IDP-01LAYER 01 - INTERFACE AND INTAKESHARED ACROSS 12 FUNCTIONS

Intelligent Document Processing

Extract, validate and process data from documents, in any format and in either language.

The leak it closes

Every process in the business that begins with a person reading a document and retyping what it says into a system. Invoices, contracts, bills of quantities, claim forms, bank statements, shipping sets, certificates and site reports.

Retyping is slow, but the real cost is that nobody validates what was typed. Duplicate payments, price variances, missing exclusions and expired certificates all survive because the document was treated as true rather than tested against the record.

The four steps it runs

Ingests from every channel, including email, WhatsApp, scanner, portal and shared drive, and handles scans, photographs, PDFs, spreadsheets and email bodies in Arabic and English.
Classifies the document against the process it belongs to and extracts at line level rather than at header level.
Validates what it extracted against the purchase order, the contract, the policy or the rule set, and tests for duplicates, tolerance breaches and missing items.
Posts clean records to the system of record and routes every exception to a named person with the source page, the field in question and the reasoning attached.HUMAN IN THE LOOP

Where the same capability shows up

Accounts payable invoice capture and three-way matching
Construction and MEP bills of quantities priced against a cost library
Letter of credit document sets tested against UCP 600
Clinical coding and claim documentation sufficiency
Mortgage bank statements and payslips normalised into an affordability view
Statutory and inspection certificates tracked to expiry

Primary measure

Straight-through rate: documents processed with no human touch

Secondary measure

Field-level extraction accuracy and cost per document

Systems it reads and writes

SAPOracleMicrosoft Dynamics 365SalesforceDocument managementEmail and WhatsApp intakesZATCA and FTA e-invoicingVision and OCR models

Discuss this capability Back to the stack

KNW-02LAYER 02 - KNOWLEDGE AND DATASHARED ACROSS 12 FUNCTIONS

Enterprise Knowledge & Evidence Agents

Deliver trusted answers from approved enterprise knowledge, with the source cited every time.

The leak it closes

Answers that sound right and cannot be traced. Policy, precedent, procedure and prior work sitting in systems nobody searches because searching them takes longer than asking a colleague.

The organization pays repeatedly for knowledge it already owns, and its most valuable asset walks out of the door with every departure. In a regulated business, an uncited answer is also an audit finding waiting to happen.

The four steps it runs

Indexes policies, contracts, precedents, engagement history and operating procedures with the permission and conflict model applied at ingestion, not at display time.
Interprets the question in the context of who is asking, what matter they are on, and what they are entitled to see.
Answers with the source cited at clause or paragraph level, and states explicitly where the retrieved precedent differs from the situation asked about.
Refuses and logs any request that would cross a confidentiality or conflict wall, rather than answering it with a caveat.HUMAN IN THE LOOP

Where the same capability shows up

Firm knowledge and precedent for legal and advisory teams
HR policy desk answering from the approved handbook
Contract playbooks used during review and negotiation
Standard operating procedures served at the point of work on a shop floor
Audit and inspection evidence packs assembled with references

Primary measure

Share of answers carrying a verifiable citation

Secondary measure

Reuse rate of existing work, and time to a defensible answer

Systems it reads and writes

SharePoint and Microsoft 365Google WorkspaceiManage and NetDocumentsContract repositoriesHRIS handbooksGoverned vector index

Discuss this capability Back to the stack

WFA-03LAYER 03 - REASONING AND ORCHESTRATIONSHARED ACROSS 12 FUNCTIONS

Custom AI Workflow Agents

Build AI agents around your workflows and your business rules, not around a vendor template.

The leak it closes

The gap between how a process is documented and how it actually runs. Approval matrices, tolerance thresholds, exception routes and the unwritten rules that decide what really happens when something does not match.

Generic automation breaks on the exception, and the exception is where most of the cost lives. Some tool handles the easy eighty percent and the team still carries the difficult twenty, which was always the expensive part.

