start here — here's your OS, how you plug data in, and the week from prototype to live
beth / nate / ryan · first open lands here · then Automation → Reporting → Client Portal
90 seconds · beth room
seven stops. click through once before the meeting. the OS does the talking.
if a stop is still prototype, say so — then show the week-to-live. honesty is the close.
prototype → live
here's the product, here's how you plug data in, here's how pilot becomes production in about a week.
4,134 tracker rows · Peet's · DICK'S · Albertsons (aliased Grocery-A). Not synthetic demos — their coverage, scored into the OS so you can feel the real tool.
Day 0
4,134 rows from Peet's, DICK'S, and Albertsons already power Coverage, Automation, Reporting, and Client Portal. Sample chips stay honest until your feeds connect.
Day 1–2
Meltwater (or CSV drop on Onboarding). Consent-gated. Coverage stops being a paste tab — same path the three pilots used.
Day 3–4
Lock client rubric. Robot Intern codes; humans govern in Queue — and every accept or correction already recomputes the real books today, live, not someday.
Day 5–7
metric_snapshot already IS the books — that part of the engine runs today. This is just the day your own feed's rows start filling it instead of the pilot set. Reporting and Client Portal read the exact same rows either way.
Integrations holds every hook and every client-scoped consent check is real and DB-enforced today. Vault-backed credential storage is the target architecture — Integrations' key field is honest that it's prototype-only until that lands.
open Integrations →Kill the spreadsheet OS. One crew on the Workflows. One set of books for agency and brand. AI that drafts — humans who ship.
Praytell Impact OS captures coverage, codes it against your rubric, scores Discovery–Authority–Decision, and proves it on one set of books — the Workspace for the agency, the Client Portal for the brand.
the build log · what we built · why
one page for the new hire or the CFO in diligence — the product thesis, the six capability areas, one line of business reason each, and where the full spec lives in the repo.
product thesis
praytell already runs a serious measurement practice — it's just trapped in a spreadsheet. the OS swallows that practice whole (rubrics, campaigns, reporters, hot markets, crisis) and turns earned coverage into audited business proof while you sleep.
one books of record. the workspace for us, the client portal for the brand, the same immutable metric_snapshot under both — so nobody argues about whose spreadsheet is right anymore. we argue about what to do next.
the robot codes, humans govern, offense stays honest. every hero number carries a method label, there are no made-up exposure stand-ins, and nothing publishes to a client without a human saying yes.
each area earns its keep against a business requirement (BR-#) and a workstream (WS#-#) — the tags on the cards. the state chips are honest to a fault: "live" means wired end to end, "partial" means the screen is real but part of the plumbing is still fixtures.
kill the spreadsheet OS. reclaim roughly 10% billable and make coverage + queue the real system of record — not a 26-tab tracker nobody trusts by friday.
requirement · BR-16.3 · BR-16.8–16.12 · WS2 · WS3
surface · Today · Coverage · Govern · Reporting · Outcomes
one books of record with the brand. the overview a client opens is the exact snapshot nate approved in reporting — method labels and all, no second version floating around.
requirement · BR-16.5–16.7 · BR-16.18–16.20 · WS4
surface · Overview · Coverage · Outcomes · AI Visibility · Receipts · How this works
adding a data plug is configuration, not a report build. google ads, meta ads, ga4 and meltwater all speak one MetricPoint contract — consent and write-only secrets baked in.
requirement · BR-12.x · BR-16.13–16.17 · WS5
surface · Connect (agency) · Connections (client)
ai coding only bills when it clears a senior-analyst bar. the gold-set gate is what turns "robot intern" into a real feature instead of a slide.
requirement · BR-3.x · BR-16.22 · O3 · WS3
surface · Robot Intern · Queue · Health · Gold set
foresight without the overclaiming — bands not lines, sample-size guards, internal-only until it's back-tested. if offense lies, the measurement layer's credibility dies with it.
