Agent Product Page · AI Outreach Agent
HERALD — AI Outreach Agent
PERSONALISED OUTREACH.
AT REAL SCALE.
HERALD turns an approved prospect list into conversations. It writes from your approved messaging, sequences across channels, and gets out of the way the moment a human should take over.
HERALD value proposition
HERALD runs approved multi-channel outreach at scale, classifies every reply, and hands genuine interest to a human before it can be mishandled.
PROBLEMS IT SOLVES
Follow-up that stops after the first email
Personalisation that dies as soon as volume increases
Opt-outs handled inconsistently across channels
Sending domains burned by unpaced volume
Positive replies sitting unread for days
COMPLIANCE AND REPUTATION CONTROLS
Outbound AI without controls is spam at scale, and the damage lands on your domain and your brand before anyone notices. HERALD is built the other way round: every send has to pass a stack of reputation, consent and content checks before it leaves, and any check can stop it.
Reputation & deliverability
- SENDER REPUTATION SCORE
- A rolling score per sending identity built from delivery, bounce, complaint, opt-out and engagement rates. Falling scores throttle sending automatically before volume damage occurs.
- DOMAIN HEALTH MONITORING
- SPF, DKIM and DMARC alignment, blocklist checks and bounce-category tracking on every sending domain, with alerts to the owning human when anything degrades.
- SEND-FREQUENCY CONTROLS
- Per-domain, per-mailbox, per-contact and per-company caps, plus warm-up ramps for new senders and cooldown windows so one prospect is never hit from multiple sequences at once.
- AUTOMATIC THROTTLING AND PAUSE
- Bounce or complaint rates crossing your thresholds slow or halt the affected sequence and sender rather than continuing and burning the domain.
Consent & lawful basis
- LAWFUL-BASIS TRACKING
- Every contact record carries the basis it is being contacted under, the source of that basis and the date it was captured. No basis recorded, no send.
- JURISDICTION RULES
- Channel and consent rules applied per jurisdiction — SMS and marketing email only go where the recorded basis and local rules allow it.
- AUTOMATIC SUPPRESSION
- Opt-outs, complaints, hard bounces, competitor and do-not-contact entries suppress instantly and permanently, across every channel and every sequence, not just the one that triggered it.
- SUPPRESSION PRECEDENCE
- Suppression is checked immediately before each send, not only at list build — a record suppressed mid-sequence stops on the spot.
Message safety
- AI MESSAGE RISK DETECTION
- Each generated message is screened before sending for pressure tactics, misleading urgency, regulated-advice language, unapproved pricing or terms, and anything that reads as spam.
- HALLUCINATION CHECKING
- Every factual statement in a message must trace to an approved template variable or a recorded signal with a source. Unverifiable content is stripped or the message is held for review — HERALD never invents a statistic, result or reference.
- BRAND VOICE ENFORCEMENT
- Messages are checked against your approved tone, terminology, banned-phrase list and claims policy, and rejected back for regeneration when they drift.
- TEMPLATE VERSIONING
- The exact approved template version and every variable used are recorded on the sent message, so any message can be reconstructed and explained later.
Every message passes seven checks before it sends
01
PERSONALISATION CHECK
The message must carry real, record-backed personalisation — no empty merge fields, no generic filler dressed up as research. A message that would read as mail-merge is rejected before it sends.
02
FACTUAL ACCURACY CHECK
Every factual statement must trace to an approved template variable or a recorded signal with a source and date. Anything unverifiable is stripped or the message is held.
03
BRAND VOICE CHECK
Tone, terminology, banned phrases and claims policy are checked against your approved voice profile; drift sends the message back for regeneration.
04
COMPLIANCE CHECK
Lawful basis, channel permission for the jurisdiction, required identification and opt-out wording, and regulated-language screening — all confirmed for this specific recipient.
05
SUPPRESSION CHECK
Re-checked immediately before the send, not just at list build: opt-outs, complaints, hard bounces, do-not-contact and competitor lists across every channel.
06
FREQUENCY CHECK
Per-contact, per-company, per-mailbox and per-domain caps, cooldown windows and cross-sequence collision detection, so nobody is contacted twice at once.
07
RISK SCORE
A combined score across the six checks plus model confidence. Below your threshold the message is queued for a human instead of sent — the default is to hold, never to guess.
