Compliance
Gates before sends
Consent, suppression, DNCR and calling windows are evaluated by deterministic code before any channel fires.
Accelerate · Outbound control plane
An outbound control plane that books qualified calls.
Accelerate sources against your ICP, verifies eligibility, qualifies with evidence, runs sequences, escalates to voice and books into GoHighLevel — with every transition recorded.
Deterministic compliance gates
Idempotent execution
Replaceable vendors
Sequence run SQ-2041
BookedSarah Whitlam
-
Eligibility Email eligible · consent evidence on file
Passed -
Qualification 78/100 · evidence verified · confidence 0.86
Qualified -
Sequence Email 1 delivered · positive reply received
Replied -
Voice DNCR checked · calling window respected
Completed -
Booking Tue 10:30 · synced to GoHighLevel
Synced
Every transition recorded
Idempotent sends
Full history in Postgres
Audit · booking synced
appt_8812 · webhook verified
Why it’s serious
An engine, not a bot with an API key
A vendor list is not a product. Accelerate owns the parts that fail silently elsewhere — state, eligibility, retries, evidence — so nothing important depends on a chatbot’s mood.
Execution
Idempotent by default
Every send, call and booking carries an idempotency key — a retry can never double-touch a prospect.
Judgment
Evidence-backed qualification
Scores arrive structured, with evidence and confidence. Low confidence routes to a human, not onward.
Architecture
Vendors stay replaceable
Apollo, Retell, Twilio and GoHighLevel sit behind adapters. The engine keeps its own record either way.
Scope
Everything outbound needs, owned in one place
Each subsystem exists because its absence has a known cost — bad leads, double sends, legal exposure, or a CRM that quietly drifts from the truth.
01Prospect acquisitionSearch, import, dedupe and enrichment — source provenance kept for every data point.
02QualificationRubric, structured score, confidence and manual override — never a vague “good lead”.
03ComplianceConsent, suppression, DNCR, quiet hours and immutable evidence before any channel fires.
04Sequence orchestrationDurable state machine with delays, retries, branching and rate limits — not fragile cron.
05Channel adaptersEmail, SMS and voice normalized into one internal event model, so vendor logic never leaks.
06Conversation intelligenceReply intent, objections, summaries and next-action suggestions — triage without the triage.
07BookingAvailability lookup, reservation and conflict handling, synced to your calendar.
08CRM syncContacts, conversations, pipeline and appointments mirrored for the operator team.
09ExperimentationVersioning, attribution and controlled rollout — so “self-improving” is measured, not claimed.
10ObservabilityEvents, webhook replay and audit history — a real answer to “why did this happen?”
Lead lifecycle
One lead, end to end
Follow a prospect through the engine. Every state transition is decided by code; every judgment inside a step is made by a model — and both leave a record.
Book a pilot reviewStage 01 · Prospect acquisition
The prospect enters with provenance.
Leads arrive from a defined ICP query — never a scraped pile. Duplicates are removed before a single enrichment credit is spent.
Prospectprs_8f2a91
- Name & title
- Sarah WhitlamOperations Director
- Company
- Meridian ElectricalIndustrial services · AU
- Source
- Apollo search icp_au_opssourceRecordId retained for replacement later
- Created
- 12 Feb, 09:38 AESTcontact points stored with verification state
Dedupe runs first — paid enrichment is never consumed on a duplicate.
Stage 02 · Enrichment
Every claim keeps its source.
Company-site research fills in services, sector, locations and recent activity. Each field is stored as a claim with origin, time and confidence.
DataClaims4 stored
- Role
- Operations DirectorApollo · 12 Feb 09:41 · confidence 0.94
- Sector
- Industrial servicesWebsite research · 12 Feb 09:44 · 0.81
- Signal
- New Brisbane depot — careers pageWebsite crawl · 12 Feb 09:47 · 0.88
- Verified, single sourceContactPoint · provider ID kept behind adapter
Field-level provenance means any statement in outreach can be traced back to where it came from.
Stage 03 · Compliance gate
Eligibility is decided before outreach, per channel.
A deterministic rules engine evaluates each channel against consent evidence, suppression state and calling windows. The result is explicit — with reasons.
EligibilityDecisionevaluated 09:52
- EligiblelegalBasis: ConsentEvent cev_3092 · inbound enquiry, 3 Nov
- SMS
- Held — consent requiredgate blocks the SMS branch until evidence exists
- Voice
- PendingDNCR check queued · window weekdays 09:00–18:00 AEST
“Apollo has an email” never equals “consented”. Eligibility is recomputed, not assumed.