requirement · BR-13.x · O11 · O12 · WS6 (rollout)
surface · Opportunities · Media Contacts · Forecasting
corpus pooling and model calls hard-fail before they can embarrass anyone. every table carries a data-rights tier, every llm call routes through the policy-checking gateway. that's by design, not politeness.
requirement · BR-14.x · BR-15.x · CLAUDE.md law 6d
surface · data_rights_registry · model_policy · ai-gateway
beth's public ai bet, made clickable — a visible crew (conductor, intern, clipper, steward, scribe, mirror, scout, connector, forecaster) with autonomy dials and an inbox, and a hard rule that nothing auto-publishes to a client.
requirement · BR-17.1–17.10 · WS7
surface · Automation · Approvals · workflow recipes
what the state chips mean
these are repo paths, not urls — open them in the same clone the OS is built in. the v2 docs are historical (we banner them); v3 wins any fight.
v3 baseline
docs/BRD-v3.mdthe business baseline — what praytell requires, and why · BR-16.x + BR-17.xdocs/PRD-v3.mdthe product baseline — ia, journeys, pipeline, kpi dictionary · §0 indexes the addendasupplementary specs · read alongside PRD-v3
docs/PRD-v2.1-addendum-admin-portal.mdintegrations admin + the connector framework · BR-12.xdocs/PRD-v2.2-addendum-offense.mdthe offense honesty law — forecasting, opportunities, media contacts · BR-13.xdocs/PRD-v2.3-addendum-corpus-sovereignty.mdrights tiers, the pooled corpus, model policy on egress · BR-14.x / BR-15.xdocs/PRD-v3.1-addendum-swingset-floor.mdswingset floor — conductor + crew, inbox, autonomy dials · BR-17.xbuild ops
docs/EXECUTION-PLAN-v3.mdbuild order — workstreams ws0–ws7 + acceptancedocs/BACKLOG-v3.mdthe workstream backlog — maps ws#-# to e#-### ticketsdocs/PLAYBOOK.mdthe doctrine narrative — the in-app playbook mirrors itdocs/decisions/spike adrs — vendor picks recorded before they hit productionhistorical · superseded but preserved
docs/BRD-v2.mdsuperseded by brd-v3 · br-1…br-15 still in force where v3 stays quietdocs/PRD-v2.mdsuperseded by prd-v3 · the original coding-engine + gate math detaila new Praytell hire
start with the thesis and the six areas, then hit the in-app tour from the button up top — it walks the actual screens in the order nate uses them.
an investor / partner in diligence
the six areas are the whole surface map. the state chips are honest — anything marked partial or prototype isn't on the invoice yet. cross-check against the BR-# / WS# tags in the linked docs.
an engineer joining the build
read CLAUDE.md first, then PRD-v3 §0 for the doc map, then whichever addendum owns your area. every BR-# on the cards below tells you which addendum you answer to.
one thesis. six areas. every hero number can point at a method, a snapshot, and a business reason — never just a slide.
what we heard · what we audited
this isn't a pitch deck fantasy — it's what Peet's, Albertsons, and DICK'S already run by hand, plus how the big shops solve (and don't).
~10% of billable time still dies in trackers — coding, paste tabs, link chasing. Beth wants that time on media and ideas.
DAD exists on paper. Clients want lower funnel, GEO, foresight — without the exposure-multiplier cosplay.
Hourly → outcomes needs audited shared books. No shared system of record yet = no honest contract conversation.
enterprise measurement practice — trapped in Excel.
| client | shape | what it proves |
|---|---|---|
| Peet's Coffee | ~1.1k rows · 11 sheets | Hand Impact Score (tier, highlight, KM, CTA, visual, spokes) · REACH/4 · listicle · crisis · TOP OUTLETS |
| Albertsons (aliased Grocery-A) | ~560 rows · 2025–26 | Reporter column = relationship gold · original vs syndication · initiative tags |
| DICK'S Sporting Goods | ~6.7k rows · 26 sheets | Media Impact Score rubric · Always On + campaigns · HM keywords · Meltwater still carrying Ad Equivalency · broadcast |
4,134 real rows power this prototype. The OS absorbs campaigns, rubrics, reporters, listicles, hot markets, crisis — without asking analysts to learn a foreign model.
holding companies build internal OS layers because vendors monitor — they don't run the agency.
