Relevance AI

Relevance AI: release discipline and a deployment team, inside the vendor's own cloud

Among the software vendors on this site, Relevance AI comes closest to Nexus on the question that decides most stalled pilots, which is whether anyone from the vendor will help get the agent live. A sponsor reading the agents page will find a promise that matters at exactly this stage: Relevance places a deployment team inside the customer to reach go live, and that team then coaches the customer's own staff until they can build fast unaided¹. Around that offer sits a platform its documentation describes as low or no code, built for single agents and for teams of several agents working together², which the product page sums up in three verbs: build, run and manage agents at scale³. In the marketing, all of this travels under the banner of an AI Workforce, with the homepage claiming to be its home⁴.

This site ranks Relevance AI fifth. On release discipline it is among the better documented platforms here, and the embedded team addresses the staffing gap. It ranks behind Dify for this buyer because a hosting option inside the customer's own estate is not publicly stated on the pages we opened, and behind the top three because its support coverage and commercial tiers are shaped around smaller go to market teams, as the sections below show.

Who helps a stalled pilot over the line

The delivery offer is the most interesting part of the package for a sponsor in this position. Besides the embedded deployment team¹, the agents page says the vendor works with the customer to map workflows and pick the agent use cases that suit the organization best⁵, which is the scoping step many stalled pilots skipped. On timing, the vendor goes further and says certain use cases are ready within hours⁶, a claim a sponsor should test against the real complexity of the stalled pilot.

How large the deployment team is, how long it stays and what it costs are not publicly stated on the pages we opened. The pricing page links a dedicated account manager to the Enterprise tier⁷, and the support page limits calls with dedicated staff to customers on Enterprise⁸, so the help described above is most likely tied to the top plan.

Building, from plain language to code

The build flow in the documentation runs in five steps. Agents are created to complete tasks on their own, using Invent, a Marketplace clone or a blank start⁹; they get tools from a builder that needs no code¹⁰; knowledge is uploaded or synced from SharePoint, Google Drive, Notion or public websites¹¹; agents are joined into workforces on a visual canvas¹²; and guardrails are added as triggers, approval workflows and escalations¹³. Invent, the generator at the front of that flow, writes the prompt, the tools and the evaluations from a description in ordinary words¹⁴, and the homepage credits it with helping a company's people work out which tasks to automate in the first place¹⁵. Templates and the no code builder are offered as a quick route, with deployment measured in minutes¹⁶.

Engineers are not left out. A developer who prefers an editor to a canvas can work on the very same agents from Claude Code, Codex or Cursor¹⁷, adding the platform with the one line MCP command the product page prints¹⁸. Under the canvas, multiple agents are coordinated through a node and edge graph¹⁹, and custom tool logic can be written in Python or JavaScript that executes inside a sandbox²⁰. The people the vendor expects to run agents day to day are the operations and revenue staff who sit nearest the work²¹, and the agents page pictures teams of specialist agents assembled by those same employees in everyday language, without programming²².

The release gate

For a sponsor who lost a pilot to a change board, the release mechanics are where Relevance AI earns its place. Every edit triggers the agent's full evaluation suite before the change is allowed to publish²³, and once the agent is in service Relevance continues to grade a sample of real runs to detect drift²⁴, which gives a change board evidence on both sides of release. Agents carry version states, with a pinned version available alongside draft and active ones²⁵, so production can stay on a release the business has approved while the next one is worked on.

Human control over agent actions is spelled out on two pages. The enterprise page has risky actions pause until a human approves them, at thresholds the customer chooses²⁶, and the vendor says it lets customers define quality standards and guardrails and watch over every agent²⁷. The product page breaks this down into features: approvals that can be required for an individual tool²⁸, approvals that fire only under set conditions²⁹, escalation through email or Slack³⁰, and a single view for data security, permissions and access control³¹.

Security file

The certifications a reviewer will ask for first are present, though the list is shorter than at the large clouds. The enterprise page displays SOC 2 Type II among its trust marks³², a status confirmed in the documentation FAQ³³, GDPR sits beside it³⁴, and even the entry level Pro plan advertises both³⁵. The enterprise controls include data encrypted at rest³⁶ and a pledge not to train on the customer's data³⁷; the documentation adds that Project Information is kept confidential, the vendor's own staff included³⁸. A changelog entry adds the ability to mask or block PII such as phone numbers and email addresses³⁹.

Identity and access controls for the enterprise tier include SSO with SAML⁴⁰ and role based access control⁴¹. The enterprise page also promises a 99.9% uptime SLA⁴², a commitment several larger vendors on this site do not state as plainly.

Where the data sits and how activity is recorded

Two items on the enterprise list speak to a compliance office, one being audit logs⁴³ and the other a data residency control⁴⁴. The changelog goes further, recording that both execution traces and audit logs can be pushed to storage the customer controls⁴⁵ using OpenTelemetry⁴⁶, and the enterprise page names Delta Share and OTEL as export routes⁴⁷. For a security operations team that wants agent activity in its own tooling, this is a practical feature.

