Fifteen AI-native revenue workflows, built the way enterprises actually deploy them: every agent writes back to a system of record, passes a governance gate before it acts, and publishes a metric you can hold it to. No demos. Reference architectures.
Every workflow in this series runs its actions through the same four-stage control plane. This is the difference between an enterprise deployment and a weekend demo — and it is the first objection any CRO, CISO, or RevOps lead will raise.
Does this agent hold an explicit scope for this action, on this object, in this system? Scopes are declared per-agent, not inherited.
Every recommendation carries its sources: the signals, records, and transcript spans that produced it. No evidence, no action.
Actions above a declared risk threshold route to a named human owner. Low-risk actions auto-pass with logged rationale.
Every pass, flag, and block is written to an immutable log with actor, evidence hash, and outcome. Reviewable by ops, quotable to security.
Verdicts: pass / flag / block — enforced on every workflow below
All fifteen ship as working builds: reference architecture, operator control panel, eval metrics, failure modes — and a live scenario demo that runs through the gate.
Fuses intent, hiring, tech-stack, and funding signals to select the play — not just the lead. Human gate before send; disposition written back to CRM.
Research agents maintain living account dossiers, ICP fit scores, and auto-rebalanced territories with RevOps sign-off trails.
Inbound enrichment, qualification, routing, and booking under an enforced SLA, with escalation when confidence drops.
1:1 account pages, decks, and ads at scale — under a brand-governance layer with named approvers.
Job-change detection, multithreading-gap alerts, and auto-drafted re-engagement plays when a champion moves.
MEDDPICC or BANT scoring from call transcripts, email, and CRM hygiene — flagging at-risk deals before pipeline review, with evidence attached to every score.
Rep commit vs. signal-derived prediction, with variance explanations a CRO can read.
Battlecards auto-refreshed from win/loss transcripts and web monitoring — versioned so sales trusts them.
Pricing guardrails, legal-clause boundaries, e-signature handoff. Governance gates are the whole product.
Transcript mining into positioning and messaging change recommendations, closing the loop to marketing.
Health scoring from product telemetry, support, and sentiment — renewal-risk playbooks triggered with CSM approval.
Value-realization decks from usage data — on-brand, human-reviewed before any customer sees them.
Identity resolution, dedupe, and enrichment governance across CRM, MAP, and warehouse. Unglamorous, desperately in demand.
PQL scoring from product events with an explainable "why now" attached to every routed lead.
Natural-language querying over the revenue warehouse with attribution logic — answers that are defensible, not just fluent.
Every architecture names a capability ("reasoning model with 100k context"), never a vendor lock. Swap the model; the workflow holds.
An agent that only produces a Slack message is a toy. These write dispositions, scores, and merges back to CRM, MAP, and warehouse.
Each build ships with the ways it breaks and the controls that catch it. That section is what tutorial content never includes.
Most "AI SDR" builds automate the send. This one automates the decision before the send: fusing intent, hiring, tech-stack, and funding signals into a play selection — which sequence, which persona, which angle, and whether to reach out at all. A human approves the play; the agent handles the mechanics; every disposition writes back to CRM.
How the flow runs — live
Suppression is a valid exit at any stage — logged with a reason, never silent.
Teams bought AI email tools, 10x'd volume, and watched reply rates collapse while domains got flagged. The bottleneck was never sending capacity — it was knowing which account deserves which motion this week, and having proof for why.
Signals stop being a column in a spreadsheet and become the input to a routing decision. A funding event at an ICP-fit account with a fresh VP hire triggers a different play than lone website intent. Low-conviction accounts get suppressed — deliberately doing nothing is a valid, logged output.
| Decision | Choice | Rationale |
|---|---|---|
| Plays are a curated library | Agent selects; humans author | The agent never invents a motion. Marketing and sales leadership own the playbook; the agent's job is pattern-matching signals to it. This keeps message quality governable. |
| Suppression is an output | "Do nothing" is logged with a reason | The costliest outbound failure is burning a future buyer with a bad touch. Below-threshold conviction produces a logged suppression, reviewable in the same audit trail. |
| Signal conviction decays | Time-decay per signal class | A funding round is warm for 90 days; website intent for 7. Fusion scores decay on class-specific half-lives so stale signals can't trigger fresh sends. |
| First touch is human-gated | SDR approves message 1; follow-ups auto-run | The highest-risk moment is the first impression. Gating only that touch keeps throughput high while keeping a named human accountable per account. |
Toggle the signals observed for a sample account. The fusion agent scores conviction, the selector picks a play — or suppresses — and the gate rail logs the full decision.