Internal OS (e.g. ECOS-class) + AI agent demos on top of measurement governance they own.
Crawl + alerts. Strong volume. Weak on your rubric, shared books, outcome pricing, client portal honesty.
Encode how Praytell already measures · plug commercial data · one books with brands · Automation crew you can click · refuse fake ad-rate stand-ins for earned value.
Parity is operating leverage + measurement honesty — not crawl size. That's the board-safe claim.
the real business case
Not a vibes deck. Five cash levers with named mechanisms. Assumptions are labeled — if a number isn't auditable, it doesn't belong here.
Tracking burns ~8–12% of billable on accounts like Peet's / DICK'S / Albertsons. Reclaiming that is table stakes. The OS is how those hours become sold counsel, how books unlock outcome pricing, and how brands renew when their CFO asks 'so what.'
Order is intentional: protect revenue first, then expand, then reprice.
A brand that can't defend earned to finance churns or compresses scope. One saved renewal on a mid AOR ($250–600k/yr) pays for the OS build many times over.
proof in product · Reporting → Client Portal handshake · Receipts methodology · no exposure-multiplier stand-ins
Reclaim ~10% monitoring → redeploy into media engagement / Opportunities / counsel that bills. On a 4-person account team at ~$180k fully loaded, that's ~0.4 FTE of capacity (~$70k) sold or used to take another book without headcount.
proof in product · Dashboard hours-saved · Queue keyboard clear · agreement gate
Outcomes / efficiency and GEO panels become paid modules or retainer lines once plugs exist — not free 'extra slides.' Typical wedge: +$3–8k/mo per brand for measured outcomes + AI-answer share.
proof in product · Impact ladder rungs 4–7 · Wilson GEO bands
Hourly → deliverable/outcome pricing only works when books are trusted. Model: move 2 accounts off pure hourly; capture 8–15 pts margin on those books by pricing the outcome, not the paste.
proof in product · Contract / attainment · cost-to-serve on Dashboard
Holding-co pitches already claim 'impact OS.' You either show a working spine (Workflows → Reporting → Client Portal) or you lose on theater. One incremental win at Peet's-class fees resets the year.
proof in product · Demo spine · tracker-vs-OS gut punch · LIVE honesty
Beth · P&L
Cost-to-serve next to attainment. Capacity without headcount. A path to outcome pricing that finance will sign — not a slogan.
Nate & Ryan · delivery
Judgment on edge cases. Offense time instead of paste tabs. Same Reporting send Casey already trusts.
Casey · brand seat
Earn Rate → EIV ladder → GA4 sessions when plugged. Flow-through she can take to her CFO without inventing blended ROI.
New business
Impact OS in every qualified pitch. Honesty + leverage — not crawl size — is the wedge vs holding-co theater.
BR-16.19 — climb only as far as connected data allows. Never invent a blended ROI.
1 · Detection
Coverage landed & deduped
counted
2 · Quality
DAD + Earn Rate vs rubric
scored
3 · Attention
EIV-Reach (modeled)
modeled
4 · Traffic
Earned sessions (GA4)
measured
5 · Interest
Branded search lift
estimated / measured
6 · AI answers
GEO presence + citations
counted + interval
7 · Efficiency
Earned vs PR $ / vs paid $
mixed · labeled
8 · Contract
Attainment vs promise
snapshot-backed
Cut cost by killing the spreadsheet OS. Grow revenue by retaining, expanding, repricing, and winning — with books both sides trust. That's the case. Everything else is decoration.
requirements → build
business requirements name the outcomes. the PRD names the surfaces, tables, and gates. nothing in the UI is decorative.