What is missing for some buyers is a deployment model inside their own perimeter. Running Relevance in the customer's VPC, in its own data center or as self hosted software is not publicly stated on the pages we opened. A pilot that stalled on data location will not be unblocked by the published options alone, and the residency control⁴⁴ should be explored in detail with the vendor.

Connections

Relevance quotes several different connector numbers across its site, so a sponsor should treat any single figure with care. On the integrations page the headline offers upward of 2,000, from Salesforce to Slack⁴⁸, while the directory beneath it counted 2,728 on the day we looked⁴⁹; the homepage puts the number of connectable apps above 1,000⁵⁰; the Enterprise column of the pricing page lists 2,000 plus⁵¹; and on the product page the native connector count is over 1,000, supplemented by MCP and by connectors the customer can build itself⁵². Whatever the exact count, a sponsor should name the systems the stalled pilot needs and check each one.

To run unattended, an agent can be woken by a custom signal, a webhook, something happening in a connected app, or simply the clock⁵³. That range covers most of the ways a production process hands work to an agent.

Prices and plan rules

Prices are published in the documentation. List prices make an early budget straightforward. The Pro entry point is $19 a month⁵⁴, a rate that assumes annual billing and becomes $29 when paid month by month⁵⁵; Team is $234 or $349 a month on the same two terms⁵⁶; and paying annually comes out 33% cheaper⁵⁷. Usage is counted in Actions, with Pro carrying an allowance of 30,000 a year, equivalent to 2,500 monthly⁵⁸; beyond that, Actions are sold at $80 per thousand⁵⁹ in thousand unit increments only⁶⁰, while additional Vendor Credits run $20 per ten thousand⁶¹. The vendor says Vendor Credits, which pay for model and tool consumption, are charged at wholesale cost without a margin⁶², and a Pro customer can avoid them by supplying its own API keys⁶³. Team adds agents for calls and meetings⁶⁴. The production tier most corporate sponsors will need is priced by negotiation, comes with a named account manager⁷, and appears on the public pricing page without any amount⁶⁵.

Several plan rules affect how a pilot grows. Anyone who priced a pilot on an old plan should recheck, because the documentation says the Free plan is now closed to every account⁶⁶ and the support page says Business was discontinued, with its subscribers transferred to Team⁶⁷. On Pro only a pair of people can build and everything must fit in a single project⁶⁸, which is tight for a cross functional team. Three smaller rules complete the picture: subscriptions are not transferable across Organizations⁶⁹, buying a top up requires a paid plan to be active⁷⁰, and rollover of unused allowances on Enterprise has to be agreed with the account manager⁷¹.

Support coverage, the limit to weigh most

For an agent that will run a production process, support hours deserve close reading. The support page sets office hours of nine to five, Monday to Friday, on Sydney time, and rules out weekend cover⁷². Service level commitments are withheld from free customers entirely⁷³. A reported defect is forwarded to the engineers, yet the support page says no reply will follow⁷⁴, and Harley, the AI agent that fields support questions, carries a beta warning about possible bugs and errors⁷⁵. A North American or European operations team running a weekend process should negotiate support terms explicitly in the Enterprise contract.

Customers and fit

Use cases on the homepage cover research, operations, human resources and customer support as well as the commercial functions of sales, marketing and customer success⁷⁶. Each tier names its buyer: Pro the go to market operator working alone⁷⁷, Team the groups building at scale⁷⁸, and Enterprise the companies aiming to grow without adding headcount by deploying an AI Workforce⁷⁹. Customer examples on the homepage include Send Payments, whose AI first approach is credited with freeing 40 hours every week in operations around the world⁸⁰; Zembl, reporting a 30% rise in customer conversions after putting AI on sales coverage day and night⁸¹; and Qualified, credited with $7M of pipeline from more than 35 agents⁸². A testimonial on the same page comes from Autodesk, attributed to a senior manager in marketing⁸³. Relevance also claims to have taken thousands of teams through the launch of an AI workforce⁸⁴, and it quotes a G2 score of 4.5 stars⁸⁵; both are the vendor's own figures.

Verdict for a team with a stalled pilot

Fifth place goes to Relevance AI for a specific mix of strengths and gaps. It offers something most software platforms here do not, a vendor team placed with the customer to reach production before handing over the skills¹, and its release gate is well designed, with evaluations rerun on each change²³, sampled scoring of production runs²⁴, pinned versions²⁵ and per tool approval rules²⁸. A security reviewer's opening checklist is largely answered by the SOC 2 Type II report, GDPR compliance and an uptime commitment of 99.9%^32,34,42.

Its place behind the four platforms above it comes down to three points: a deployment inside the customer's own estate is not publicly stated on the pages we opened, support runs on Sydney weekday hours⁷², and the published plans are sized for smaller go to market teams^77,68. Against Nexus specifically, the gap is commercial as much as technical: the scope and price of the deployment help are not publicly stated, whereas Nexus folds its consultants into a per role price⁸⁶.