## Implementation

- Setup time: An hour to sign up, clone a marketplace agent, and see it run. A week or two to get a first agent doing real work against real data, most of which is spent writing instructions, connecting integrations, and discovering the edge cases where the agent is confidently wrong.
- Learning curve: Low to begin, steep in the middle. The visual tool builder is approachable for a non-engineer, but reliable agents require prompt discipline, structured output enforcement, and an appetite for reading traces. Multi-agent Workforces and custom API steps are a genuine technical exercise, not a no-code one.
- Onboarding: Self-serve on Free, Pro, and Team, with documentation, a changelog, a community, and a marketplace of clonable builds. Enterprise adds a dedicated account manager and hands-on implementation. There is also an expert partner directory for teams that want the build done for them.
- Migration: Nothing migrates cleanly from another agent platform; prompts and logic have to be rebuilt, though tools exposed as API endpoints make it possible to call Relevance AI from an existing automation stack rather than moving everything at once. Anyone on a pre-September 2025 subscription is on the older credit model until renewal, and should recompute their volume against Actions before assuming costs are comparable. Bring-your-own API keys should be configured before any high-volume run, not after the first bill.

## Platform, API & security

- Platforms: Web application, REST API, Model Context Protocol (MCP) server, Slack, Microsoft Teams, Android app, Embeddable tools and agents
- API: Full REST API for running tools and agents, plus every tool exposable as its own HTTP endpoint or embeddable component. An MCP server lets external clients call Relevance AI capabilities, and the MCP gateway governs which tools an agent can reach. OpenTelemetry export streams traces and audit events to customer-owned destinations on Enterprise.
- Compliance: SOC 2 Type II, GDPR
- Data residency: Data residency options are offered on Enterprise plans; the company is Australian and operates a public trust center.
- SSO: SSO and SAML on Enterprise plans, alongside role-based access control and audit logs.
- Security notes: PII masking, encryption at rest, human-in-the-loop approval gates, version control on agents, and a stated commitment not to train on customer data. AWS Bedrock Guardrails can be applied to bring-your-own Anthropic models for prompt-injection and PII filtering. Note that the access-control features (RBAC, SSO, audit logs) are Enterprise-only, so self-serve plans run without them.

## Support

- Channels: Email and contact support, In-product help, Community forum, Dedicated account manager on Enterprise
- Documentation: Substantial documentation reorganized in March 2026 into Getting Started, Build, Integrations, Enterprise, and Access and Administration, plus a detailed public changelog that is unusually useful for tracking what actually shipped.
- Community: An active Relevance Community plus an expert partner directory of consultants who build on the platform, which is where most of the practical agent-design knowledge circulates.

## Company

- Founded: 2020
- Founders: Daniel Vassilev, Jacky Koh, Daniel Palmer
- Headquarters: Sydney, Australia
- Ownership: Independent, venture-backed (OnSearch Pty Ltd trading as Relevance AI)
- Employees: ~100 (est. 2026)
- Funding: Roughly $37 million raised in total, including a $24 million Series B led by Bessemer Venture Partners in May 2025.

Funding history:

- Series A (2023): $15M. Led by Insight Partners, with Peak XV and King River Capital participating.
- Series B (2025): $24M. Led by Bessemer Venture Partners; Insight Partners, Peak XV, and King River Capital returned.

Timeline:

- 2020: Founded in Sydney by Daniel Vassilev, Jacky Koh, and Daniel Palmer, initially building vector search and tooling for unstructured data.
- 2023: Pivots to AI agents and no-code agent tooling, and raises a $15 million Series A led by Insight Partners.
- 2024: Agent platform matures with the tool chain builder, knowledge retrieval, and the first GTM-focused templates.
- 2025: Raises a $24 million Series B led by Bessemer Venture Partners in May and launches Workforce, the multi-agent team system, alongside prompt-based agent creation.
- 2025: On 8 September, replaces the single-credit model with Actions plus Vendor Credits, discontinues the Business plan, and enforces seat and project limits per tier.
- 2026: Enterprise features arrive in sequence: OpenTelemetry export in January, the Evals framework in February, agent performance observability in March, then Confluence knowledge sync, an Android app, and eval cost breakdowns through mid-year.
- 2026: Public pricing page reduced to the Enterprise plan alone with a Talk to sales call to action, while Free, Pro, and Team remain live in the product and documentation.

## Integrations

HubSpot, Salesforce, Slack, Microsoft Teams, Gmail, Outlook, Google Calendar, Notion, Google Drive, SharePoint, Confluence, Zendesk, Freshdesk, Snowflake, WhatsApp Business, Firecrawl, Lusha, Airtop, OpenAI, Anthropic, Google Gemini, AWS Bedrock, Pipedream (roughly 2,000 further services)

## FAQ

### What is Relevance AI used for?

Building AI agents that do repeatable business work without code. Typical builds are research and enrichment on inbound leads, drafting outbound messages grounded in something specific about the account, summarizing calls and extracting commitments, triaging support tickets, and cleaning CRM data against a company's own definitions. Agents are given tools, knowledge, and approval rules, and can be grouped into Workforces where a manager agent delegates to specialists.