| BRD | PRD / build | why it matters |
|---|---|---|
| BR-16.6 · one books of record | metric_snapshot + Reporting → Client Portal handshake | Zero “whose number?” fights on approved reports. |
| BR-3.2 / O3 · ≥90% agreement gate | Robot Intern + Queue + gold-set CI | AI coding is billable only when it clears the same bar as senior humans. |
| BR-16.x · Playbook literacy | Workspace Playbook + Client Portal How this works | Every new hire and every pitch can explain the OS in one scroll. |
| BR-17 · Automation | Conductor + Agents + Approvals + autonomy dials | Beth's public AI bet becomes clickable — not slideware. |
| BR-14 / BR-15 · rights & model policy | data_rights_registry + gateway-only LLM egress | Corpus pooling and model calls hard-fail before they can embarrass you. |
| O1 · −75% tracker labor | Ingest → code → score → snapshot; Today/Queue is the human job | Sheets can sync. They are not truth. |
research · methods we stand on
We didn't invent trust theater. We surveyed peer-reviewed stats, marketing science, and AMEC industry law — then encoded only what survives a Beth room.
The exposure-multiplier proxy — the Barcelona-banned fake ad-rate stand-in for earned value, formally rejected by AMEC. Point-estimate forecasts without intervals. Silent model hops. Dual-axis chart lies. If it can't survive peer review or industry standards, it doesn't ship client-facing.
AMEC · Barcelona Principles; industry ban on exposure-multiplier proxies
where · EIV ladder · forbidden-phrase lint · Receipts methodology
why · Praytell needs a board-safe posture that formally rejects the exposure-multiplier proxy. AMEC is the peer org CMOs already recognize — we encode it, we don't debate it in the deck.
McKinsey Consumer Decision Journey (Court et al., 2009) + Praytell DAD practice
where · placement scores · Earn Rate · Client Portal tiles
why · Funnel language execs already use — made measurable against your rubric, not a slide metaphor.
Percentile rank on log10(audience+1); stationary monthly reference distributions
where · PRD §7 Discovery · score versions stamped · Coverage tracker
why · Raw UVPM isn't comparable across outlets. Log + percentile is the defensible normalize — freeze distributions so June books don't drift in July. Currently: the live SQL recompute (app.recompute_metric_snapshot) uses a simpler self-cohort percentile, not yet the stationary reference distribution this entry describes — see the D.A.D. realignment note in the build log. (The v2 per-placement TS module that once carried this doctrine was retired — decision D-007; the reference-distribution upgrade lands in the SQL reducer when it lands.)
100 · D^α · A^β · Dec^γ — exponents declared per client, published, versioned, never fitted
where · score → metric_snapshot · Reporting / Client Portal / Dashboard
why · One hero number for the room — weights transparent, history recomputed with stamps, never a black-box 'impact score.'
Wilson (1927) binomial proportion CI — still standard for rates
where · AI Visibility / GEO panels · citation propensity
why · Presence and citation shares are proportions. Wilson intervals beat naive ±sqrt(pq/n) for small n — and we alert only when CIs are disjoint (no fake week-over-week panic).
Blind double-coding · ≥ min(90%, human−human − 3 pts) · confusion matrices
where · Queue · Health · CI regression on prompts/models
why · LLM coding is billable only when it clears a senior-analyst bar. Exact match for categoricals; ±1 band for ordinals — classical IRR discipline.
Matched-cohort P25/P50/P75; shrinkage by cohort_n; James–Stein intuition
where · Offense · Forecasting bands · internal_only until gate
why · Ranges, never lines. Small cohorts shrink toward portfolio prior so we don't overclaim on n=3.
Quantile regression (Koenker); GBM pinball loss; back-test coverage gate ≥45%
where · PRD §7.7 · BR-13 honesty gate before client-facing
why · When history is deep enough, upgrade the band model — still intervals, still gated by back-test, still no point promises.
