How Closer works

an admin drops call recordings, transcripts, pricing docs, positioning notes, and battle cards into the Workspace. It's a one-time setup, then top-ups when things change. More material = sharper personas and objections.

Closer runs a series of AI passes to identify recurring buyer types, product mentions, common objections, and standout ‘best line’ moments. Each extracted item shows the source quote it came from, so admins can audit before approving.

a rep opens a persona, hits ‘Call now’, and holds a real spoken conversation. The AI plays the buyer, follows the extracted objection patterns, and pushes back the way real prospects do. A call takes 3–10 minutes.

the moment the call ends, the transcript is saved, the rubric is scored, and standout lines can be pinned to Best Lines. Managers coach on specific turns; reps rerun the scenario. Data also improves what Closer knows about the org over time.

when the AI says a persona is ‘skeptical about ROI’ it can point to the exact line where a real prospect said something skeptical about ROI. This is the single most important design choice in the product - how hallucination is kept out of training.

each card shows voice tags (blunt, warm, distracted…), scenario (cold call, renewal, demo), and a 1–5 difficulty badge. Reps pick intentionally - a new rep starts on 2, a senior rep runs 5 to stay sharp.

clicking any field opens a drawer showing the source quotes and the workflow steps that produced it. Admins verify before approving; managers use it to explain why a persona pushes back the way it does.

if there's no evidence for a trait the field stays blank rather than being filled with a plausible guess. Better to show nothing than teach reps something the buyer never actually said.

Closer uses ElevenLabs voice models over WebRTC. That combination is what makes the back-and-forth feel like a phone call rather than a walkie-talkie - the buyer starts replying almost immediately, the way a human would.

the AI doesn't wait politely for the rep to finish. It cuts in when the rep says something questionable, sometimes goes quiet mid-answer to see how the rep fills the space, and shifts tone based on how the call is going. Reps say this feels ‘uncannily realistic’.

every call is transcribed live and stored against the call record. Nothing is lost - a rep who fumbled at minute 4 can find that exact turn afterwards and hear it back.

the transcript is clickable - click any line to jump the audio to that moment. Managers coach on specifics (‘let's talk about 02:34’) instead of vague impressions. Reps can end a call at any point with no scoring hit.

practice calls, real calls and hiring calls are all transcribed and graded automatically. Coverage isn't capped by manager hours, so a team can push through hundreds of scored calls a day instead of the handful a manager would ever get to.

Discovery, Objection handling, Talk ratio, Listening, Framing, Next step, Tone, and Business acumen. The axes never change - that's what makes a 72 in January mean the same as a 72 in June, across reps and across managers.

when the grading is done for you, manager hours move to the part that actually changes behaviour: talking through specific moments with the rep. The AI proposes, the human still decides what to coach on.

any moment in any transcript can be pinned as a ‘best line’ - a great objection handle, a sharp reframe, a clean close. Closer auto-flags candidates from scored calls; a manager approves what goes in.

a manager clicks any line and leaves a note attached to that exact turn. The rep sees notes in-context when they open the call - no more ‘the note says minute 4 but which part?’

every score, turn, and annotation has an icon that opens a scoped chat. The AI knows the surrounding context - a rep can ask ‘why did I score 58 on Framing?’ and get an answer grounded in what they actually said, not a generic tip.

the AI coach is deliberately narrow. It answers about the specific item you clicked. It won't hallucinate cross-history analysis or invent a story about a rep's career - that's what keeps its answers trustworthy.

after coaching, the rep can rerun the same scenario against the same persona. Compare the two transcripts side-by-side to see whether the coaching landed.

the admin generates a link, the candidate opens it in a browser, and runs 2–4 short AI calls asynchronously - no scheduling, no admin present. Takes 20–40 minutes on the candidate's own time.

the recruiter sees each candidate's transcripts, scores per rubric axis, and highlights. No ‘I liked their energy’ black box - you're looking at the same evidence you'd have for one of your own reps.

candidate scores use the same rubric as internal reps, so you see where a candidate would land on your leaderboard before hiring. A much stronger signal than a résumé plus a 30-minute chat.

the gauntlet is a series of calls, not one. Because the candidate does several in a row and can retry after AI feedback, you see whether their scores climb between attempts. A candidate who starts at 55 and ends at 78 tells you something a single interview never can: they take coaching.

screening 60 candidates the old way costs 30+ hours of interviewer time. Screening 60 through the gauntlet costs zero upfront - you only spend time on the ones the evidence justifies.

Overview (headline metrics and ‘Ask Closers AI’), Analytics (activity, scores over time, rubric heatmaps), Members (each rep's page), Leaderboard, Recruitment (the hiring gauntlet), and Leads.

Personas (view, edit, regenerate the AI buyers), Knowledge (raw docs Closer has ingested), and Workspace which splits into Company (positioning, values), Calls (upload with paste/upload/audio), and Products (the auto-extracted catalog).

training assets, product info, team performance, and hiring inform each other. When a product changes, the personas that mention it update. When a new objection is common on real calls, it shows up in practice. Splitting these across tools breaks that feedback loop.

the Overview cockpit has a natural-language search bar that answers across the workspace - ‘which reps are weakest on objection handling this month?’, ‘what's our most common lost-deal reason?’ Grounded in the actual data, not made up.

extracted products land in a review stage first. Admins bulk-approve, edit, or delete; only approved products show up to reps. Select-all and stage filters make triaging dozens of extractions fast.

each product has an overall confidence score and a per-field breakdown - the AI might be 100% sure of the name and 55% sure about top objections. Click the pill to see which fields are shaky and why.

Target Buyer, Key Benefits, and Top Objections are each generated by their own pass, in order, because later fields depend on earlier ones. Each field has a ‘Regenerate’ button and a workflow trace showing exactly which sources produced it - no black box.

AI Actions (Reanalyze missing / Force refill), Status (Approve / Archive), and Destructive (Delete, with confirmation). Bulk reanalyze runs 3 in parallel with live counts and Cancel. ‘Reanalyze missing’ only touches empty fields - anything a human wrote is preserved.

the database itself enforces that a query from Org A physically cannot return rows belonging to Org B - even if the app code had a bug, the boundary holds. Stricter than app-layer filtering.

recordings are not URL-accessible. To play a clip, you must be authenticated as a member of that org, and the URL is a short-lived signed link. No ‘public share’ mode by design.

a nightly cleanup job purges completed and failed live-call audio on a TTL. Recordings don't accumulate indefinitely - the moment their training value expires, they're removed.

recordings, transcripts and docs uploaded by one org are used only for that org's own personas, scoring and coaching. Nothing is pooled to train Closer's models or a third-party foundation model. Your calls train your reps, not somebody else's AI.

the full privacy policy and terms of service are linked at the bottom of every page. For compliance conversations that need specific wording, that's the canonical source.

Closer uses ElevenLabs realtime voice models over WebRTC. That combination gives the sub-second latency and natural-sounding turn-taking. OpenAI's Realtime API is present in the codebase but not the default path.

English is fully supported with the best voice quality. Other languages work but voice naturalness varies; teams selling primarily outside English should trial with their target language first.

the 8-axis rubric is the default, but each org can shape its own scorecard - rename axes, adjust weights, add criteria that reflect how you actually judge a great call. Custom rubrics are a core part of the product, not an add-on.

Closer runs as a managed product; self-hosting isn't available. If it's a hard requirement, flag it early so we can talk about the roadmap rather than promising a timeline on the call.

each org has a plan plus usage-based consumption for AI calls. Detailed pricing is handled by the sales team, not in the product docs, so avoid quoting numbers on discovery calls.