Relevance AI

Relevance AI: a bought platform that engineers can still drive

Relevance AI sits at the "buy" end of this site's spectrum, with a door left open for engineers. In the docs it is a low-code and no-code environment where single agents and multi-agent teams are assembled¹; on the product page the scope widens to the whole life of an agent, from build through operation to management at scale². Its home page brands the company around the phrase "The home of the AI Workforce"³. It ranks fifth here. For a platform leader, the questions are how far a business team can get on its own, how much control engineering keeps, and what the platform does not cover.

What a build looks like

The docs walk through five steps. A new agent starts life one of three ways: generated by Invent, copied from a marketplace template, or assembled by hand⁴. Tools are added through a no-code tool builder⁵. Reference material arrives by upload or by sync, with Drive, SharePoint, Notion and public web pages named as sources⁶. Several agents can then be wired together into a workforce by drawing the links on a canvas⁷. The final step in the docs covers control, where a builder defines escalation paths, approval steps and the triggers that start work⁸.

Invent is the headline feature. A user types what the agent should do, in ordinary words, and Invent returns a working draft with its prompt written, tools attached and evals ready⁹. The home page says Invent also helps people work out what to automate¹⁰. The agents page goes further, promising whole teams of specialists that ordinary staff can create by writing in everyday language¹¹.

Under the canvas, Relevance models orchestration as a graph of nodes and edges¹²; LangGraph uses the same structure but hands it to developers as code¹³. Where no-code runs out, a tool can execute Python or JavaScript inside a sandbox¹⁴. Four start conditions are offered for an agent run: schedules, app events, webhooks and custom signals¹⁵.

Quality controls

Relevance AI is one of the few contenders on this site that publishes a built-in evaluation gate. No edit reaches production untested, because the product re-executes the agent's full set of evals ahead of each publish¹⁶. Testing continues after release too, with a sample of real runs scored on an ongoing basis so that drift shows up early¹⁷. Version control is built in, with separate draft, active and pinned states for each agent¹⁸.

A platform team building on a framework would normally have to assemble that loop from separate tools. Here it is part of the product.

The engineer's path

Engineers are not locked out of the no-code surface. Developers can operate the platform from coding assistants such as Claude Code, Codex or Cursor¹⁹, connecting through an MCP install command printed on the product page²⁰. That can give a platform team a way to manage agents alongside its code and to script changes, while business users keep the visual builder.

Implementation support

Relevance places its own deployment staff inside the customer for the launch, then teaches the in-house team to carry on building without them²¹. Before any build, that team also works with the customer to chart current workflows and choose where agents will pay off²². Speed claims vary by page, from a few hours for some use cases²³ to minutes when starting from the builder or a template²⁴, and both come from Relevance itself.

Among the other contenders, this comes nearest to how Nexus works, since Nexus also sends an engineer to build the first procedure alongside the client²⁵. The difference sits in the commercial model and the timeline commitment. On the Nexus side the commitments are explicit: a first agent live in 4 weeks inside a 3-month proof of concept, and a price per role automated²⁶. No timeline commitment or price for the deployment team appears in the Relevance pages gathered for this review.

Integrations

Counting connectors depends on which page you trust: the integrations banner claims over 2,000²⁷, and the searchable directory beneath it totals 2,728²⁸. Other pages use other numbers, 1,000-plus apps on the home page²⁹, 2,000-plus for the Enterprise tier on the pricing page³⁰, and 1,000-plus native connectors with MCP and custom ones on the product page³¹. The figures differ between pages; the directory total is the most specific.

Either way, the catalog is large, second on this site behind the more than 4,000 integrations Nexus lists³².

Security and certifications

Relevance AI publishes its certifications. For the SOC 2 Type II question, Relevance gives a written answer in its FAQ³³ and repeats it as an enterprise badge³⁴. There is a GDPR badge as well³⁵, and compliance with both SOC 2 and GDPR is listed even among the features of the entry Pro plan³⁶. Three further commitments sit on the enterprise page: three-nines availability under an SLA³⁷, data encrypted when stored³⁸, and an undertaking to keep customer data out of model training³⁹. A changelog entry describes masking or blocking personal data such as emails and phone numbers⁴⁰, and the docs go as far as saying a customer's project information is kept confidential from Relevance staff⁴¹.

The uptime SLA is notable. Its uptime figure appears in the fact files gathered for this review.

Governance

Enterprise buyers get access by role⁴² and federated sign-on through SSO or SAML⁴³. Risky actions stop and wait for a person to approve them, and the customer decides where the risk line falls⁴⁴. The controls can be fine-grained, with approval rules set per tool⁴⁵ or by condition⁴⁶ and escalations delivered over email or Slack⁴⁷, all managed from one pane that also covers access, permissions and data security⁴⁸. Relevance's enterprise pitch is that each agent gets its own quality bar, guardrails and monitoring⁴⁹.

For a platform leader, the approval rules are the most useful detail. They let IT require a human decision on specific tools without blocking the rest of the agent.

Deployment and data

Two more enterprise items deserve a mention for compliance teams: a data residency control⁵⁰ and audit logging⁵¹. Logs need not stay with the vendor, because audit records and run traces can flow to storage the customer controls⁵² using OpenTelemetry⁵³, with Delta Share named as a second export route⁵⁴.