Shapley-value attribution (game theory → MTA literature); Markov path models as alt
where · Outcomes efficiency when GA4/Ads consent lands — labeled mixed methods
why · We don't fake last-click ROAS from earned alone. When plugs exist, attribution math must be named — Shapley/Markov are peer-discussed; we stage them behind consent + labels.
Every hero number on Client Portal can point at a method: industry law, peer-reviewed interval, or versioned index — not 'the model said so.'
same row · two worlds
one Peet's headline — left is the sheet, right is what Nate clears in Queue.
Yahoo Life · Tasting Table · 6 Most Overpriced Coffee Chains In 2026
Google Sheet · today
Impact OS · Queue
Same clip. Left burns billable hours. Right is one decision — then it hits the snapshot.
nine stages · one spine
coverage in → audited proof out. offense loops learning back. tap a stage.
Coverage happens
Earned placement lands in the wild — article, broadcast, listicle, syndication.
system of record
Postgres + RLS is the security model. Append-only where trust lives. If a number can appear on an invoice, it came from here.
raw / placement
every coverage item — url, outlet, reach, fetch status
coding (+ coding_current)
append-only field codes; corrections supersede, never mutate
review_task
human queue — SLA, seconds_spent → cost-to-serve
score
Discovery · Authority · Decision · Earn Rate versions
metric_snapshot
immutable period books — Reporting and Client Portal both read this
report_instance
approved/sent proof; exported handshake for the portal
agent_run / agent_run_step
Automation ledger — every LFG is replayable
data_rights_registry
every table tagged for corpus sovereignty (T1–T4)
Automation
Beth bet publicly on Automation. Nate and Ryan still live in Sheets. Competitors demo agents. We built a crew you can see, pause, and approve — without breaking measurement trust.
Agent slideware doesn't move margin. A visible Conductor + Agents does — if every LLM call still goes through one gateway, client model_policy still hard-fails, and client-facing truth never auto-publishes.
Conductor routes a Run (new coverage, weekly reporting, Opportunities sweep, manual LFG). Clipper → Intern → Steward (or Scribe → Steward for Reporting). Handoffs paint on the Workflows. Approvals is the human gate. Autonomy dials: Observe · Draft · Auto-internal · Never-auto-client.
Analysts clear exceptions, not paste tabs. Account leads approve narratives grounded in snapshots. Beth sees cost and approvals on the Dashboard. Brands never see the Workflows — they see approved proof. Mind-blow without trust-blow.
handoff recipes
runs in your cloud
Not locked to Lovable hosting. Normal multi-tenant app — lift the same Postgres + RLS + workers into their environment.
ECS/Fargate or EKS · RDS/Aurora Postgres · Secrets Manager · SQS workers
Cloud Run or GKE · Cloud SQL Postgres · Secret Manager · Pub/Sub jobs
stays · Schemas, RLS policies, coding engine, gateway law, Automation contracts, Workspace + Client Portal UI
rewires · Connection strings, auth adapters (Cognito / Identity Platform), Secrets Manager, CI deploy, Meltwater webhooks
easiest · Keep managed Postgres (Supabase or RDS/Cloud SQL) and lift app + workers into their VPC. Don't rewrite into Dynamo/Firestore.
hard part · SSO (Okta/Azure AD), private LLM endpoints, VPC peering, their change-control — not the React screens.
Medium — weeks for a competent platform team, not a weekend.
interactive journey
click each step — what they told us, what we built, what it returns to the agency
one composition — capture to advise. click a stage to see the surface + the win.
Integrations + Coverage
BR-16 · kill spreadsheet OS
Meltwater and backups land as RawItems. Coverage is one table — filters replace 26 sheets.