Stage 04 · AI qualification
A structured verdict, not a vibe.
The model answers one precise business question against your rubric and returns machine-validated output. Evidence is required; confidence decides routing.
// Qualification · prs_8f2a91 · rubric v7 · model routed via OpenRouter
{
fitScore: 78, // 0–100 against ICP rubric
need: "yes",
decisionMaker: true,
timing: "soon", // depot expansion underway
evidence: [
"Owns OH&S compliance for 3 sites",
"Careers page: new Brisbane depot, Q3"
],
confidence: 0.86,
recommendedAction: "continue"
}
Low confidence routes to manual review — a human decides, and the override is logged. Prompt version and model are stored with every decision, so results stay replayable.
Stage 05 · Sequence orchestration
A durable state machine, not a chain of timers.
Each step runs with its own execution ID and idempotency key, so a retry can never double-send. An inbound reply cancels scheduled steps immediately.
Sequence v3 · enrollment enr_2041
Step 1 EMAIL intro delivered ✓
Wait 2 business days
Reply positive → QUALIFY detected, pending steps cancelled
opt-out → SUPPRESS suppression written first
none + SMS eligible → SMS
Wait 1 business day
Voice eligible → AI CALL gated by eligibility + DNCR
Book qualified → BOOKED
else → END
Global and per-domain rate limits, lead cooldowns and unsubscribe handling are explicit configuration — not emergent behaviour.
Stage 06 · Booking & CRM sync
The call lands in the workflow you already run.
Availability is checked against your calendar, the slot is reserved with conflict handling, and the appointment is mirrored into GoHighLevel for the team.
Bookingbkg_5c12
- Slot
- Tuesday 10:30 AESTreserved after conflict check · GHL calendar
- CRM sync
- GoHighLevel appointment appt_8812internal and vendor IDs both persisted
- Attribution
- Sequence v3 · email template v2outcome linked to the exact wording sent
- Outcome
- Booked — awaiting attendedattended and converted tracked downstream
Your CRM stays the operator view: contacts, inbox, pipeline and calendar — without becoming the system of record.
Stage 07 · Audit & outcome
“Why was this person contacted?” answers itself.
Sequence, eligibility decision, consent evidence, step executions, provider call IDs and webhooks are all in the record — as data, not anecdotes.
AuditEvent trailprs_8f2a91
- 09:38
- Prospect created — Apollo icp_au_ops
- 09:52
- Eligibility: email eligible cev_3092
- 09:52
- Qualified 78/100 · rubric v7 · confidence 0.86
- 09:53
- Enrolled in Sequence v3 · email step scheduled ike_…
- +2d
- Positive reply · pending steps cancelled
- Tue
- Voice call completed after DNCR pass · booked · synced appt_8812
When a prospect asks to be removed, suppression writes first — then pending work is cancelled, across every channel.
Control & intelligence
AI supplies judgment. Code keeps authority
“AI generates judgment and language. Code controls authority and side effects.” That boundary is the product principle — it makes the system safer to run, cheaper to debug and possible to trust.
AI works inside bounded steps
Interprets enrichment and research
Scores fit against the rubric, with evidence
Classifies reply intent and objections
Drafts and adapts sequence messages
Holds a voice conversation with a narrow goal
Deterministic code decides
Channel eligibility, consent and suppression
Workflow state — never inferred from transcripts
Scheduling, retries, branching and rate limits
Idempotency keys before every external side effect
Bookings, CRM sync and the audit trail
Guardrails on every model call
Structured outputSchema-validated; malformed outputs are rejected, not parsed hopefully.
Evidence requiredHigh-impact claims must point to the input that supports them.
Confidence thresholdLow confidence routes to manual review instead of proceeding.
No compliance authorityA model may summarize consent context; it can never overrule a gate.
Versioned & replayablePrompt version and model stored per decision — re-run from frozen input to compare.
// The contract every qualification must satisfy
type Qualification = {
fitScore: number; // 0–100
need: "yes" | "no" | "unknown";
decisionMaker: boolean | null;
timing: "now" | "soon" | "later";
evidence: string[];
confidence: number; // 0–1
recommendedAction:
"continue" | "manual_review" | "stop";
}
Most of the system is ordinary TypeScript. Agent frameworks appear only where tool-calling or language-heavy reasoning clearly earns their place.