No signals observed. Suppression logged — this account stays untouched, and that decision is auditable.
Governance audit log
Agents don't remove operators — they change what operators watch. This is the console spec: who owns it, what they check on what cadence, what pages them, and how they pull the plug.
Intent providers change taxonomies; a topic rename silently zeroes a signal class. Mitigation: per-class volume monitors alert when a signal source deviates ±40% from its 30-day baseline.
Correlated signals (funding → hiring → press coverage) triple-count one underlying event. Mitigation: fusion agent clusters signals by root event before scoring; one event contributes once.
SDRs rubber-stamp first-touch approvals at volume, hollowing out the gate. Mitigation: approval-rate telemetry — if an SDR approves >95% in under 5 seconds each, the sample rate of required reviews increases and the pattern is surfaced to the manager.
The agent keeps selecting a play whose reply rate quietly died two months ago. Mitigation: per-play performance floors — plays under threshold for 30 days are auto-benched and flagged to their human author.
Research agents maintain a living dossier per account — firmographics, initiatives, stack, org changes — score ICP fit against a versioned rubric, and propose territory rebalances that RevOps signs off as diffs. Territories stop being an annual spreadsheet war.
How the flow runs — live
Fit scores recompute when facts change; the territory map only moves through an approved diff.
Territory planning happens once a year on stale data. By Q2, half the “A-tier” accounts have reorged, churned a champion, or pivoted — and reps prospect against a map of last year's market.
Dossiers refresh continuously with cited evidence. Fit scores recompute when the underlying facts change, and rebalance proposals arrive as diffs with rationale — RevOps approves the diff, not a 4,000-row spreadsheet.
| Decision | Choice | Rationale |
|---|---|---|
| Dossiers carry citations | Every claim links to its source with a fetch date | An uncited dossier is a hallucination risk wearing a suit. Citations make dossiers auditable and let reps verify a “fact” before quoting it to a prospect. |
| Rebalances ship as diffs | Propose changes; never rewrite the map | Wholesale re-carving destroys rep trust and pipeline continuity. Diffs with rationale keep every change explainable and reversible. |
Three curated scenarios — one that passes, one that flags for a human, one that the gate blocks. Watch the verdict, the evidence rows, and the audit log.
Governance audit log
Agents don't remove operators — they change what operators watch. Console spec: who owns it, what they check, what pages them, and how they pull the plug.
An outdated initiative stays in the dossier after the source page dies, and a rep leads a call with it. Mitigation: Citations carry fetch dates and class-specific TTLs; expired claims are re-verified or dropped, never silently kept.
Volatile signals make fit scores oscillate, thrashing territory proposals weekly. Mitigation: Hysteresis — rebalance proposals require a sustained score change across two refresh cycles before they're raised.
The agent over-weights PR language and inflates fit for loud companies. Mitigation: Rubric weights favor verifiable operational signals — headcount, stack, filings — over announcement volume.
Inbound leads get enriched, qualified, routed, and offered a meeting inside an enforced SLA — with the agent escalating to a human the moment confidence drops, not after the lead has gone cold.
How the flow runs — live
Every lead exits with a logged decision: booked, routed to a named human, or politely declined.
Median B2B inbound response is measured in hours; conversion decays in minutes. Routing rules break silently, leads sit in queues, and nobody owns the gap between form-fill and first touch.
Every lead gets a decision in minutes, and every path is logged. The SLA is enforced by the system — misses page a human instead of vanishing into a queue nobody watches.
| Decision | Choice | Rationale |
|---|---|---|
| Framework is segment config | BANT-lite for velocity motions; full rubric for enterprise | One rubric can't serve a $99 self-serve signup and a 2,000-seat inquiry. Segment-level config with a shared extractor keeps it governable. |
| Confidence-gated automation | The bot books only what it's sure about | A wrong auto-booked meeting costs an AE hour and buyer goodwill. Below threshold, the lead routes to a human — with the evidence already assembled. |
Three curated scenarios — one that passes, one that flags for a human, one that the gate blocks. Watch the verdict, the evidence rows, and the audit log.