Monitoring stops eating ~10% billable. Humans do judgment, not paste.
select a problem from the Lex Overview brief — see BRD → PRD → business return
what we heard
Peet's / Albertsons / DICK'S trackers — thousands of hand rows, Impact Scores typed, Meltwater paste tabs, exposure-multiplier ghosts.
brd
BR-16.8–16.12 · Coverage as system of record
prd surface
Ingest → code → score → snapshot · Tracker + Queue · Sheets sync optional
business impact
Analysts reclaim monitoring hours for media engagement, Opportunities, and relationships.
Open Coverage →step through the job-shaped IA — each stop is a real surface
Today
Portfolio wall — what needs a human in the next hour.
Leadership glanceability without opening six tools.
named clients. real pain. exact fix.
today · Peet's / Albertsons / DICK'S trackers — thousands of hand rows, Meltwater paste tabs.
os · Ingest → code → score → snapshot. Coverage is one table. Today / Queue is the human job.
today · DAD on paper; lower funnel hard; GEO unclear; exposure-multiplier proxies still haunt the room.
os · Earn Rate + D/A/D · EIV ladder · AI Visibility · efficiency vs paid · Forecasting as bands.
today · No trusted shared metric for outcome pricing.
os · Immutable snapshots → approved reports → attainment. Beth sees cost-to-serve on the Dashboard.
memorize this once — ⌘K (Ctrl+K) gets you anywhere instantly; the Client Switcher at the top of the rail anchors whose work you're doing.
Workspace · agency
| Today | see what needs a human in the next hour (dashboard + action items) |
| Coverage | find / filter / export placements (the tracker) |
| Queue · Govern | code coverage vs rubric; the keystone — your accept/correct recomputes metric_snapshot |
| Automation | watch the Agent Constellation; clear Approvals; Summon a crew; tune autonomy dials |
| Approvals | human gate — approve a Robot Intern draft and it writes a real coding row |
| Reporting · Reports | AI draft + visible diff → named approve → send; QBR narrative grounded in snapshots |
| Opportunities | act on signals; ranked Media Contacts pitch matches with one-click draft; share-of-voice alerts |
| Media Contacts | journalist relationships; outreach drafts from Connector |
| Forecasting | calibrated forecast bands; GEO outlet targets; send to contract draft (persists for real) |
| Integrations | plug Meltwater, GA4, Google Ads, Meta; credentials in Vault |
| Health | anomaly narration in plain English; connector self-heal classifier |
| Data Governance | data rights registry; corpus pooling boundaries |
| Onboarding | wizard creates a client and lands it in the switcher immediately |
| Rubric | edit coding rubric; field weights; gold-set review |
| Admin | Account Pulse churn-risk board; Cross-Client Patterns; Forecasting review queue (admin seat only) |
| Playbook | learn the OS / pitch the OS / meet the crew |
Client Portal · brand
| Overview | 30-second truth (Earn Rate, Pulse, Spotlights) |
| Coverage | every placement, client-safe |
| Outcomes | lower funnel & efficiency as data connects |
| AI Visibility | AI-answer visibility with intervals |
| Connections | plug GA4 / budget / ads (with consent) |
| Reports | approved reports + methodology |
| How this works | what the numbers mean |
same OS — different seats, different rhythm.
Daily · Nate / Ryan / analysts
Switch client context (⌘K or Switcher) → Today → clear Action Items / Queue → spot-check Coverage → act on Opportunities alerts. Queue clears are the keystone: each accept/correct recomputes that client's books.
Weekly · account leads
Outcomes (same as client) → Reporting report with AI draft + diff review → named approve → send → digest lands in Client Portal.
Monthly · Beth + leads
Admin: Portfolio Dashboard · Account Pulse churn-risk · hours-saved / cost-to-serve · goal pace · (when data depth allows) Forecasting bands for contract conversation.
