Overview

Meetings are compounding faster than sends: efficiency is improving, not just volume. Click a point to open the leads behind it.

Constraint: reply → positive reply is the tightest non-structural stage. Fixing message quality moves more than fixing volume.

Top-right = scale. Bottom-right = the expensive quadrant — high volume behind a message that isn't landing.

Compliance Doc Review has a significant positive slope (p = 0.004) while still under the 2% line on some days. The slope is the signal.

Winner gates (all four): positive reply ≥ 45% · meeting rate ≥ 0.55% · unsubscribe ≤ 0.8% · cost per meeting ≤ $75 · minimum 2,000 delivered.

Engagement combines opens, clicks, reply, response speed and follow-up engagement. AI fit combines ICP alignment, company size, industry relevance, technology fit, operational complexity and automation opportunity. Intent is weighted from website sessions, calendar views, deep clicks and multi-threading.

Read click and reply, not open. Apple Mail Privacy Protection manufactures opens; the filter removes pre-fetches but the metric remains directional only.

Tue–Thu 07:00–09:00 local is the band. Friday afternoon sends are near-worthless — reallocate that capacity.

Median 19.6 h, but 31% of replies land inside 4 h — and those convert at 2.2×. The follow-up SLA should be 1 h, not 1 day.

Break-even is touch 5. Touches 6–7 yield 0.21% of replies and cost 0.96% in unsubscribes — cut them.

Human beats AI on positive reply rate by 9.3 points — real. AI-assisted beats human on reply rate by 0.6 points — not real; the intervals overlap.

The decision chart. Upper-left wins. Human-written cannot carry volume at $71/QL no matter how well it converts.

The deliverable sentence: AI wins at Director level and below $50M revenue; humans win at C-level and above $75M. Assisted wins everywhere in between.

v3.2 regressed and was auto-rolled back within 36 hours. That rollback is what makes weekly prompt shipping safe.

Trigger-event personalization is worth 2.4× role-only. Company-level is barely better than role-only — it costs enrichment budget for almost nothing.

Invoice / AP keying is up 47% this quarter and no live campaign addresses it. This chart is the strategic output of the whole dashboard: it says which agent to build next.

Positive share is up 6.9 points since June while negative fell 3.4 — the market is warming to the message, not just receiving more of it.

"We'd build this in-house" is up 31% — the fastest-growing objection in the programme, and the one the next enablement asset has to answer.

Do not compute CAC on a hatched cohort. Median sales cycle is 74 days; a cohort is immature until day 90.

No-show triples past 7 days out. Cap booking windows at 6 days — a calendar setting, not a messaging problem.

Sample size is fixed before launch. Peeking and stopping early is how teams accumulate winners that don't replicate.

A winner with an empty promoted column after 14 days becomes a recommendation card. That field is what closes the loop.

Rates are fully loaded, not salary. Minutes-each is measured against the manual baseline recorded before the workflow was automated, not estimated.

Cost rose 15% over the year while value rose 142%. That gap is the whole case for automation, and it is the only chart on this page a CFO needs.

Calendar viewed, not booked is the highest-lift signal in the system (6.3× on meeting booked) and it decays within 24 hours.

Two metrics are below target: verified email rate and meeting rate. Both are fixable at the top of the funnel — verification before send, and CTA placement.

Guard: any combination that drops a cell below n = 30 renders "not enough data" instead of a rate.

Managed IT on AI-assisted is the strongest cell in the grid — 51.8% positive reply at $56 per meeting. Insurance on AI is the weakest at 22.4% and $210.

Every view encodes its filters in the URL, so a filtered dashboard can be pasted into Slack and land on the same numbers.

Freshness is a first-class metric. A stale mart is worse than a missing one, because nobody notices.

Every alert names an action. An alert that tells you a number changed but not what to do about it gets muted within a fortnight, and then the real ones get muted with it.

One gap: the external SDR account has no MFA. External accounts scoped to a vertical still see contact-level PII.

The finding worth keeping: the mockup's palette is not broken, its ordering is. Re-ordering the same seven hexes lifts the worst adjacent colourblind separation from ΔE 7.0 to ΔE 14.1 — a fail to a clear pass, zero hex changes. But that best order puts green in slot 2, which would render "unsubscribe" and "negative reply" green. Semantics beat the gate.

Two rules apply to all of them: never a dual y-axis, and every rate carries its denominator in the tooltip.

Generation mode is the most-used filter in the dashboard, which is why it sits in the top bar rather than the drawer.