Skip to content

System Architecture

How the Xiilio AI Revenue System Works

ONE SYSTEM. SHOWN OPEN.

The operating model the agents follow, the governance around them, how the system joins your existing stack, and which parts are ours versus a named third party.

System Overview

HOW AGENTS PASS WORK

01

QUALIFY

Agents capture need, timeline and budget in conversation, and score every lead on live signals — so only real opportunities move forward.

02

ROUTE

Qualified conversations are handed to the right next step automatically: a booked Free Strategy Audit, an outreach sequence, or a task for the owning human.

03

ESCALATE

Agents work inside tight guardrails — they can't make claims, quote prices or commit your business without human approval. Positive replies and edge cases go straight to a person.

04

REPORT

Every action is audit-logged and traceable, with daily briefings by email or Slack, so you always see what an agent did and why.

What Connects Them

UNIFIED EVENT ARCHITECTURE

Every connected system publishes its events onto one shared bus. Any event can trigger an AI workflow, and any workflow can call any agent. That is what makes the agents a system rather than six separate tools — a lead created in your CRM, a call answered by ARIA, a card charged in Stripe and a conversion fired by an ad platform all arrive in the same normalised shape, with the same identity resolution, permissions and audit trail behind them.

OAuth and credentials

Each system is authorised with scoped OAuth or least-privilege API credentials, stored encrypted, rotatable and revocable one connection at a time.

Permissions

Per system and per object: read-only, write with approval, or write automatically. Defaults are read-only until you widen them.

Webhooks

Inbound webhooks are signature-verified and de-duplicated; outbound webhooks let your own systems subscribe to Xiilio events.

Retry logic

Failed calls retry with exponential backoff and jitter, respect vendor rate limits, and stop at a defined ceiling rather than hammering an API.

Error handling

Anything that cannot be delivered goes to a dead-letter queue, is surfaced in the command centre, and is replayable once the cause is fixed. Nothing is silently dropped.

Audit logs

Every event, workflow run, agent action and configuration change is logged with timestamp, actor, inputs and before/after values, and is exportable.

Role-based access

Who can view, edit, approve, execute or connect is set by role. Approval rights are separate from build rights.

Data mapping

Field-level mapping between systems with type coercion, transformation rules and one authoritative source per field, changeable without engineering work.

Event history

A searchable timeline of every event and every workflow it triggered, filterable by system, record, agent or outcome.

Workflow versioning

Workflows are versioned, diffable and roll-backable. Each run records the version it executed, so past behaviour stays explainable.

Human approval

Any step can be marked approval-required, with a named approver, a timeout and an escalation path if nobody responds.

Event coverage depends on what each vendor's API and webhook surface exposes and on the permissions you grant. Workflows described here are configuration examples, not guaranteed outcomes.

Governance Layer

GUARDRAILS, APPROVALS AND THE RECORD

These are the same controls published in full on the Responsible AI centre and the security disclosure, including the limits of each one.

Human oversight

Autonomy without a human decision point is how automated systems cause commercial damage at speed.

  • — Approval gates on consequential actions: first-touch outbound to a new segment, budget shifts above your configured cap, pricing statements, and anything customer-facing in a regulated context.
  • — Human takeover on any live conversation, at any point, with full context handed over.
  • — Xiilio staff review conversation and outreach samples for quality and safety; reviews are logged.
  • — Every recommendation the platform makes is presented with evidence, confidence and risk so the approver can judge it rather than rubber-stamp it.
  • — You choose the mode per capability: recommend only, approve each action, or auto-execute inside limits you set.

Oversight is only as good as the reviewer. Where you enable auto-execute, actions happen without a person seeing them first — that is your decision to make, and the caps are yours to set.

Agent permissions

An agent with broad system access can do broad damage, quickly, without malice.

  • — Least privilege by default: each agent is granted named scopes on named systems, and nothing else.
  • — Write actions are separated from read actions; an agent can be granted read-only access to a system it reports on.
  • — Spend, volume and frequency caps are configured per agent — an agent cannot exceed its budget or send limit even if instructed to.
  • — Connections are authorised by you via OAuth and can be revoked by you at any time, from your side, without involving us.
  • — Permission changes are recorded in the audit log with who changed what, and when.

Permissions govern what the platform can do inside connected systems. They cannot restrict what those systems then do with a legitimately authorised action.

Audit logs

If you cannot reconstruct why an automated system did something, you cannot defend it to a regulator, a client, or your own board.

  • — Append-only logging of agent actions, approvals, rejections, overrides, pauses and permission changes, with actor and timestamp.
  • — Each significant action records the reasoning summary and the evidence it relied on, so 'why did it do that?' has an answer.
  • — Administrative actions in the Xiilio admin area are logged separately and cannot be edited from the interface.
  • — Logs are exportable for your own compliance records and for procurement review.

Logs capture what the platform did and on what basis. They are not a certified immutable ledger, and retention follows the periods published on the Security page.

Escalation

The value of an AI system is decided by what it does at the edge of its competence.

  • — Defined triggers: low confidence, a complaint, a legal or regulatory question, a pricing or contractual request, repeated misunderstanding, an explicit request for a human, or detected distress.
  • — Escalation carries the full transcript, summary, intent, sentiment and suggested next action to the human picking it up.
  • — Named routing per client — enquiry type to owner — rather than a shared inbox nobody reads.
  • — Escalation targets and unhandled escalations are tracked and reported, not silently dropped.
  • — Incident escalation for platform issues follows the process published on the Security page, including ICO notification where the threshold is met.

Escalation hands the matter to your team or ours. Response times depend on the humans available; we do not offer a 24/7 answering guarantee.