Governance audit log
Agents don't remove operators — they change what operators watch. Console spec: who owns it, what they check, what pages them, and how they pull the plug.
A wrong company match on a generic domain routes an enterprise lead into the SMB queue. Mitigation: Low-confidence enrichment downgrades to human routing — the system never guesses a tier.
Auto-booking against stale availability double-books AEs and burns credibility on day one. Mitigation: Booking writes are transactional with a live availability re-check at confirm time.
Bot traffic floods qualification, pollutes SLA metrics, and buries real leads. Mitigation: A pre-qualification filter layer quarantines suspected bots and excludes them from SLA math.
Generates 1:1 account pages, deck variants, and ad copy at scale — inside a brand-governance layer where templates, claims, and tone are locked by marketing, and every asset traces to an approved source block.
How the flow runs — live
The agent rearranges approved truth into personalized layouts — it never invents a claim.
ABM personalization at scale means either 3 accounts done well or 300 done with a mail-merge that fools no one — and the moment generation scales, brand and legal lose the ability to review what ships.
Personalization comes from account evidence slotted into governed templates. Claims come from an approved library. Marketing reviews launch waves with spot-samples, not individual sentences.
| Decision | Choice | Rationale |
|---|---|---|
| Claims are a library, not a generation task | The agent composes from approved statements | Product claims and compliance lines are where generated content becomes a legal problem. Library composition caps the blast radius at layout, not truth. |
| Per-wave approval granularity | Approve a campaign wave with spot-samples, not each asset | Asset-level review recreates the bottleneck automation was meant to remove. Wave-level review keeps humans meaningful without throttling throughput. |
Three curated scenarios — one that passes, one that flags for a human, one that the gate blocks. Watch the verdict, the evidence rows, and the audit log.
Governance audit log
Agents don't remove operators — they change what operators watch. Console spec: who owns it, what they check, what pages them, and how they pull the plug.
Over-specific references — “saw your CFO changed 12 days ago” — read as surveillance, not relevance. Mitigation: An evidence-use policy classifies every signal as usable-explicitly, usable-implicitly, or never-referenced.
One winning template gets cloned across segments until it saturates and dies everywhere at once. Mitigation: Per-segment performance floors plus enforced template diversity quotas.
Pages reference an initiative the account abandoned last quarter. Mitigation: Assets inherit dossier citation TTLs; an expired fact triggers regeneration or takedown of the asset.
Watches the people, not just the accounts: job changes, promotions, single-threaded deals, and dark spots in the buying committee — and drafts the re-engagement play the moment a champion lands somewhere new.
How the flow runs — live
The agent watches and drafts. The rep owns every touch — relationship capital is theirs, not the system's.
Champions change jobs mid-deal and the CRM finds out at the loss review. Multithreading gaps are invisible until the one contact goes silent. Relationship risk is the most predictive deal signal nobody instruments.
The graph makes relationship risk a first-class field: thread count, role coverage, champion stability on every deal. A champion move triggers two drafted plays — protect the old deal, open the new door — both waiting on rep approval.
| Decision | Choice | Rationale |
|---|---|---|
| Two-source verification | Job changes confirmed across two sources before any action | False job-change signals trigger embarrassing outreach. Two-source verification trades a day of latency for credibility. |
| Plays draft, never send | Relationship moves are always human-approved | The agent's job is watching and drafting. The rep's job is the relationship. The gate keeps that boundary structural, not aspirational. |
Three curated scenarios — one that passes, one that flags for a human, one that the gate blocks. Watch the verdict, the evidence rows, and the audit log.
Governance audit log
Agents don't remove operators — they change what operators watch. Console spec: who owns it, what they check, what pages them, and how they pull the plug.
Tracking drifts from professional signals into territory that damages trust if surfaced. Mitigation: A hard signal allowlist — public professional data only — reviewed by legal, with every lookup logged.
Contacts change roles internally and the map silently misleads deal strategy. Mitigation: Engagement-pattern anomaly detection prompts re-verification of stale nodes.