Integrations holds every hook. Credentials in Vault. Consent before sync.
| Plug | Unlocks |
|---|---|
| Meltwater / Muck Rack | Coverage ingest — RawItems → placements |
| CSV / Sheets drop | same path the Peet's · DICK'S · Albertsons pilots used |
| GA4 | earned sessions on Outcomes (aggregate only) |
| Search Console | branded search interest next to coverage |
| Google Ads / Meta Ads | earned-vs-paid efficiency |
| Budget CSV | pacing without touching ad accounts |
AI drafts. Humans ship. Approved drafts write real coding rows. Client-facing never auto-publishes.
Observe
crew watches and summarizes — nothing writes without you
Draft
writes land in Approvals for approve / reject / edit; approved Intern drafts write coding rows
Auto-internal
safe to apply inside Workspace; still gated for Client Portal
Never auto-client
hard stop — brand surfaces only after a named human
the single fact that changes a new hire's mental model about what their Queue work is for.
When you code an article in Queue — accept, correct, or override the Robot Intern — the system recomputes that client's metric_snapshot. Your review isn't busywork, it's the actual books. The numbers that land on Client Portal, in approved reports, and in Forecasting inputs all came from this loop. Conductor pre-sorts “needs your brain” vs “Conductor already checked these” so you spend judgment on exceptions, not trivialities. Bulk-accept clears the obvious ones in one keypress.
open Queue →same books in Workspace and Client Portal — method labels travel with every number
| KPI | What it means | Method |
|---|---|---|
| Earn Rate | Composite quality of earned work | scored |
| Discovery / Authority / Decision | DAD pillars | scored |
| Placements | Count of coverage items | counted |
| Reach | Audience estimate | modeled |
| EIV-Reach | Value of attention | modeled |
| Earned sessions | GA4 attributed | measured |
| Branded search | GSC interest | estimated / measured |
| GEO share | AI citation visibility | counted + interval |
| Efficiency | Earned outcomes vs PR $ / vs paid $ | mixed — always labeled |
| Goal pace | Progress to agreed goals | derived from snapshots |
Agency-only: agreement %, queue depth, cost-to-serve, feed lag.
Holding companies build internal OS layers because monitoring vendors don't run the agency. You get shared books, rubric coding with a gate, and Automation — a crew you manage with clicks.
| Need | Vendor default | Impact OS |
|---|---|---|
| Monitoring volume | Strong | Plug Meltwater — don't boil the ocean |
| Journalist DB | Cision/Muck Rack | Media Contacts from your wins + optional sync |
| Quality coding | Manual / generic AI | Your rubric (Peet's/DICK'S scores) + gate |
| Earned-value honesty | Still leaks exposure-multiplier proxies | EIV ladder + lint ban |
| Lower funnel | Rare | GA4 + Ads plugs |
| AI visibility 2026 | Vendor prompt sets | Client-owned AI Visibility prompts |
| Client portal | Weak in PR tools | Client Portal = same books; real client_admin login |
| AI agents | Slideware / hidden cron | Automation — visible crew, Constellation, Approvals writes real rows, Summon custom crew |
| Predictive | Overclaim risk | Calibrated bands + admin review gate; persists to campaign_forecast when accepted |
| Outcome pricing | Blocked | Audited snapshots → approved reports → campaign_forecast contracts |
Leapfrog: not more clips than Cision. Operating leverage + honesty + shared books + a crew you can manage.
The Workflow Coordinator routes. The Coding Assistant codes. The Intake Assistant enriches. The Quality Checker QAs. The Report Writer drafts. The AI Visibility Monitor runs GEO. The Opportunity Scanner, Outreach Assistant, and Forecasting Assistant draft growth work — humans send.
