NEW The NOVA engine now understands Saudi dialects with higher accuracy
Platform · The AI Engine

An engine born for Arabic, deciding within your boundaries

The NOVA engine reads the message as written: dialect, shorthand and all: extracts the intent and entities, and decides the next step within permissions you drew. Anything beyond the boundary escalates to human approval, and every step lands in the audit trail.

The responsible review pipeline

No action executes without its path

Every decision the engine proposes travels the same pipeline: draft, then permission check, then human review where needed: then a recorded execution.

The same pipeline applies to every agent in every flow: no exceptions, no side doors.

Dialect & intent understanding

Reads the message as written

Most engines think in English and translate: meaning gets lost on the way. The NOVA engine starts from the Arabic itself.

Local dialect understanding

Najdi, Hijazi, Gulf: and business fus'ha, with high accuracy. Your customers and employees never have to write the way "the system understands".

Entity extraction in Arabic

Order numbers, amounts, dates, names: even from unformatted text, so a message turns directly into fields a workflow can act on.

Decisions within boundaries

Permission before intelligence

Intelligence without explicit boundaries is operational risk. In NOVA, every agent's permissions are drawn before it acts: at the action and data level.

Permissions per agent

What an agent executes on its own, and what it escalates to human approval: you define both per system and per action, not as one platform-wide setting.

Beyond the boundary, nothing runs

An action outside its permission never auto-executes: it escalates for approval with full context, and the decision: approval or rejection: is recorded in the audit trail.

Permission map: operations workspace
Illustrative matrix: what each agent can reach in each system: allowed, approval-gated, or denied
Support agentOps agentFinance agent
Order systemAllowedAllowedDenied
Customer recordsAllowedApprovalDenied
Invoices & paymentsDeniedApprovalAllowed
Knowledge baseAllowedAllowedAllowed
AllowedApprovalDenied
Knowledge grounding

Answers from your knowledge, not thin air

The engine's answers are grounded in the knowledge sources your organization connects to the workspace: and each agent's access to each collection is explicit and visible.

Sources you connect

Your policies, your guides, your systems: the engine grounds itself in what the workspace authorizes, so you always know where an answer came from.

Out of scope stays out

Whatever an agent has not been granted simply never appears in its answers: a design rule that supports data minimization and audit readiness.

The decision mechanism

How the engine decides: in four steps

From incoming message to recorded trail: the same four steps, every time.

Reads the message as written

Before any decision, the engine turns free text: dialect, shorthand and all: into structured understanding: what the sender wants, and what data the message carries.

  • Detects and normalizes the dialect before analysis
  • Extracts the intent: what does the sender actually want?
  • Captures the entities: numbers, amounts, dates, names

Picks the next step inside the boundaries

The decision is never open-ended: the engine matches the proposed action against the agent's drawn permissions: within the boundary it proceeds, beyond it it stops at a human.

  • Matches the proposed action against the agent's permissions
  • Within the boundary it proceeds; beyond it, it escalates for approval
  • The full decision context is presented to the human reviewer

Works in your systems through connectors

Execution happens inside your actual systems: the order system, customer records, email: within each integration's scoped access.

  • Executes through integrations within scoped access
  • Separate development and production environments for testing before launch
  • A failed step halts the path: no blind execution

Writes everything to the audit trail

Every understanding, decision and execution leaves a reviewable trace: you don't have to trust what happened, you can read it.

  • Who decided, what was executed, when, and why
  • Human approvals and rejections are part of the record
  • Review the trail any time: before the audit and after it
FAQ

The questions we hear in every demo

No: and that is not its purpose. The engine is designed to take over the repetitive steps: reading messages, extracting data, executing routine actions within explicit permissions. Sensitive decisions stay with your team through human approvals: so your people focus on what genuinely needs their judgment, and the organization sees everything that happened in the audit trail.

Najdi, Hijazi, and Gulf dialects, alongside business fus'ha: with high accuracy. It also extracts entities in Arabic: order numbers, amounts, dates, names: even from unformatted text, so your customers never have to write the way "the system understands".

It never executes automatically. An action beyond the drawn boundaries escalates to human approval with its full context: what was requested, why the agent proposed it, and what it would affect. The whole path: approval or rejection: is recorded in the audit trail.

From the knowledge sources and systems your organization connects to the workspace: your policies, guides, and data: within explicit access permissions for every agent. Whatever an agent has not been granted simply never appears in its answers; a design rule that supports data minimization and audit readiness.

Try the engine on your own messages.

Book a live demo where we run messages from your real work: and you watch the engine understand, decide within boundaries, and record every step.