Every profile update pings reps until they mute the whole system. Mitigation: Alerts are ranked and budgeted per rep per week; below-threshold changes batch into a digest.
Scores every open opportunity against MEDDPICC or BANT — your choice of framework — using call transcripts, email threads, and CRM field hygiene as evidence. Flags at-risk deals before pipeline review, and never updates a record without passing the gate.
How the flow runs — live
Scores auto-pass; stage-change suggestions always stop at the gate for a named human.
Deal qualification lives in a rep's head and a stale CRM picklist. Managers inspect 40 deals in a 60-minute call, relying on whoever tells the best story. Slipped deals were "surprises" that had been visible in the transcript for six weeks.
Every deal carries a framework score computed from primary evidence — what the buyer actually said, what fields actually changed, who is actually multithreaded. Pipeline review starts from the flags, not the anecdotes. Reps argue with evidence, which is the point.
| Decision | Choice | Rationale |
|---|---|---|
| Framework is configurable | MEDDPICC (default) or BANT | Enterprise teams run MEDDPICC; velocity and mid-market teams still run BANT. The rubric is a config object, not hardcoded logic — the same evidence extractor feeds both. |
| Evidence granularity | Transcript spans, not summaries | A score of "Champion: weak" links to the exact 40-second span where the champion hedged. Summaries hallucinate; spans are checkable. |
| Write-back scope | Score fields only; stage changes are suggestions | The agent never moves a deal stage. It proposes; a human disposes. This is the single design choice that gets enterprise security to yes. |
| Model requirement | Any reasoning model, ≥100k context | Vendor-neutral. Transcript batches fit in context; no fine-tuning required. Swap providers without touching the rubric. |
Pick a framework, pick a sample deal, run the inspection. Watch the gate rail: the score write auto-passes; the stage-change suggestion routes to approval. Every event lands in the audit log.
Governance audit log
The scorer is only trusted while its precision is monitored. This console spec defines who watches what, on what cadence, and where the manual overrides live.
A buyer says the right words on a call; the deal still dies. Verbal signals inflate scores. Mitigation: behavioral evidence (stakeholders added, security review started, redlines returned) is weighted above verbal evidence in the rubric config.
Reps learn what the extractor listens for and coach the language on calls. Mitigation: precision sampling audits scores against outcomes quarterly; criteria weights are recalibrated against actual win/loss, not call content.
BANT scoring on a 9-month enterprise cycle produces confident nonsense — budget rarely exists at Stage 1. Mitigation: framework is set per pipeline segment, not per org; the config warns when deal size and framework disagree.
Managers ignore a number they can't interrogate. Mitigation: no score ships without the Risk Narrator paragraph and clickable evidence spans. If the narrator's confidence is low, the deal is flagged for manual review instead of scored.
Runs the forecast twice — once from rep commits, once from deal evidence — and makes the variance the artifact: every gap between what reps say and what signals show gets a written explanation a CRO can interrogate.
How the flow runs — live
Both forecasts are preserved. The tension between them is the product.
Forecast calls run on assertion: reps defend numbers, managers apply gut haircuts, and the CRO triangulates vibes. When the quarter misses, nobody can reconstruct which assumption broke.
Two forecasts, one variance report. Each gap names its driver — “3 commit deals show no paper-process evidence” — and snapshots are versioned, so misses become diagnosable: either the model was wrong, or the pipeline was.
| Decision | Choice | Rationale |
|---|---|---|
| Read-mostly by design | The agent never edits rep commits | A system that overwrites rep numbers gets gamed or ignored. Preserving both views keeps the tension visible — and the tension is the product. |
| Explanations over scores | Every variance ships with prose and evidence links | A CRO won't act on “model says 0.62.” A CRO will act on “these 4 commit deals carry zero economic-buyer engagement.” |
Three curated scenarios — one that passes, one that flags for a human, one that the gate blocks. Watch the verdict, the evidence rows, and the audit log.
Governance audit log
Agents don't remove operators — they change what operators watch. Console spec: who owns it, what they check, what pages them, and how they pull the plug.
Leadership treats the signal number as truth; reps stop maintaining CRM because “the model knows.” Mitigation: Both errors are published every quarter. The model earns authority publicly or loses it publicly.