| Agent | Does | Waits for you when… |
|---|---|---|
| Conductor | Routes every run; pre-sorts Queue into 'needs your brain' vs 'already checked'; watches budget/policy | Something needs a human gate; policy would be violated |
| Robot Intern | Codes coverage vs rubric; approved drafts write real coding rows (not cosmetic) | Confidence is low; Steward flags; or you correct in Queue |
| Clipper | Enriches placements + spots coverage gaps | You confirm a miss |
| Steward | QA before books move; designed to gate snapshots (live wiring in progress) | Always in draft by default — never auto-approves |
| Scribe | Drafts Reporting packages with visible diff; QBR narratives; voice/tone check before send | Named human approves before client send |
| Mirror | GEO / AI citation checks with Wilson intervals | You publish AI Visibility |
| Scout | Opportunity signal detection + ranked Media Contacts pitch matches with grounded reasons | Offense stays internal until you act |
| Connector | Media Contacts outreach drafts; share-of-voice alert wiring | You hit send |
| Forecaster | Forecasting bands (never a point number); accepted forecasts persist to campaign_forecast; GEO outlet/term targets | Sample size + intervals clear; admin review before client-facing |
| Brief | TL;DR packs + pitch notes with visible diff | Human voice check |
moved out of roadmap — these are real, not planned.
| Feature | Where | What it does |
|---|---|---|
| Client Switcher + ⌘K | Rail + global | Searchable command-palette-backed switcher. Pinned / recent / at-risk sections. Health-tier tinted trigger. ⌘K jumps to any client, page, or action. |
| Agent Constellation | Automation | Live visual of who's working, on what. Summon an agent for a session-scoped custom crew. |
| Approvals writes real coding rows | Approvals | Approving a Robot Intern draft now writes a coding row — not cosmetic. |
| Queue → snapshot recompute (keystone) | Queue (Govern) | Your accept/correct moves metric_snapshot. Conductor auto-triage + bulk-accept. |
| AI draft with visible diff | Reports | TL;DR + Pitch Notes show exactly what a human edited vs the AI original. |
| QBR narrative generator | Reports | Grounded quarter narrative from snapshots, reusing the draft-diff pattern. |
| Health anomaly narration | Health (/health) | Plain-English 'what's actually happening' strip. Always honest about live vs sample. |
| Predictive pitch matching | Opportunities → Media Contacts | Opportunities signal → ranked Media Contacts targets with grounded reasons → one-click draft. |
| Forecast persistence | Forecasting + Admin | 'Send to contract draft' writes campaign_forecast (immutable). Admin review queue before client-facing. |
| GEO recommendation engine | Forecasting | Ranked outlet/term targets. Honest about cells that didn't clear the sample floor. |
| Account Pulse + Cross-Client Patterns | Admin | Churn-risk board + pooled agency-wide insights (sovereignty-safe, always discloses boundary). |
| Client login to Client Portal | Client Portal | client_admin / client_viewer roles work. Real auth, same approved books. |
| Onboarding wizard | Onboarding | Really creates the client; lands it in the switcher immediately. No manual wiring. |
| Voice / tone checker | Drafts (Reports) | Wired into Reporting before send. |
| Share-of-voice alerts | Opportunities | Fires when a competitor surges above threshold. |
| Connector self-heal classifier | Health / Integrations | Tells you which plugs need attention before data goes stale. |
honest about the real remaining gaps — not aspirational.
D.A.D. realignment (July 2026 framework)
The new D.A.D. deck redefines discovery/authority/decision as Share of Voice / Share of Citations / Ranked Position (by LLMs) — genuinely different metrics from what recompute_metric_snapshot computes today (reach+sentiment / feature-prominence / CTA-detection). Phased plan is scoped; needs a decision on how we source competitor volume before Phase 1 touches the live formula. See the gap analysis for the full breakdown.
Steward agent wiring
Steward QA before books move is the architectural design — the live end-to-end hook isn't fully wired yet.
Integrations action wiring
Integrations UI exists; full live credential-vault loop for adding new connectors needs completion.
Onboarding feed + tier persistence
Wizard creates the client; feed configuration and rubric-tier selection don't fully persist yet.
Forecasting GBM v1
Quantile regression upgrade (pinball loss, back-test coverage gate) — staged behind sufficient history depth.
Multi-touch attribution (Connect layer)
Shapley/Markov connect layer — staged behind GA4/Ads consent + data depth. Labeled when it ships.