Sending stops automatically when
- Bounce rate rising above your threshold for a sender, sequence or domain
- Spam complaints increasing beyond the configured rate
- Domain reputation or authentication health deteriorating
- A recipient asking not to be contacted, in any channel or any wording
- Compliance confidence falling — a jurisdiction, consent or basis question it cannot resolve
- AI confidence falling below the threshold on generation or reply classification
What it optimises for, in order
REPUTATION › ENGAGEMENT › QUALIFIED CONVERSATIONS › MEETINGS › VOLUME
HERALD is tuned in that order, deliberately. Reputation always outranks output: if protecting your domain means sending less this week, it sends less. Volume is the last thing it optimises for and never a target on its own.
Hard stops and human gates
- Human approval for sensitive and regulated industries — per-message review mode can be required for legal, insurance, financial and health-adjacent messaging
- First send from any new sequence, template or domain is human-approved before volume is allowed
- Any message failing a risk, hallucination or brand-voice check is queued for a person, never auto-sent
- Complaint or regulatory language in a reply escalates to a human immediately and freezes the sequence
- Deliverability thresholds, sending caps and suppression rules are set by you, and HERALD cannot raise them itself
These controls reduce risk; they are not a legal opinion or a guarantee of compliance. You remain the data controller for your own outreach, and your lawful basis, retention and jurisdiction decisions are yours to set. Xiilio holds no regulatory approval or certification beyond what is stated on the security page.
INPUTS, LOGIC, OUTPUTS
Inputs
- Approved prospect records from NOVA
- Your approved templates, sequences and messaging rules
- Buyer signals and CRM fields permitted for personalisation
- Suppression and do-not-contact lists
- Inbound replies, bounces and out-of-office responses
Processing Logic
- Select the sequence and channel that fits the prospect
- Compose from an approved template using permitted variables only
- Check suppression, consent and sending caps before every send
- Pace sending inside warm-up and deliverability limits
- Classify each inbound reply: positive, negative, OOO, bounce, opt-out — with sentiment and intent scored separately
- Qualify meeting-ready replies against your criteria before a diary slot is offered
- Stop the sequence and route positive replies to the owning human
Outputs
- Sent messages with the exact template version and variables used
- Reply classifications written to the CRM record
- Suppression list updates on every opt-out
- Owner notifications with the full thread attached
- Channel and sequence performance data for CORTEX
HERALD specification
Connects To
- Google Workspace / Microsoft 365 sending domains
- HubSpot, Salesforce or Pipedrive
- LinkedIn outreach tooling
- SMS and WhatsApp Business gateways (where consent exists and the channel is permitted)
- Slack or email for reply alerts
- NOVA (input) and ARIA (inbound handover)
What Triggers It
- An approved prospect list handed over by NOVA
- A form fill or content download on your site
- A CRM stage change, e.g. deal marked 'nurture'
- A reply, bounce or opt-out arriving
- A scheduled follow-up falling due
Decisions It Makes
- Which sequence and channel fits this prospect
- What to personalise, from approved variables and signals only
- When to send, and when to hold back for deliverability
- Whether a reply is positive, negative, an OOO or an opt-out
- When to stop and hand the conversation to a person
Acts Autonomously On
- Send approved sequence steps within caps
- Pause or slow sending to protect deliverability
- Stop a sequence on reply, bounce or opt-out
- Suppress a contact permanently
- Notify the owning human of a positive reply
GUARDRAILS, IN WRITING
An autonomous agent is only useful if its limits are explicit. Here is what HERALD will never do, and where a human always signs off.
What It Cannot Do
- Send to anyone who has opted out or is on a suppression list
- Write claims, statistics or promises outside approved messaging
- Negotiate terms, discounts or scope over email
- Sign contracts, agree discounts or commit your business to anything
- Invent facts, case outcomes, guarantees or performance claims
- Give legal, medical or regulated financial advice
- Act outside the systems and data scopes you explicitly grant it
Human Approval Points
- Every sequence, template and messaging variant is approved before it sends
- New sending domains and warm-up plans are signed off
- Positive replies always go to a human — HERALD never closes
- Optional per-message approval mode for regulated messaging
WHEN THINGS GO WRONG
Failure Handling
- Every run is idempotent — a retry can never duplicate a record or a send
- Transient failures retry with backoff; persistent failures stop the run rather than guess
- A failed step never silently continues — the workflow halts in a known state
- Failures raise an alert to the owning human with the run ID and the exact failing step
- Errors, retries and halts are written to the same audit trail as successful actions
- If deliverability signals degrade, sending throttles automatically before volume damage occurs
Escalation To Humans
- Every positive reply goes to a human — HERALD never negotiates or closes
- Complaints and regulatory language escalate immediately
- Replies it cannot confidently classify go to a person, not a guess
- Requests for pricing or terms are handed over untouched
EXAMPLE OUTPUT
Sequence event log — a representative example of what HERALD writes back into your systems.
sequence : 'Litigation intake' · step 2 of 5
channel : email · sending domain warmed, within daily cap
personalised : {first_name}, {firm_type}, {signal: 2 open vacancies}
template : approved v4 — no claims, no pricing
reply : positive — 'can you send times for next week?'
action : sequence stopped · owner notified in Slack
suppression : n/a
audit : logged 14:02 UTC · reviewable in the client portalEXAMPLE DASHBOARD
HERALD — outreach dashboard. Every figure below is a simulated demo value, not a reported client result.