Reps under-commit to beat the model and bank easy wins. Mitigation: Variance is reported in both directions; chronic under-commit patterns surface by rep.
A broken transcript sync silently degrades the signal model for weeks. Mitigation: Input freshness is printed on page one of every report — a forecast with stale inputs says so before it says anything else.
Battlecards that update themselves from win/loss transcripts, product releases, and pricing-page diffs — versioned like code, so a rep quoting a card knows exactly when each claim was last verified.
How the flow runs — live
Cards ship like software: diffs, changelogs, release approval, rollback.
Battlecards are written at launch, trusted for a month, and quietly wrong for a year. Reps learn to ignore them — then lose deals on objections the last ten transcripts already answered.
Cards become living documents. New objections from calls propose updates; pricing-page changes open diffs; PMM approves releases like pull requests. A rep sees “verified 6 days ago,” not “created 2024.”
| Decision | Choice | Rationale |
|---|---|---|
| Cards are versioned artifacts | Diffs, changelogs, release approval, rollback | Trust dies from silent staleness. Versioning makes freshness visible — and makes rollback possible when a claim turns out wrong. |
| Transcript claims quarantine | Competitor claims heard on calls are hearsay until verified | Buyers repeat competitor FUD inaccurately. A hearsay lane prevents laundering unverified claims into official positioning. |
Three curated scenarios — one that passes, one that flags for a human, one that the gate blocks. Watch the verdict, the evidence rows, and the audit log.
Governance audit log
Agents don't remove operators — they change what operators watch. Console spec: who owns it, what they check, what pages them, and how they pull the plug.
Unverified transcript claims leak into cards as facts and a rep repeats them to a buyer who knows better. Mitigation: The hearsay quarantine plus PMM verification gate — nothing crosses without evidence.
A competitor redesigns their site; the diff monitor silently returns no changes for months. Mitigation: Zero-change-for-N-days alerts per monitor — silence is treated as a signal, not a comfort.
One angry, memorable loss reshapes the whole positioning. Mitigation: Objection patterns require minimum incident counts before proposing card changes.
Assembles proposals and SOWs from approved clause and pricing blocks — discount authority, legal boundaries, and payment-term rules enforced at composition time, then handed to e-signature with the approval chain as the audit trail.
How the flow runs — live
Policy violations can't be drafted — the un-draftable discount never becomes a negotiation.
Proposal assembly is where deals slow down and margin leaks out: reps clone old docs with stale terms, discounts get invented under quarter-end pressure, and legal finds the non-standard clause after signature.
Documents compose from the current block library only. Discounts beyond authority route to the right approver automatically; clause deviations route to legal before the buyer ever sees the doc.
| Decision | Choice | Rationale |
|---|---|---|
| Composition-time enforcement | Policy violations can't be drafted, not merely caught later | Catching a bad discount at approval is a delay. Making it un-draftable removes the negotiation-by-fait-accompli pattern entirely. |
| Tiered approval routing | Risk decides the approver, not the org-chart default | Flat “manager approves everything” creates rubber stamps. Discount depth, clause deviation, and payment terms each map to named authorities. |
Three curated scenarios — one that passes, one that flags for a human, one that the gate blocks. Watch the verdict, the evidence rows, and the audit log.
Governance audit log
Agents don't remove operators — they change what operators watch. Console spec: who owns it, what they check, what pages them, and how they pull the plug.
Big deals demand terms outside the library, and the workflow gets bypassed entirely. Mitigation: An explicit manual-mode lane with mandatory legal pairing — visible and logged, not shadow paperwork.
Pricing blocks lag a packaging change and quote last quarter's SKUs. Mitigation: Blocks carry effective dates; expired blocks hard-fail assembly rather than silently composing.
Tiered approvers click approve without reading, and the gate hollows out. Mitigation: Approval dwell-time telemetry — the same rubber-stamp detection pattern used on outbound first-touch review.
Mines every closed deal's transcripts, threads, and timeline into structured cause analysis — then closes the loop by proposing specific positioning, product, and process changes with the evidence attached and minimum-incident thresholds enforced.
How the flow runs — live
Deal-level archaeology first; themes only past minimum evidence. Anecdotes don't reshape strategy.
Loss reasons in CRM are a picklist lie — “price” means the rep stopped asking why. Real causes live in transcripts nobody re-reads, so the same losses repeat quarterly with fresh optimism.
