Stop shipping party tricks. Recruiter approval rate per draft is the only metric that matters.

Because it is the only number that combines three things at once: model quality, surface design, and recruiter trust. A draft that the recruiter approves with one tap is a draft the agent got right enough that no edit was required. A draft that gets edited heavily is a coin flip on whether the recruiter ships it at all. A draft that gets rejected is a wasted token call. Approval rate is denominated in recruiter time, which is the actual constraint in TA. Model accuracy on a held-out candidate matching set is interesting trivia; it does not predict whether the team will still be using the agent in week 4. Approval rate per draft does.

Internal sampling across the teams running the workspace pattern in production. The breakdown is roughly: one persona pass, two sourcing dossier batches, three first-touch outreach drafts (warm-up, follow-up one, follow-up two), one panel scheduling thread (which often expands to four messages once a panelist needs to reschedule), one prep-doc draft for the hiring manager, and two post-panel debrief notes. That arithmetic lands at fourteen per interview booked. Half of those 14 do not require an edit; half are tweaked and approved. The total recruiter time across all 14, when the queue is one-tap, is about 30 seconds. The same 14 in Gmail and a spreadsheet runs 25 to 40 minutes of recruiter clicking.

Heavy review UI. Anything that asks a recruiter to read a paragraph of justification before approving a draft. Anything that requires switching tabs to review a draft. Anything that surfaces a model confidence score the recruiter is supposed to interpret. Anything that buries the approve button below a fold. Anything that asks for a comment before approval. Each of those adds friction measured in seconds, and at 14 touchpoints per interview, every extra second is a 14x tax. By week 2 the recruiter has quietly opened Gmail in another window and started writing the touchpoints by hand, because Gmail is honest about what it costs.

Because Gmail is the floor. It is the surface every recruiter knows, every recruiter has open, and every recruiter trusts. When an agentic tool gets heavier than Gmail in a recruiter's hand, the recruiter reverts to Gmail. The job of the agent is not to be more sophisticated than Gmail; it is to be lighter than Gmail at the task of getting a touchpoint out the door. The one-tap approval queue surfaced inside Gmail itself is the only shape that consistently beats Gmail on recruiter time. A separate dashboard, a Slack bot that requires a thread, a web review interface with side-by-side diffs: all of them lose to plain Gmail by week 2.

A recruiter opens Gmail at 9:02 am. Fourteen drafts are sitting in a queue label, generated overnight or while the recruiter was in a panel. Each draft has a one-line summary at the top (candidate name, action, claim or context). The recruiter taps approve on seven that look fine. Two need a small wording tweak; the recruiter edits inline and taps send. Three are scheduling drafts that fire Cal.com invites; the recruiter taps approve and watches the invites go. One is a debrief note the recruiter wants to think about; it stays in the queue. One was bad; the recruiter taps reject and writes a one-word reason. Total elapsed time, around 30 seconds. The recruiter then opens the next thing on their day.

Three counters per recruiter per week. Counter one: drafts created by the agent. Counter two: drafts approved with no edit. Counter three: drafts approved after edit. Approval rate per draft is (approved no edit + approved with edit) divided by drafts created. Healthy teams sit between 78 and 88 percent. Below 70 percent the agent is creating noise, the recruiter is overwhelmed, and adoption is at risk. Above 92 percent the agent is probably underdrafting and missing the long tail of touchpoints; the recruiter has reverted to Gmail for the harder cases. The shape that matters is the no-edit ratio inside the approval count: when no-edit approvals trend up week over week, the agent is learning the recruiter's voice.

Useful as a sanity check, not as a north-star metric. The benchmarks usually score the model on a held-out set of candidate-to-job pairs labeled by a panel. Two problems. First, the labels reflect the panel's recruiter, not yours; voice and bar differ across teams. Second, the benchmark scores the candidate match in isolation, while the actual unit of work is the touchpoint (which combines match plus phrasing plus timing plus recruiter voice). A model that scores 91 percent on a benchmark and lands a 62 percent recruiter approval rate is shipping fewer interviews than a model that scores 84 and lands an 86. The recruiter is the eval set that pays the bill.

Three constraints. First, the action is the default; the approve button is the largest tap target and the keyboard shortcut is one keystroke. Second, the context is collapsed; the recruiter sees a one-line summary and the draft body, not a paragraph of justification. Third, the queue is in Gmail and on the web with the same shape; switching surfaces does not switch UX. When a recruiter wants to inspect the claim ledger behind a match, the link is one tap deeper, and the deeper view never blocks approval. The recruiter can approve without reading the deeper view. That is the difference between a queue and a review UI.

Both, with one calibration. Agencies have a higher draft count per booked interview because the candidate funnel is wider and more outreach happens before a callback. Their 14 touchpoints baseline is closer to 22, and the recruiter time at that baseline is around 45 seconds. In-house teams at Series A through C usually run lighter funnels (8 to 14 touchpoints per booked interview) and the 30-second baseline holds. The metric, recruiter approval rate per draft, is identical. The math under it scales with funnel shape.

10xats is one option among several agentic recruiting stacks. The reason it shows up in this conversation is that the workspace is built around the approval queue from day one, every write tool produces a draft rather than an action, and the queue surfaces in Gmail, Slack, and the web view with the same shape. That choice is what lets the 30-second number hold across 14 drafts. The alternative pattern, where the agent fires actions directly and asks the recruiter to audit afterward, breaks at the same week-2 cliff that any heavy review UI breaks at. There are several teams shipping the workspace pattern in 2026; 10xats is one of them.