Atsky’s own demonstration · Voice AI architecture

From a conversation
to a controlled
workflow.

Explore how voice input, configured policies and system integrations fit together in an agentic workflow. Skyla is Atsky’s own demonstration of that design, with First Notice of Loss as the example.

▶ Try FNOL Demo Live How It Works ↓

Policy-controlled workflows Data-handling controls Real-Time Voice Policy-Constrained Architecture Multi-Tenant SaaS

Agentic Runtime Fabric — SKYLA Architecture

Layer 1 — Listen

Predictive Flow Manager

STT / TTSProsodyIntent PredictionInterruptions

↓

Layer 2 — Think

Neuro-Symbolic Reasoner

Neural EmbeddingsCausal GraphsRAGWorld State

↓

Layer 3 — Act (ARF Core)

Policy-Constrained Planner

OPA / RegoSafety GovernorConfidence GatesStep Order

↓

Layer 4 — Integrate

AIS Registry + GER Audit

SAPSalesforceGuidewireAPI integration

The ARF principle: LLMs propose. Deterministic logic decides. The design checks proposed actions against configured policies before execution.

CX Journey Orchestrated by SKYLA

One caller. Every system. One session.

SKYLA acts as the intelligent orchestration layer between the caller and all enterprise systems — executing every step deterministically, in the configured order, without a human agent in the loop.

Measure latencyIntent to system action

HandoverEscalation route to define

Shared contextContext carried between steps

PolicyConfigured checks to validate

GERAudit record on every invocation

The Problem We All Recognise

We've all been stuck on hold.
It's fatal for businesses.

The problem is not that AI cannot talk.
It's that it cannot be trusted to act.

Waiting, repeated questions and disconnected systems can make a contact-centre journey frustrating. A useful automation assessment starts with the causes and their measured impact in your operation.

Voice triggers verified actions. Real-time updates across all enterprise systems. Deterministic, policy-bound workflows the AI cannot deviate from.

⏱

Handling time to investigate

Where agents move between disconnected systems, measure the time spent looking up and re-entering information while keeping the caller informed.

⚠️

Compliance steps skipped under volume

Check whether workload and handoffs affect validation, disclosures and escalation steps. Establish the actual gaps from operational evidence.

🎲

Generic AI guesses — regulations don't

Model output alone does not enforce workflow controls. Define explicit validation, access rules and escalation paths, then test them against the intended use.

The Market Gap

Listening vs. Doing

Most voice AI today

Focused on sounding human

Better speech recognition and transcription

Probabilistic guessing at intent

Human remains the execution layer

Cannot guarantee step order or field validation

Compliance is a post-hoc checkbox

Customers care if the refund is processed — not if the bot understood the word "refund."

→

SKYLA

Focused on acting correctly

Voice triggers verified, policy-bound actions

Deterministic execution — not probabilistic guessing

AI executes across systems; humans review only edge cases

Step order enforced by symbolic rules — unskippable

Compliance is architectural — PII redacted before any LLM call

Define allowed actions, refusal conditions and the path to a human, then test them.

The Core Innovation

Three algorithmic breakthroughs.
One agentic runtime.

SKYLA is not a better chatbot. It is a Neuro-Symbolic agentic system — a hybrid engine blending neural intelligence with causal logic to constrain workflow actions in real-time voice.

Layer 1

Listen

Predictive Flow Manager

A multimodal model that fuses audio cues, prosody, and semantic context to predict intent before the user finishes speaking. Handles interruptions, corrections, and non-linear disclosures natively.

Solves the interruption problem. Natural conversation for elderly users, complex accents, and emotionally charged calls like FNOL.

Layer 2

Think

Neuro-Symbolic Reasoner

A hybrid engine blending Neural Embeddings (for nuance and emotion) with Causal Graphs (for logic and facts). Maintains world state across multi-step workflows — even through interruptions and corrections.

Drastically reduces hallucinations. Understands the difference between a user "venting frustration" and "changing a database field."

Layer 3

Act

Policy-Constrained Planner

Compliance as code, not a suggestion. By combining Symbolic Policies (OPA/Rego) with LLM scoring, SKYLA creates a deterministic Safety Governor. The design checks proposed actions against configured policies before execution.

Configured policies define which actions the workflow may execute. Execution records support review of those decisions.

From Call to Resolution

One call. Every system. No rip-and-replace.

📞

Caller speaks

Natural voice — interruptions, corrections, and accents handled natively

→

🎙

Intent predicted

STT + Predictive Flow Manager resolves intent before utterance completes

→

🛡

Policy cleared

OPA evaluates the configured tenant policies before execution

→

⚙️

System executed

AIS Registry writes to CRM, claims, scheduling — schema-validated before commit

→

✅

Confirmed live

Voice + SMS confirmation to caller. GER audit record created atomically.

Measure the whole path from intent to confirmed action. API compatibility, integration work and recovery behaviour depend on the selected systems.

Where SKYLA Works

Initial focus: regulated, high-volume,
rules-driven voice processes.

Where unsafe automation fails fastest — healthcare, insurance, public services. If it works safely here, it works everywhere.

