Lever sits in a different shape than the rest of the ATS market. It treats every opportunity as a CRM record first and an application second, which means an AI agent can work the top of the funnel without the recruiter having to manually push leads into a pipeline. The practical workflow for a fast-moving talent team is this: an agent sources and drafts through the Lever Data API, the recruiter approves in Lever, and the reasoning behind every action lives somewhere the next interviewer can actually find it.
The workflow, step by step
1. Sourcing. Point an agent at Apollo or LinkedIn Recruiter for the raw stream. Apollo's filters get a shortlist of, say, 200 senior backend engineers in Berlin with Postgres experience. The agent scores each profile against the role brief and writes a one-paragraph rationale. For deeper coverage of this layer, see AI candidate sourcing.
2. Outreach drafting. Gem still runs the best outreach sequencing on top of Lever. An agent drafts the first-touch message using the candidate's actual GitHub or recent talk, not a mail-merge token. ChatGPT or Claude handles per-candidate personalization; Gem handles cadence and reply detection. The recruiter reviews 30 drafts in ten minutes instead of writing 30 messages in three hours.
3. Push into Lever. The agent creates the opportunity through the Lever Data API, attaches the sourcing rationale as a note, and tags the source channel. Lever is the system of record from this point. The opportunity ID is the canonical pointer.
4. Screening and interview kits. When a candidate books a screen, an agent reads the resume plus the brief and drafts an interview kit: five questions tuned to the candidate's background, plus a rubric. The recruiter edits and ships it. If you are also writing the JD with an agent, the loop closes nicely with AI job description writing.
5. Decision capture. After each interview, the panelist's scorecard goes into Lever. The agent's reading of that scorecard, plus the cross-panel synthesis, lives separately so the next recruiter sees the reasoning, not just the score.
Worked example: a Series B hiring two senior PMs
A 60-person product company opened two senior PM roles in February. Sourcing agent pulled 340 Apollo leads against the role brief, scored them, and surfaced the top 45 with written rationales. Gem ran the outreach sequence; reply rate landed at 18%, on the higher end of the CandE benchmark range for warm outreach. Of the 31 replies, 22 made it to screen. The agent drafted interview kits per candidate. Both roles closed in 11 weeks, vs. the team's prior average of 17. The recruiter spent her time on judgment calls, not data entry.
Where the workflow breaks: persistent state
Every step above leaks context. The sourcing rationale lives in a Notion doc the agent wrote once and forgot. The outreach personalization logic lives in a prompt the recruiter is iterating in ChatGPT. The interview kit lives in a Google Doc attached to the Lever opportunity. Three weeks later the hiring manager asks "why did we pass on Priya?" and nobody can answer, because the agent has no memory across sessions and the reviewer's "no" lives in a Slack thread that scrolled away. One way to solve this is a workspace like Dock that holds the sourcing rationale, the screening notes, and the reviewer decision in rows pointed at the Lever opportunity by lever_opportunity_id. Lever stays the system of record for the candidate; Dock holds what the agent interpreted around them, with a stable agent identity on every write so you know which agent wrote which note.
Why it matters
Lever's strength is the CRM model, and AI agents amplify CRMs more than they amplify pipelines. But amplification without memory just produces faster forgetting. The teams that compound results are the ones whose agent reasoning persists across the hire, the role, and the quarter.
If you are still mapping the broader landscape, start with how to do hiring with AI in 2026 or compare against the Greenhouse-shaped workflow.
FAQ
Does Lever have a native AI assistant in 2026? Lever ships AI features for summarization and search inside the product, but the heavy lift, sourcing, scoring, outreach drafting, interview kits, still runs best through the Data API with an external agent. Native features handle the inside-the-record workflow; the API is where the funnel work happens.
Can I use ChatGPT or Claude with Lever directly? Not natively. You connect them through the Lever Data API, either by writing a small integration or by routing through Gem, Bardeen, or a workspace that already speaks both sides. The model handles language tasks; the integration handles reading and writing Lever records.
What's the difference between using Gem and using a raw agent on Lever? Gem is a sequencing and engagement layer optimized for outreach cadence and reply detection. A raw agent handles the per-candidate reasoning Gem cannot: scoring against a brief, drafting personalized first touches, synthesizing interview feedback. Most teams run both.
Where should the agent's reasoning live? Not in Lever notes. Lever is the system of record for candidate data, not for agent reasoning. Keep the rationale, the screening synthesis, and the reviewer decisions in a workspace built for persistent agent state, with pointers back to the Lever opportunity ID.