### How much does Relevance AI cost?

There is a free plan with 200 Actions per month and a one-time grant of 1,000 Vendor Credits. Pro is $19 per month billed annually ($29 month to month) for 2,500 Actions and about $20 of Vendor Credits. Team is $234 per month billed annually ($349 month to month) for 7,000 Actions, about $70 of Vendor Credits, five build users, and five projects. Enterprise is quote-only. Extra Actions cost $80 per 1,000 and extra Vendor Credits $20 per 10,000.

### What is an Action in Relevance AI?

One tool run. Sending an email, calling an API, or executing a multi-step workflow each count as an Action, and a run counts even if it fails. This is the meter that catches people out, because a multi-agent Workforce handling one request can consume many Actions, and retries are billable. Vendor Credits are the separate meter covering the underlying model spend.

### Does Relevance AI have a free plan?

Yes. It gives 200 Actions per month, 1,000 Vendor Credits granted once at signup rather than monthly, one build user, and one project. That is enough to build and test an agent, and not enough to run one on a schedule. Note that the marketing pricing page no longer lists it; you reach it through the signup flow or the documentation.

### Why does the Relevance AI pricing page only show Enterprise?

The company has moved its public merchandising upmarket. As of August 2026 the pricing page shows the Enterprise plan and a Talk to sales button only. The Free, Pro, and Team plans still exist, are still documented, and can still be bought self-serve, but the vendor is clearly prioritizing sales-led deals. Any comparison quoting a $599 Business tier is out of date; that plan was discontinued on 8 September 2025.

### Can I use my own OpenAI or Anthropic API key?

Yes, on any paid plan. Connecting your own keys bypasses Vendor Credits entirely, so you pay the model provider directly and Relevance AI only meters Actions. For heavy or long-context workloads this is a meaningful cost reduction and it is worth configuring before the first high-volume run rather than after the bill arrives.

### Relevance AI vs Clay: which do I need?

They solve different problems and are often bought together. Clay is a spreadsheet-shaped tool with waterfall enrichment across many data providers, unmatched for building and enriching a prospect list. Relevance AI has no comparable data waterfall and expects you to bring the data; its strength is agents that reason, act, and escalate. Clay assembles the list, Relevance AI works it.

### Do I need to be technical to use Relevance AI?

Not for the first agent. Cloning a marketplace template and connecting an integration is genuinely no-code. Building reliable agents is a different matter: it takes prompt discipline, structured output enforcement, patience with tracing, and comfort with API steps once the workflow branches. Teams that succeed usually have one person who owns agent quality as an ongoing responsibility.

### Is Relevance AI secure enough for customer data?

It holds SOC 2 Type II, supports GDPR obligations, encrypts at rest, offers PII masking, and states it does not train on customer data. The caveat matters: RBAC, SSO, audit logs, data residency, and the evals and observability tooling are all Enterprise-only, so a self-serve buyer is running without the access controls and without a way to systematically measure whether agents are behaving.

### Can agents act without a human approving each step?

Yes, and you should decide deliberately where that line sits. Approval gates pause specific high-stakes actions for human sign-off at thresholds you define, and escalation rules route cases the agent should not decide to a person with context attached. Everything else runs unattended. Most teams start with approvals on anything customer-facing and loosen from there once traces show the agent is reliable.

### Who owns Relevance AI?

It is an independent, venture-backed company based in Sydney, trading as Relevance AI under the legal entity OnSearch Pty Ltd. It has raised roughly $37 million in total, most recently a $24 million Series B led by Bessemer Venture Partners in May 2025, with Insight Partners, Peak XV, and King River Capital also on the cap table. It has not been acquired.

### What happens if an agent fails halfway through?

Runs sit in a job queue, so failed steps retry rather than disappearing, and the Tasks view has dedicated tabs for approvals, escalations, and errors with day-bucketed timelines. Full tracing shows the exact step, model call, and cost where things went wrong. The billing consequence is that failed and retried runs still consume Actions, so a flaky integration costs money as well as time.

## Editorial verdict

Relevance AI is one of the better-engineered agent platforms available to a small company, and the parts that matter most for production use, tracing with per-step cost, approval gates, escalation rules, a job queue with retries, and an honest two-meter bill, are more mature here than in most of its peers. The friction is commercial rather than technical. Pricing jumps from $19 to $234 per month with under three times the Actions, the Business tier that used to bridge that gap was retired in September 2025, and the governance features that make agents safe to leave running unattended (evals, observability, audit logs, SSO) are all locked behind a quote-only Enterprise plan. The public pricing page now advertises nothing else, which tells you where the company's attention has gone. Buy the Free or Pro plan for supervised, human-in-the-loop work where a person reviews output, and expect real value from it. Do not plan an unattended, customer-data-touching deployment on a self-serve plan, because the tools to prove it is behaving are not sold at that price.

---

Source: SaaSTracker (https://saastracker.org), an independent editorial project. This profile is compiled from public information, carries no peer reviews or paid placement, and was last reviewed 2026-08-23. Awards are judged on published criteria: https://saastracker.org/methodology