Messages sent
3,180
last 30 days · illustrative
Delivered
98.2%
within warm-up caps
Replies
214
all classified
Positive → human
46
sequence stopped automatically
Opt-outs
19
suppressed instantly
Qualified conversations
31
met your criteria · illustrative
Meetings booked
12
from qualified conversations
Complaint rate
0.02%
below your configured ceiling
Domain health
Healthy
SPF/DKIM/DMARC aligned · not blocklisted
Median handover time
4 min
reply to owner notification
KPIs HERALD IS HELD TO
Deliverability rate
Messages accepted and delivered vs sent
Reply rate
Replies per hundred contacted
Positive reply rate
Replies classified positive and routed to a human
Opt-out rate
Unsubscribes per hundred contacted — watched as a quality signal
Qualified conversations
Replies that met your qualification criteria, not just replies received
Meetings booked
Qualified conversations that became a diary slot
Pipeline and revenue attributed
Value of opportunities sourced through outreach, read from your CRM
Complaint rate
Spam complaints per hundred delivered — a hard reputation ceiling, not a soft target
Domain health
Authentication alignment, blocklist status and reputation trend per sending domain
Time to human handover
Minutes between a positive reply and owner notification
Technical Architecture
HOW HERALD CONNECTS
HERALD reads from your existing systems, applies its own governed processing, and writes results back. Nothing runs outside these lanes.
Upstream · sources
- NOVA approved prospect queue
- CRM records and permitted fields
- Approved template library
- Inbound mail, LinkedIn and SMS webhooks
HERALD · processing
- Sequence orchestrator
- Template + personalisation renderer (approved variables only)
- Consent and suppression gate
- Deliverability pacer
- Reply classifier
Downstream · destinations
- Sending domains (Google Workspace / Microsoft 365)
- LinkedIn and SMS channels
- CRM activity timeline
- Slack / email owner alerts
- CORTEX performance feed
- The consent and suppression gate sits in front of every channel — no send path bypasses it
- Positive-reply routing is a hard stop, not a scored preference
SECURITY & AUDIT LOGS
Security
- Data encrypted in transit (TLS 1.2+) and at rest
- Least-privilege access — the agent only sees the records its job needs
- Row-level security on every customer record in the Xiilio platform
- UK/EU data handling with GDPR rights (access, export, deletion) honoured
- No customer data used to train third-party foundation models
Logged On Every Action
- Timestamp (UTC) and agent identity
- Trigger that started the run
- Every system read and write, with the record ID touched
- The decision taken and the rule or score behind it
- Human approvals, overrides and rejections
- Template version and every personalisation value used in a sent message
Full detail on encryption, data handling and your GDPR rights is on the Trust & Security page.
How an insurance broker would use HERALD
A commercial broker has 600 lapsed renewals sitting in their CRM. HERALD works them with an approved three-step sequence, personalised only from data the broker already holds. Opt-outs are honoured instantly, negative replies close themselves, and the two or three genuinely interested people each week land in the account manager's inbox with the full thread attached.
This scenario is written to show how HERALD is deployed. It is not a named client, a delivered engagement or a reported result.
Live Demo
SEE HERALD RUNNING
The homepage system builder shows HERALD's outreach stage. Metrics there are simulated demo values.
See HERALD inside a live pipelineHERALD FAQs
HERALD — QUESTIONS WE GET ASKED
- Does HERALD write its own emails?
- It composes from templates and messaging you have approved, filling approved personalisation variables from data already held. It cannot invent claims, statistics or promises, and every template version used is recorded in the audit log.
- How are opt-outs handled?
- Immediately and permanently. An opt-out stops all sequences for that contact and adds them to the suppression list, which is checked before every send across every channel.
- What happens when someone replies?
- HERALD classifies the reply. Positive replies stop the sequence and go straight to the owning human with the full thread. HERALD does not negotiate, quote or close.
- Will this damage our sending domain?
- Sending is paced within warm-up and daily caps, bounces are handled and suppressed, and deliverability is monitored. New sending domains go through an approved warm-up plan before volume increases.