For a platform leader the gap is hosting, since running Relevance in your own data center, your VPC or any self-managed form is not publicly stated on the pages we opened. A team whose policy requires agents to run inside its own network will find Dify, LangGraph and CrewAI document that option^55,56,57.

Pricing

Relevance AI publishes self-serve prices. A pilot can start cheaply, since Pro costs from 19 dollars a month⁵⁸; that is the yearly-contract price, and monthly billing is 29 dollars⁵⁹. Moving up to Team raises that to 234 dollars a month on an annual contract, or 349 dollars paid month to month⁶⁰. Annual billing saves 33 percent⁶¹. Usage is counted in Actions, of which Pro bundles 30,000 a year, roughly 2,500 a month⁶²; beyond that, Actions are sold only in lots of 1,000⁶³ at 80 dollars a lot⁶⁴. Model and tool spend runs through a second meter, Vendor Credits, at 20 dollars for 10,000⁶⁵, which the company says it charges at wholesale cost without margin⁶⁶. Even on Pro, a team may supply its own model API keys⁶⁷. The Team tier unlocks agents that make calls and join meetings⁶⁸.

Enterprise carries no public price⁶⁹; buyers negotiate it and are assigned their own account manager⁷⁰. Its stated audience is firms that want revenue to grow faster than staff numbers by adding an AI workforce⁷¹.

Fine print on plans and support

Note that there is no longer a free tier, for new or current accounts⁷². Plans cannot be reassigned from one Organization to another⁷³, and top-ups are only sold to accounts already on a paid plan⁷⁴. Pro and Team customers can read the rollover rules in the docs, but on Enterprise the docs defer to the account manager⁷⁵. The Pro limits, two people who can build and a single project⁷⁶, suit a trial better than a company-wide rollout.

Support terms are spelled out. Phone time with staff is an Enterprise privilege⁷⁷, unpaid accounts have no service level⁷⁸, and help is available Monday to Friday, 9am to 5pm in Sydney, with nothing at weekends⁷⁹. A US team should note the time zone. Relevance has sunset its Business plan and moved those accounts to Team⁸⁰. First-line help comes from Harley, an AI support agent the company warns is in beta⁸¹, and while bug reports do reach engineering, the support page says they go unanswered⁸².

Who it suits

By the docs' own labels, Pro suits a lone go-to-market operator⁸³ and Team suits groups building in volume⁸⁴. According to the product page, the daily owners of agents should be the ops and revenue people who do the underlying work⁸⁵. Use cases on the home page span seven functions, from sales and customer success through marketing, support and operations to research and HR⁸⁶.

Customer examples are published on the home page. Send Payments, a featured customer, reports reclaiming 40 hours every week across its operations worldwide⁸⁷. For Zembl, AI sales cover at all hours is tied to a 30 percent rise in conversions⁸⁸. At Qualified, the home page says, more than 35 agents across the company produced a 7 million dollar pipeline⁸⁹. An Autodesk senior marketing manager is quoted in support of the product⁹⁰. The site also claims thousands of teams helped⁹¹ and points to its 4.5-star average on G2⁹².

Where Relevance AI fits in a build-versus-buy plan

For a platform team that would otherwise build on a framework, the most valuable parts of Relevance AI are the ones that are tedious to assemble by hand. The eval suite that re-runs on every change¹⁶, the draft, active and pinned versions¹⁸, and per-tool approval rules⁴⁵ together form a release process. On LangGraph, the same team would write its own authentication handlers⁹³ and pay for tuned evaluators by the run⁹⁴. Relevance packages those concerns into the product.

Some work stays with engineering. When the catalog lacks a system, the answer is a custom connector³¹ or a sandboxed code tool in Python or JavaScript¹⁴, and either one puts a developer back in the loop. The MCP connection²⁰ gives that developer a scripted path into the platform. Hosting is the main decision engineering cannot make here, since a self-hosted or VPC option is not publicly stated on the pages we opened.

Run cost is easy to model. Actions are metered⁶² and extra blocks cost 80 dollars per 1,000⁶⁴. Because Vendor Credits are billed at cost⁶⁶ and own keys are allowed⁶⁷, the model bill can be forecast from expected usage. What the forecast will not include is the time business staff and engineers spend building and maintaining each agent, which is the line Nexus folds into a price set per role automated²⁶.

Three contract details deserve a question before signing. Subscriptions cannot be moved between organizations⁷³, which matters for groups with several legal entities. Rollover terms on Enterprise must be confirmed with the account manager⁷⁵. And support follows Sydney business hours with no weekend cover⁷⁹, so a US team should ask how Enterprise escalations are handled overnight.

Verdict

Relevance AI is the most complete no-code platform on this site. It publishes SOC 2 Type II³⁴, a 99.9 percent uptime SLA³⁷, an eval gate on every change¹⁶ and fine-grained approval rules^45,46, and engineers can work on it from their own coding tools¹⁹. It ranks fifth because self-hosting is not publicly stated, support hours follow Sydney time⁷⁹, and the deployment team comes without a published timeline or price. A platform leader who wants the speed of a bought product together with a committed delivery date and a price set per role automated will find that combination at Nexus²⁶.