Each closed deal gets a written cause analysis citing its evidence. Themes only emerge past minimum incident counts, and they arrive as routed proposals: “move security review earlier — 7 of 9 enterprise losses stalled there.”
| Decision | Choice | Rationale |
|---|---|---|
| Deal-level before pattern-level | Archaeology first, themes second | Aggregating bad per-deal analysis produces confident nonsense. Each deal's causal story is reviewable on its own before it feeds a pattern. |
| Minimum-N for themes | No theme ships on fewer than 5 incidents | One vivid loss shouldn't reshape strategy. Thresholds keep the engine from amplifying anecdotes into initiatives. |
Three curated scenarios — one that passes, one that flags for a human, one that the gate blocks. Watch the verdict, the evidence rows, and the audit log.
Governance audit log
Agents don't remove operators — they change what operators watch. Console spec: who owns it, what they check, what pages them, and how they pull the plug.
Analyses narrate inevitability instead of identifying decision points. Mitigation: Every analysis must cite the earliest evidence of the losing factor — the question is “when could we have known,” not “what went wrong.”
Analyses read as blame; reps stop cooperating and the evidence dries up. Mitigation: Process-and-positioning framing in public themes; individual coaching insights route privately to managers only.
Only closed deals get mined — but stuck, no-decision deals hold the real lessons. Mitigation: A quarterly no-decision cohort analysis runs alongside win/loss with the same rigor.
Fuses product telemetry, support sentiment, invoice behavior, and stakeholder changes into account health that means something — and triggers renewal-risk or expansion playbooks with CSM approval, months before the renewal date makes it urgent.
How the flow runs — live
Every score shows its drivers and their trend. Arguable beats accurate-but-ignored.
Health scores today are a traffic light nobody trusts: green accounts churn, red accounts renew, and CSMs learn the score is theater. Real risk hides in trajectories — seat stagnation, champion exit, support tone shift.
Health becomes an explained number: every score shows its drivers and their direction. Risk plays trigger with enough lead time to matter, expansion signals route with the same rigor, and the CSM stays the owner of every touch.
| Decision | Choice | Rationale |
|---|---|---|
| Explainability over accuracy | A slightly worse model CSMs trust beats a black box they ignore | Adoption is the metric. Driver-level explanations make the score arguable — and arguable means used. |
| Segment-specific weighting | Enterprise health ≠ SMB health | Login frequency predicts SMB churn and means nothing at 5,000 seats. Weights are per-segment config, versioned, recalibrated against actual renewal outcomes. |
Three curated scenarios — one that passes, one that flags for a human, one that the gate blocks. Watch the verdict, the evidence rows, and the audit log.
Governance audit log
Agents don't remove operators — they change what operators watch. Console spec: who owns it, what they check, what pages them, and how they pull the plug.
The model misses relationship-driven exits that telemetry can't see. Mitigation: Champion-tracker signals (WF-05) feed health directly; relationship risk is a first-class driver, not a footnote.
Every dip triggers a play until CSMs mute the sentinel entirely. Mitigation: Play budgets per CSM per week; sub-threshold changes batch into digests instead of interrupts.
Manual overrides hide risk to protect the renewal forecast. Mitigation: Overrides expire, require reasons, and are reported alongside the un-overridden score — hiding risk becomes visible work.
Builds the value-realization story per account — usage against goals, ROI evidence, roadmap fit — as an on-brand deck the CSM edits and owns. Prep drops from a day to an hour, and no unreviewed slide ever reaches a customer.
How the flow runs — live
The deck ships to the CSM, never to the customer. Human review is the product's contract.
QBR prep eats a CSM day per account, so QBRs happen for the loudest 20% — and half of prep is hunting numbers that were true last quarter. The accounts that quietly needed the story never get one.
Every account gets a draft deck assembled from live data against its own success plan. The CSM's day becomes an hour of judgment — sharpening narrative, cutting what doesn't serve the meeting — and coverage stops being rationed.
| Decision | Choice | Rationale |
|---|---|---|
| Draft, never final | The deck ships to the CSM, not the customer | A wrong number in front of a customer costs a renewal. Human review is enforced by the gate, not by policy hope. |
| Goals-anchored, not activity-anchored | Decks argue value against the success plan | Usage graphs without goal context are noise. If no success plan exists, the deck's first page says so — which is its own finding. |
Three curated scenarios — one that passes, one that flags for a human, one that the gate blocks. Watch the verdict, the evidence rows, and the audit log.