Compliance
Compliance is a subsystem, not a disclaimer
Australian outreach rules are engineered into the flow — consent is established and evidenced before a channel fires, and eligibility is decided by code the model cannot overrule.
Consent first
A commercial message is never the vehicle for asking consent — a lawful basis exists and is evidenced before sending.
Immutable evidence
ConsentEvent stores source, scope and time. A purchased list is a contact, never consent.
Instant suppression
Unsubscribe writes suppression first, then cancels queued steps — well inside the five-working-day expectation.
Calling windows
Timezone-aware scheduling respects telemarketing hours and DNCR status before any call is placed.
Engineering rule: a language model may explain a compliance state, but it must never be able to override a deterministic not eligible decision.
Regulator-aligned engineering notes (ACMA and Do Not Call guidance, researched September 2026) — not legal advice. Final outreach basis, scripts and targeting are approved with the client’s adviser.
Eligibility decision
Gated before send
Email
Eligible
Eligible — consent evidence cev_3092 on file
Sender identity & unsubscribe configured
SMS
Held
Consent required — no evidence captured yet
Gate holds the branch until basis exists
Voice
Allowed
DNCR check completed before dialling
Within allowed window · caller ID + disclosure script set
ConsentEvent · source / scope / capturedAt / evidenceRef
SuppressionEntry · reason + source event
Operator view
Your CRM stays the operator view
GoHighLevel is where your team lives. The engine keeps the canonical record underneath, so the CRM can be replaced without losing history.
| Domain | Control plane · Postgres | GoHighLevel |
|---|---|---|
| Prospect identity | Canonical + vendor IDs | Mirrored contact |
| Consent & eligibility | Canonical + immutable evidence | Optional summary fields |
| Workflow state | Canonical state machine | Pipeline / opportunity summary |
| Messages | Normalized log of every send & reply | Operator-facing conversation |
| Appointments | Canonical booking + GHL ID persisted | Calendar appointment |
| Outcomes | Canonical booked / attended / converted | CRM reporting |
Webhook payloads stored raw for replay
Idempotent contact upserts
Scoped OAuth · encrypted credentials
Execution truth lives in the control plane
GoHighLevel mirrors it for the team
Architecture
Vendors are adapters. The control plane is the product
Postgres stores truth, the workflow runtime executes it, and vendors perform channel-specific work. Swap any box on the edge — the engine stays.
Lead sources
ApolloICP search & dataset, behind an adapter
CSV importDeduped before enrichment credits are spent
Web formsInbound enquiries with consent captured at source
Accelerate control plane
TypeScript · Postgres · durable workflows
Prospect stateidentity, contact points, status
Eligibility & consentevidence, suppression, windows
Qualificationrubric, score, evidence, confidence
Sequences & auditexecutions, retries, every event
Channels
Emaildedicated sending infra · SPF / DKIM / DMARC
SMSTwilio · consent-gated
VoiceRetell · eligibility-gated · stop tool built in
Operator CRM
GoHighLevelcontacts, inbox, pipeline, calendar
Sync contractidempotent upserts · both IDs persisted
Replaceablebusiness logic lives above the CRM
AI models — routed via OpenRouter for bounded tasks only: qualification, reply intent, drafting, summarization. Model choice is a setting, not an architecture.
structured outputs · versioned promptsAuditEvent · prs_8f2a91append-only
12 Feb 09:38
System
Prospect createdApollo search icp_au_ops · dedupe passed
prs_8f2a91
12 Feb 09:52
Policy
Email eligibleconsent evidence cev_3092 · legal basis recorded
elg_0312
12 Feb 09:52
Model
Qualified 78/100rubric v7 · confidence 0.86 · replayable
qlf_7741
12 Feb 09:53
System
Enrolled · Sequence v3email step scheduled with idempotency key
ike_9d02
14 Feb 16:20
System
Positive reply receivedpending steps cancelled · intent classified
msg_b41c
18 Feb 10:04
Policy
Voice allowedDNCR pass · window open · disclosure script set
rtl_55e1
18 Feb 10:30
System
Booked · synced to GHLinternal + vendor IDs persisted · webhook verified
appt_8812
Audit & outcomes
Every answer from the record
Why was this person called? The system answers with data: the sequence, the eligibility decision, the consent evidence, the step executions, the provider call ID and the resulting webhook.