Recorded on every action

  • Timestamp, agent and task
  • Input and data sources
  • Decision summary and evidence
  • Tools used and action taken
  • Policy checks
  • Confidence and risk class
  • Human approval
  • Outcome and error status
  • Cost, model and version
What each signal captures, and where the record stops →

Integration Layer

NEXUS AND YOUR EXISTING STACK

NEXUS is the layer every other agent runs on. NOVA, ARIA, HERALD, SENTINEL and CORTEX do not each hold their own private copy of your data — they read and write through NEXUS, so one record, one identity and one audit trail run across the whole stack. Availability of any given system depends on that vendor's API and your plan with them.

What the backbone guarantees

One authoritative source per field

Conflicts are resolved by a rule you set, not by whichever system wrote last.

Identity resolution

The same person or company across CRM, billing, ads and email is matched to one record instead of five.

Idempotent writes

A retried sync does not create duplicate contacts, tasks or deals.

Backpressure and rate limits

Vendor API limits are respected and queued through, rather than failing a sync mid-run.

Replayable failures

A failed sync is queued, surfaced and re-runnable once the cause is fixed. Nothing is silently dropped.

Full audit trail

Every read, write, mapping change and agent action is logged with timestamp, actor and before/after values.

Scoped permissions

Each connection uses least-privilege credentials, revocable independently, with no shared master key across systems.

Your data stays yours

Connections can be disconnected at any time and the systems keep working — NEXUS coordinates your tools, it does not hold them hostage.

How a system gets connected

  1. 01MapWe list the systems you run, what each is authoritative for, and where records currently disagree.
  2. 02ConnectAuthorise each system with scoped credentials. Nothing is ripped out or migrated.
  3. 03ReconcileDuplicate and conflicting records are surfaced for a decision before any agent writes to them.
  4. 04GovernSet write permissions per system: read-only, write with approval, or write automatically.
  5. 05RunAgents operate through the backbone, every action logged, with a dry-run period before writes go live.

Connector availability depends on each vendor's API and your own plan with them. Where a native connector is not available, NEXUS uses documented APIs or webhooks — we will tell you which before you buy, not after.

Full NEXUS connector list →

What's Ours, What Isn't

PROPRIETARY VS THIRD-PARTY

We would rather under-claim here than have a buyer discover a gap later. So, plainly:

What we built

  • The agent orchestration logic — how the six agents qualify, route, escalate and report work between each other.
  • The governance and guardrail layer — permission boundaries, risk classes, human approvals, pause controls and the append-only activity record.
  • The vertical-specific workflow design — the playbooks, qualification criteria and conversation logic for litigation law, life insurance, insurance broking and merchant services.
  • The CRM integration layer — the connectors and data mapping between the agents and your existing stack.

What we use

  • The underlying language, speech-to-text and text-to-speech models — Google Gemini, Deepgram and ElevenLabs — licensed infrastructure accessed on each provider's published API terms. We hold no bespoke agreement with any of them.
  • Managed cloud hosting, database and storage.
  • Payments (Stripe) and transactional email delivery (Resend).
  • The voice telephony layer (Vapi) that carries ARIA's calls.

Xiilio does not train or fine-tune its own foundation models, and does not offer self-hosted or private model deployment. Every inference call goes to a third-party provider.

Architecture

Xiilio is a browser-based application with a managed Postgres backend and serverless functions. There is no Xiilio-operated hardware.

  • Front endIn placeStatic React application served from the hosting provider's CDN. No customer data is stored in the bundle.
  • Application backendIn placeManaged Postgres (Supabase, via Lovable Cloud) with row-level security on customer-facing tables, plus serverless edge functions for booking, chat, payments and webhooks.
  • Self-hosted infrastructureNot heldNone. Xiilio operates no owned servers, colocation or on-premise equipment.
  • Environment separationIn placeSeparate test and production databases with separate credentials. Production data is not copied into test.

Model providers

The named models behind the agents, so you can assess them yourself.

  • Text and reasoningIn placeGoogle Gemini models, accessed through the platform AI gateway. Model versions change as providers update them; the current version is available on request.
  • Speech-to-textIn placeDeepgram, via Vapi.
  • Text-to-speechIn placeElevenLabs, via Vapi.
  • Self-hosted or private modelsNot heldNot offered. Every model call goes to a third-party provider listed above.

Subprocessors

Every third party that can process customer data on our behalf. If a vendor is not on this list, it does not have your data.

  • Lovable (hosting) & Supabase (database, auth, storage, functions)In placeApplication hosting, database, authentication and file storage.
  • Google (Gemini models, via the platform AI gateway)In placeLanguage-model inference for chat and agent reasoning.
  • Vapi, with Deepgram and ElevenLabsIn placeVoice agent orchestration, speech-to-text and text-to-speech for ARIA voice calls.
  • StripeIn placePayment processing and subscription billing. PCI DSS Level 1 service provider.
  • ResendIn placeTransactional and notification email delivery.
  • GoHighLevelIn placeCRM routing for enquiries where the client uses it. Not used for clients on other CRMs.
  • Google Analytics 4In placeWebsite analytics, loaded only after consent, with Consent Mode v2 and IP anonymisation.
  • Change noticeDocumentedWe will tell existing clients by email before adding a subprocessor that handles their data.
Read the full security disclosure →

GO DEEPER

WANT THIS WALKED THROUGH?

We'll take a technical buyer through the architecture against your own stack, connector by connector.

See Xiilio in ActionThe Platform
  • ICO Registered

We use a single first-party cookie to remember your currency and consent choice. Page-view analytics are sent to our own servers — no third-party trackers, no profiling. See our Privacy Policy.