Governance audit log
Agents don't remove operators — they change what operators watch. Console spec: who owns it, what they check, what pages them, and how they pull the plug.
The composer leads with flattering numbers and buries the decline the customer already noticed. Mitigation: The template mandates a balanced scorecard section — declines render with the same prominence as wins.
Cross-customer comparisons expose confidential patterns. Mitigation: Benchmarks only from aggregate cohorts above minimum size, framed with legal-approved blocks.
Every customer sees the same deck shape; it reads as mail-merge and devalues the meeting. Mitigation: Narrative variants per lifecycle stage; edit-distance telemetry drives template evolution.
Every GTM AI initiative dies on the same rock: the CRM has three records for the same buyer, the MAP disagrees with all of them, and the warehouse trusts none of it. This workflow runs identity resolution, dedupe, and enrichment governance as a continuous, gated, reversible process — the least glamorous agent in the stack, and the one everything else depends on.
How the flow runs — live
Every merge carries a 90-day unwind snapshot. Consent fields never take the automatic path.
Duplicate contacts split engagement history, so lead scoring undercounts. Company records fragment across legal names, domains, and subsidiaries, so territory rules misfire and attribution double-counts. Quarterly "data cleanup projects" fix a snapshot; the rot resumes Monday.
Identity becomes a continuously reconciled graph, not a quarterly project. High-confidence merges execute automatically and reversibly; ambiguous ones queue for a human with the evidence laid out. Enrichment writes obey a field-level authority matrix — no vendor overwrites a human-verified value, ever.
| Decision | Choice | Rationale |
|---|---|---|
| Merges are reversible | Full pre-merge snapshots, 90-day unwind window | The catastrophic failure of dedupe automation is a wrong merge you can't undo. Reversibility converts a career-ending risk into a Tuesday ticket, which is what lets confidence thresholds be aggressive. |
| Consent is sacred | Opt-out/consent fields excluded from auto-merge | Merging an opted-out record into an opted-in one is a compliance incident, not a data quality issue. These fields always route to a human, regardless of match confidence. |
| Field-level survivorship | Authority matrix: human-verified > 1st-party > vendor | Record-level "newest wins" destroys curated data. Each field carries a source rank and timestamp; enrichment vendors can fill nulls but never overwrite higher-authority values. |
| Continuous, not batch | Event-driven on record create/update + nightly sweep | Quarterly cleanups treat symptoms. Catching a duplicate 30 seconds after form-fill costs one merge; catching it 90 days later costs broken attribution, misrouted leads, and a confused buyer. |
Three candidate pairs from a real-world-shaped queue. Run the janitor: watch one auto-merge, one route to human review, and one get correctly left alone — with the reasoning logged for each.
rec_0041: priya.sharma@meridianlog.com · Meridian Logistics · VP Ops
rec_2210: p.sharma@meridianlog.com · Meridian Logistics Pvt Ltd · VP Operations
rec_0788: j.chen@gmail.com · (no company) · signed up for webinar
rec_1904: james.chen@vertexhealth.com · Vertex Health · Dir. Analytics · opted out
rec_0902: a.rao@nimbuspay.com · NimbusPay · CFO
rec_3315: a.rao@nimbuspay.io · NimbusPay Labs · Founder
Governance audit log
Identity automation without an operator console is how CRMs get destroyed at 2 a.m. This spec defines the watch rotation, the alarms, and the undo.
Probabilistic matching merges a parent company with its subsidiary; territory and attribution both break. Mitigation: corporate-hierarchy signals (distinct billing entities, distinct domains, legal suffixes) act as merge blockers, not just score reducers.
An enrichment sync misconfigured once quietly overwrites thousands of human-verified titles. Mitigation: field authority matrix enforced at the write layer, not in sync settings — plus a daily diff report of authority-rank violations (target: zero).
A 92% threshold tuned on last year's data performs differently after an acquisition doubles record volume from a new region. Mitigation: monthly precision/recall recalibration on a labeled sample; thresholds are versioned config with change approval, like code.