The data model behind it — 17 entities, audit-ready
Prospect
ContactPoint
DataClaim
ConsentEvent
SuppressionEntry
Qualification
Sequence
SequenceEnrollment
StepExecution
Message
VoiceCall
Booking
ExperimentAssignment
Outcome
WebhookEvent
AuditEvent
Optimization
“Self-improving” means measured
Memory is storage. Improvement needs attribution, comparison and controlled change — so every wording, rubric and sequence ships as a version.
Versioned at enrollment
A prospect entering an experiment is pinned to an exact sequence, template and prompt version.
Attributed downstream
Outcomes are tracked past the reply — booked, attended, converted — not just opens and clicks.
Human-approved rollout
AI summarizes performance patterns; people approve changes. At 8–20 bookings a month, small differences aren’t proof — so changes ship as versions, with rollback.
| Object | Versioned | Measured outcomes |
|---|---|---|
| Email template | v2 | Delivered · reply · positive reply · opt-out · booked |
| SMS copy | v1 | Delivery · response · opt-out · booked |
| Voice prompt | v3 | Connect · qualified · handoff · booked · opt-out |
| Qualification rubric | v7 | Override rate · booked · attended · converted |
| Sequence | v3 | Time-to-book · booked rate · cost per booked call |
Interaction history preserved in full
Outcomes attributed to the exact wording sent
The pilot
A 14-day pilot, defined honestly
One ICP, one primary sequence, one GoHighLevel sub-account, constrained volumes — and a complete audit trail from day one. Go-live means a controlled production pilot, not a promise of magic.
BuildDays 1–9 · schema, adapters, compliance engine, sequences, voice, booking
HardenDays 10–12 · offline QA, webhook replay, idempotency and policy tests, controlled ramp
Go liveDay 13 · production pilot segment and operator training
HandoverDay 14 · runbook, credentials, dashboards, rollback steps
Targets are pilot objectives — not results we claim. Measured unit economics replace estimates as the pilot runs.
PILOT TARGET
8–20
qualified calls per month, from a defined ICP
PILOT TARGET
10–15 hrs
of admin coordination returned weekly
PILOT TARGET
0
compliance incidents — gates run before every send
PILOT TARGET
100%
of engaged leads visible in one CRM
FAQ
Frequently asked questions
Straight answers about how the engine runs, what it owns, and what it will never do.
What exactly is Accelerate?
An outbound lead-qualification and call-booking control plane. It owns prospect state, compliance eligibility, sequence execution, bookings, CRM sync and audit history — while Apollo, OpenRouter, GoHighLevel, Retell and Twilio act as replaceable integrations.
Is this an autonomous bot that contacts whoever it wants?
No. AI classifies, scores, drafts and converses inside bounded steps. Deterministic code decides what may be sent, when, and to whom — eligibility, suppression, retries and scheduling are code and data, never model judgment.
How is consent handled under Australian rules?
Consent is established and evidenced before a channel fires, stored as an immutable ConsentEvent with source, scope and time. A purchased list is a contact, never consent. Unsubscribe writes suppression first, then cancels queued steps.
Can the AI override a compliance decision?
No. A model may explain a compliance state, but it cannot override a deterministic “not eligible” decision. That gate is code, evaluated before every send and every call.
Where does our data actually live?
In the control plane’s own Postgres: prospect records, consent evidence, every message, call, booking and outcome. GoHighLevel mirrors it for your team. If you leave GHL later, the engine keeps its own history.
Which vendors does it connect to?
Apollo for sourcing, OpenRouter for model routing, GoHighLevel as the operator CRM, Retell for voice and Twilio for SMS. Each sits behind an adapter with its own provider IDs, so any of them can be swapped later.
What does the 14-day pilot actually include?
A constrained production pilot: one ICP, one primary sequence, one GoHighLevel sub-account, limited volumes and a complete audit trail — built in nine days, QA’d offline, ramped carefully, then handed over with a runbook.
Are the pilot numbers guaranteed results?
No — they are pilot objectives, not claims. Targets like 8–20 qualified calls a month are what the pilot aims for; measured unit economics replace them as real data comes in.
14-day constrained pilot
Put your outbound on the record
See the engine, the compliance gates and the audit trail — then decide whether a pilot fits your team.
One ICP · one sequence · one CRM sub-account
Complete audit trail from day one
Runbook & rollback at handover