🏎

FNOL — Motor Claims

Dutch non-life insurance · Primary wedge · Live demo available

Identity & policy verification in real time — no hold required
Accident detail extraction from natural, non-linear conversation
Mandatory fraud check — cannot be skipped by the agent
Schema validation before claim-file creation
Assessor appointment booked and confirmed within the same call

See the FNOL Demo →

🏥

Healthcare Back-Office

Clinical admin · GGZ waiting list mediation · Appointment handling

Frees clinicians from manual verification and admin calls
GGZ wachtlijstbemiddeling — matching patients to available capacity
Identity verification and record updates under strict data rules
Multi-step workflows with deterministic rollback on failure

Request Healthcare Briefing

📋

High-Volume CX

Retail · Telecom · Public services · Rules-heavy processes

Orders placed, updated, cancelled, tracked — in one call
Identity verified and records updated across all systems
Appointments booked, rescheduled, or cancelled autonomously
Automation coverage measured against the selected workflow

Discuss Your Use Case

From demonstration to an engagement

Start with a workflow.
Agree the decision and evidence.

We agree the scope, deliverables and price before work begins. A demonstration helps frame the questions; your operating data establishes the business case.

Baseline

Current handling time,
cost and exceptions

Test

Bounded workflow
and success criteria

Decide

Evidence for the
next investment

Enterprise Grade

Design concerns for
enterprise voice AI.

🛡

Built-in AI Governance

PDP/PEP policy enforcement, PII redaction before any LLM call, data-class rules and egress controls. Policy decisions and execution records support technical review.

⚡

Integration planning

The integration design uses an AIS capability layer. Assess the required APIs, access, failure handling and testing before agreeing implementation scope and timing.

🌐

Works Across Your Entire Landscape

One AI agent framework that unifies contact centres, back-office systems, and global teams. Eliminates tool and integration fragmentation at the source.

🔓

No Vendor Lock-In

Works with your preferred LLM — OpenAI, Anthropic, Gemini, Mistral, on-premises Llama. Multi-cloud and hybrid deployment. Strategically flexible long-term.

⚙️

Configuration and integration

Tone, locale and workflow rules are configuration concerns. New systems and use cases may also require engineering, validation and operational changes.

🧠

True Contextual Memory

Redis-backed session memory keeps multi-step workflows consistent through interruptions and corrections. Session state is intended to preserve confirmed details across the workflow.

Why This Is Hard

Five design challenges
to test in your workflow.

A voice workflow combines conversation, system state, permitted actions and recovery. These design patterns need to be tested against the intended use.

Combinatorial Explosion

"Cancel my order" is not one action — it is a branching tree of dozens of condition checks, lookups, and outcomes. Standard AI gets lost in the math.

→ Design approach: Causal graph planning with bounded search

State Drift — The Amnesia Problem

In live conversations with interruptions, LLMs lose track of world state — forgetting what has been confirmed versus what is still pending.

→ Design approach: Neuro-symbolic world-state memory

The Compliance Gap

Model output and a configured policy are different kinds of evidence. Technical controls need to be reviewed alongside the system’s use, documentation and operating procedures.

→ Design approach: OPA/Rego Safety Governor — hard rules

Temporal Dependencies

Workflows must happen in order: Check → Book → Confirm. If step 2 fails, step 1 must roll back. LLMs cannot reason correctly in time across multi-step transactions.

→ Design approach: Deterministic step ordering with rollback

The Reality of Voice

Real humans interrupt, mumble, and speak over the bot. Emotionally charged calls like FNOL add unpredictability. Static chatbot loops break immediately under acoustic chaos.

→ Design approach: Predictive Flow Manager + semantic turn segmentation

SKYLA's full solution stackNeuro-symbolic memory · Policy-constrained planning · Causal graph reasoning · Semantic turn segmentation · Deterministic fallback strategies · Hierarchical multi-agent coordinator

Policy

configured rules
to test

Measure latency

intent-to-action
execution

Scope

integration work
to assess

Compliance by Architecture

Not a checkbox. Structural.

🇪🇺

Execution records for review

The execution-record design captures the tenant, session, policy decision and outcome for technical review.

🔒

Data minimisation in the design

PII redaction runs before any LLM call. OPA-backed policy enforcement governs field-level data access. EU data residency enforced at infrastructure level.

🏦

Evidence for operational review

Operational evidence needs to be assessed against the organisation’s sector, system risks and review requirements.

🏗

Multi-tenant isolation — hard boundary

The design separates tenant credentials, policies and data. Isolation must be tested against the deployment’s access model and integrations.

Illustrative execution record

Every ARF action produces an immutable, timestamped audit record — ready for regulatory review at any point.

tenant_idillustrative-insurer-demo

session_idfnol-20260417-8a3f

stepfraud_check

policy_decisionPERMIT

pii_redactedtrue

action_executedclaim_file.create

outcomeCONFIRMED

latency_ms142

eu_data_residencyENFORCED

From a conversationto a controlled workflow.

One caller. Every system. One session.

We've all been stuck on hold.It's fatal for businesses.

Listening vs. Doing

Three algorithmic breakthroughs.One agentic runtime.

One call. Every system. No rip-and-replace.

Initial focus: regulated, high-volume,rules-driven voice processes.

Design concerns forenterprise voice AI.

Five design challengesto test in your workflow.