The 70–92% queue grows faster than RevOps can clear it; ambiguous pairs rot. Mitigation: queue-depth SLO with alerting; batch-adjudication UI groups similar pairs; sustained overflow auto-raises the review threshold and reports the trade-off.
Scores product-qualified leads from real usage events and hands sales an explainable “why now” — which behaviors fired, what they predict, what to say — with consent and product-trust boundaries checked before any usage data reaches a sales record.
How the flow runs — live
The PQL entity is the workspace, not the person. Consent boundaries are gate checks, not assumptions.
PLG companies drown sales in “PQLs” that are just active users, torching trust in the motion. Meanwhile actual buying signals — a workspace hitting seat limits, an admin reading SSO docs — route nowhere.
PQLs ship with their evidence: “hit seat cap twice this week, invited a finance role yesterday, viewed pricing 3×.” Reps get a conversation starter, not a naked score — and data-use rules are enforced at handoff, not assumed.
| Decision | Choice | Rationale |
|---|---|---|
| Consent boundary is a gate check | Usage data crosses to sales only within stated terms | PLG trust is the asset. Data-use rules — per plan tier, per region — are enforced at handoff, logged, and auditable, not assumed. |
| Explainability is mandatory | No naked scores reach a rep | A score without a story gets ignored; a story without evidence gets distrusted. Every PQL carries both. |
Three curated scenarios — one that passes, one that flags for a human, one that the gate blocks. Watch the verdict, the evidence rows, and the audit log.
Governance audit log
Agents don't remove operators — they change what operators watch. Console spec: who owns it, what they check, what pages them, and how they pull the plug.
A product redesign changes event semantics; the model keeps scoring ghosts of the old flow. Mitigation: Event schema versioning — unversioned changes quarantine the affected triggers until re-validated.
Reps blast every PQL and poison the product-led goodwill the motion depends on. Mitigation: Trigger-matched outreach templates, volume caps, and product-team visibility into sales touches on PLG accounts.
Scoring individuals when the buying unit is the workspace produces nonsense handoffs. Mitigation: The PQL entity is the workspace; people are contacts within it, with inferred roles explicitly labeled as inferred.
Natural-language questions over the revenue warehouse, answered only through governed metric definitions and with the SQL always shown — so “what's our CAC payback by segment” returns the number finance would sign, not a fluent guess.
How the flow runs — live
Questions outside governed definitions get “define it first,” not improvisation. Refusal is a feature.
Every GTM team runs on numbers that disagree: marketing's CAC, finance's CAC, and the board deck's CAC are three numbers. Self-serve BI made querying easy and consistency worse; LLM-over-SQL without governance makes confident wrongness instant.
The copilot answers only through the semantic layer. Ambiguous questions get a clarifying choice — “pipeline-influenced or closed-won attribution?” — instead of a silent assumption. Every answer shows its SQL and cites its definition version.
| Decision | Choice | Rationale |
|---|---|---|
| Refusal is a feature | Ungoverned questions get “define it first,” not improvisation | The failure mode of analytics copilots is confident answers to ungoverned questions. Refusing cheaply beats retracting expensively. |
| Read-only, PR-governed definitions | The copilot can't create metrics; humans version them | Metric sprawl is how the three-CAC problem started. Definitions change through review, and the copilot always reflects the current merged truth. |
Three curated scenarios — one that passes, one that flags for a human, one that the gate blocks. Watch the verdict, the evidence rows, and the audit log.
Governance audit log
Agents don't remove operators — they change what operators watch. Console spec: who owns it, what they check, what pages them, and how they pull the plug.
An ambiguous question gets one plausible interpretation of several, and the wrong number ships to a board deck. Mitigation: Ambiguity detection forces a clarifying choice — assumptions are never silent, ever.
The warehouse schema changes under the layer; queries succeed and mean something different. Mitigation: Contract tests between layer and warehouse run on every deploy; drift pauses the copilot automatically.
Teams cite copilot answers in board materials with no audit trail. Mitigation: Every answer carries definition version + timestamp; exports include full lineage by default.
Revenue Operations and AI Agentic Workflows for B2B companies in transformation. We build the machine. You run the company.