Our read
Start with account research and drafts that a person approves, and treat autonomous replies and meeting booking as a later step. Lightspeed documents that order: small batches, every output approved, automation only once quality is stable. Qualified reports larger results, but its public case does not show where people check the work, so it cannot set your approval rules.
What the platform gives a sales team to build with
Relevance AI describes itself as a platform for building and deploying AI agents for business workflows, without coding, including research and sales outreach. Its homepage shows sales agents that map neatly onto a rollout: a Research & Enricher, a Pre-meeting Prepper, an Outbound Prospector, a Meeting Scheduler and a Post-call Actioner.
The vendor also frames autonomy as four levels. In its wording, level 2 is a copilot where a human reviews the result, and level 3 is an autopilot where proven playbooks become governed agents reacting to events and signals. The vendor adds that real business impact comes from levels 3 and 4. That is a sales position, so use the levels as vocabulary for your own phases, not as a target you must reach.
Level 1: research before any contact
The lowest-risk first job is an agent that prepares an account and sends nothing outside the company. Lightspeed describes its AI BDR, Alex, as researching prospects before it crafts outreach. Before building, the team interviewed high-performing reps, mapped current and future workflows, and co-designed the agent with sales, brand, legal and security. That design work became the reference for the build.
Canva describes a similar low-exposure placement: agents surface in Slack and work inside Salesforce on account and opportunity tasks, covering preparation, summarisation, enrichment and scoring. Its sales leaders keep coaching, negotiation and high-trust customer conversations human-led.
Editorial suggestion, not a vendor feature: have reps read the research brief before real calls and mark what is wrong or missing. Those corrections are the raw material for the next level.
Level 2: drafts that a person approves
Lightspeed calls its first stage crawl: small batches so problems show up early and safely, human approval of every output, and early discovery of failure modes. The next stage, walk, raises volume gradually while the team refines prompts and tools and keeps functions aligned. Its advice to other teams includes documenting fail cases.
Canva reports the same habit from the IT side: prototypes start in sandbox environments, with human-in-the-loop review and several iterations to tune prompts, context and orchestration. A workflow moves to production only when governance, monitoring and the user experience are in place, and ownership then shifts to the business team.
Relevance AI lists human-in-the-loop approvals among its platform controls. The sources we read do not describe how an approval step looks to a rep or what it can block, so confirm that in a demo with your own draft-and-send flow.
Level 3: autonomous replies and bookings
Lightspeed's third stage, run, allows automation only after quality is stable, with ongoing monitoring and reusable components for new agents. Alex is described as handling replies and handing over to reps. The case does not say which replies Alex sends unreviewed, so that boundary is not established by the source.
Qualified gives larger vendor-reported figures: more than 35 agents, $7 million in pipeline and $500,000 in closed revenue over six months. It says Relevance AI turned 40 to 50 BDR tasks into an agent performing at human-level quality in a week, and that three meetings were booked in one afternoon entirely by agents. The case does not detail data sources, which actions run unattended, or where people review. These are vendor-published numbers from one company, not a benchmark, and they should not be set against Lightspeed's rollout.
What the trust page covers before prospect data flows
The trust page lists a SOC 2 Type II report and a data processing addendum. Neither tells you whether your own use of prospect and customer data is compliant; they are the starting documents for your legal and security review.
The same page names subprocessors, including Amazon Web Services, MongoDB Atlas, Google Cloud for cloud APIs and AI models, and Slack, and notes that ledgerup.ai and Databricks were added in March 2026. It lists customer personally identifiable information among data collected. Relevance AI's homepage claims PII masking, audit logs and no training on customer data; treat those as vendor statements to verify. Lightspeed put Security and Legal in the build pod from the start.
Deciding when to move up a level
Editorial recommendation: move from drafts to autonomy per action, not per agent. Booking a meeting from a clear inbound request is a different risk than answering an objection. Lightspeed's gate is stable quality, and Autodesk's account of its own rollout echoes it: scale only where the data shows lift, otherwise iterate.
Before granting autonomy, settle who owns the agent, what triggers a handover to a rep, and how a bad send is found and reversed. Canva created a dedicated GTM AI Lead role to move initiatives from experiments into durable workflows.
Method and limits
2026-11-12
This dated snapshot counts distinct companies in public customer evidence. CONFIRMED means a case study, named testimonial or press source; DECLARED means the company itself told usedby; OBSERVED means a vendor listing; SIGNAL means a job post. All four count, but their strength differs. Removed and unlevelled lines are excluded. Examples are a capped selection verified within 90 days. Vendor facts are sourced records checked within 60 days. Coverage is uneven and missing evidence is unknown. These figures measure neither market share nor product performance; capture dates are not adoption dates.
