Skip to content

Technology

One reference architecture behind every system we build.

AI systems that act need more than a model and an API key. They need orchestration, real context, and controls that live in the execution path.

Reference architecture

Five layers, spanned by trust.

The Trust Layer is drawn vertically on purpose: identity, policy, evidence and approval cut across every other layer rather than sitting beneath them.

Fig. 04: Reference architectureAI deployment stack
  • 01

    Experience Layer

    Role-adaptive interfaces, workflows and approval surfaces. What a person actually sees and signs.

    • Role-adaptive UI
    • Workflows
    • Approvals
    • Review queues
  • 02

    Orchestration Layer

    Agents, routing, planning and policy-aware execution. The part that decides what happens next and whether it is allowed.

    • Agents
    • Routing
    • Planning
    • Policy-aware execution
  • 03

    Intelligence Layer

    Foundation models, domain reasoning, retrieval, classifiers and evaluators, selected per task, not per vendor relationship.

    • Foundation models
    • Domain reasoning
    • Retrieval
    • Classifiers
    • Evaluators
  • 04

    Context Layer

    The organisation's real substrate: systems of record, communication and knowledge. Context is what separates a demo from a deployment.

    • CRM / ERP
    • Data stores
    • Email & calendar
    • Documents
    • Knowledge
  • 05

    Measurement Layer

    Evaluations, performance, cost, adoption and outcome feedback: the evidence that a system is worth its autonomy.

    • Evaluations
    • Performance
    • Cost
    • Adoption
    • Outcomes
Spans every layer

Trust Layer

Identity, permissions, policy engine, evidence, audit, human approval and rollback, spanning every other layer rather than sitting beneath them.

  • Identity
  • Permissions
  • Policy engine
  • Evidence
  • Audit
  • Approval
  • Rollback

Trust Layer

Identity, permissions, policy engine, evidence, audit, human approval and rollback, spanning every other layer rather than sitting beneath them.

  • Identity
  • Permissions
  • Policy engine
  • Evidence
  • Audit
  • Approval
  • Rollback

Autonomy model

Autonomy is allocated per action, not per product.

Consequence and reversibility set the level. A single system typically operates at several levels at once.

Autonomy levelsL0 → L5

L3 · Act with approval

The system executes, but only after an explicit human approval on the specific action.

Different actions within the same system can run at different autonomy levels, based on risk, confidence and reversibility.

Fig. 02 / Autonomy dialDrag to change the posture

L3

Act with approval

Governance load

52

The system executes, but only after an explicit human approval on the specific action.

The system may

Executes the specific approved action.

A human still

Approves that exact action first.

Evidence written

Approver, policy matched, result.

Governance load is an illustrative index, not a measured figure. Autonomy is set per action, so one system usually runs at several levels at once.

Conceptual model

Value rises with autonomy, and so does the governance requirement.

Conceptual illustration of the relationship we design around. It is not derived from measured data.

Fig. 02: Value vs autonomy, conceptual Operational value Required governance
0255075100L0ObserveL1RecommendL2DraftL3ApproveL4In policyL5Closed loop

More autonomy can unlock more operational value, but governance and evidence must rise with it. Autonomy is only defensible where the evidence layer is genuinely in place. Conceptual illustration of a design principle, not measured data.

Evidence

What a system writes when it acts for you.

Pick an action and read the receipt it produces. Sample artefacts, built to the shape our systems emit.

Fig. 03 / Trust receiptSample artefact, not client data

Every consequential action our systems take writes one of these as it happens, not afterwards.

RCPT-4F2A19

Renewal outreach sent

Signing
Inputs read
  • CRM account record, read 11:04 UTC
  • Last 40 days of activity, 62 events
  • Renewal date and contract value
Actor
Closos · revenue agent
Policy matched
OUTBOUND-03 · named accounts, business hours, one touch per week
Approval
Auto, inside boundary. Owner notified.
Result
1 email queued to the economic buyer, CRM task created.
Reversal
Recall window 15 minutes. Task deletable. Full input snapshot kept.

If it cannot be proved and undone, it should not be automatic.

Technical principles

Non-negotiables.

01

Model-agnostic by design

Models are components, not architecture. Task, cost, latency and risk decide which model runs, and the choice can change without rebuilding the system.

02

Human control where consequence demands it

Autonomy is allocated per action, calibrated to consequence and reversibility rather than applied as a single global setting.

03

Least-privilege tool access

An agent receives the narrowest possible access to the smallest necessary set of tools, scoped to the task in front of it.

04

Traceable actions and evidence

Every consequential action carries its inputs, the policy it matched, the approver and the result. Traceability is produced by execution, not bolted on.

05

Reversible automation where technically possible

Where an action can be undone, the reversal path is designed before the action is permitted.

06

Evaluation before and after deployment

Systems are evaluated before release and monitored after it. Regression is treated as a first-class risk.

07

Context over generic prompting

Performance in production comes from grounded organisational context and retrieval design, not from cleverer instructions.

08

Security and governance in the workflow

Controls live inside the execution path rather than in a parallel review process that runs after the fact.

The thesis

The next generation of software will not wait to be asked. It will understand, decide, act, and prove what it did.

We’re building that generation from Dubai’s DIFC, for organisations that have to answer for what their systems do.