The four steps it runs

Encodes the workflow as it genuinely runs, including the delegation of authority, the tolerance bands, the exception paths and the escalation ladder.
Executes the steps end to end against live data instead of stopping to prompt a person at every stage.
Produces the actual work product: the journal entry, the variation claim, the renewal quote, the dispatch, the reconciliation.
Acts inside its autonomy level and stops at the decision the business has stated a person must own, with the full working shown.HUMAN IN THE LOOP

Where the same capability shows up

Month-end close and reconciliation
Change order and variation assembly
Contract renewal with an evidenced upgrade case
Intake, triage and dispatch against an SLA clock
Onboarding chains across IT, facilities, payroll and visa
Customer deduction and claim validation

Primary measure

Touchless completion rate for the workflow

Secondary measure

Cycle time end to end, and override rate on agent output

Systems it reads and writes

ERP and CRMWorkflow and approval toolingDelegation of authority matrixScheduling and field systemsNotification channels

Discuss this capability Back to the stack

MAW-04LAYER 03 - REASONING AND ORCHESTRATIONSHARED ACROSS 12 FUNCTIONS

Multi-Agent Digital Workforce

Coordinate specialised AI agents across a process that crosses several functions.

The leak it closes

The gaps between departments, where work stalls because no single function owns the critical path and each one believes the delay belongs to someone else.

The largest value in most organizations is not trapped inside a function; it is trapped between them, which is exactly why no functional tool has ever recovered it.

The four steps it runs

Splits the process into agents with separated duties, so that no single agent both prepares the work and approves it.
Hands work between agents over a defined agent-to-agent contract, carrying the state, the evidence and the reasoning across the boundary.
Runs a supervising agent that tracks the critical path across all of them and identifies which step is genuinely blocking rather than simply reporting a status list.
Escalates cross-function conflicts to the named human owner. Agents never negotiate a commercial or clinical decision between themselves.HUMAN IN THE LOOP

Where the same capability shows up

Order to cash from quotation through collection
Procure to pay from requisition through payment
Integrated business planning across demand, supply, finance and commercial
Property transaction progression across brokers, banks and authorities
Tender through delivery through handover in contracting

Primary measure

End-to-end cycle time for the whole process, not the fastest step

Secondary measure

Handoff failure rate and separation-of-duties compliance

Systems it reads and writes

Orchestration layerAgent-to-agent protocolERP, CRM and planning systemsTask and message queuesShared state store

Discuss this capability Back to the stack

VCA-05LAYER 01 - INTERFACE AND INTAKESHARED ACROSS 12 FUNCTIONS

Voice & Conversational AI

Build intelligent voice, chat and messaging experiences that resolve rather than collect.

The leak it closes

The delay between someone reaching out and someone capable answering. Most enquiries and requests go cold in the hour after they arrive, and nights and weekends are usually dark.

Speed of first meaningful response decides conversion, satisfaction and, in service businesses, whether the SLA clock was even started. Marketing spend that generates a lead nobody answers is spend that generated a lead for a competitor.

The four steps it runs

Handles voice, WhatsApp, web chat, email and in-app messaging with the full prior history of that customer, patient or employee available in the conversation.
Qualifies, resolves or books directly against live systems rather than collecting a message for someone else to read tomorrow.
Holds the right register for the relationship, including tone, language and what may and may not be said commercially.
Hands over with the diagnosis already made, and escalates anything clinical, legal, financial or emotionally sensitive to a person immediately.HUMAN IN THE LOOP

Where the same capability shows up

Instant lead qualification and viewing or trial booking
Patient scheduling, confirmation and waitlist backfill
Tenant and facility requests taken on WhatsApp
Multi-channel service desks with full contact history
Collections outreach in the right register for each customer

Primary measure

Containment rate and first contact resolution

Secondary measure

Median response time, and coverage outside working hours

Systems it reads and writes

WhatsApp BusinessTelephony and IVRWeb and in-app chatHubSpot and SalesforceBooking and ticketing systems

Discuss this capability Back to the stack

MLR-06LAYER 01 - INTERFACE AND INTAKESHARED ACROSS 12 FUNCTIONS

Multilingual & Regional AI

Create AI experiences for local languages and markets, with the regional rules encoded rather than translated.

The leak it closes

The assumption that a global product plus a translation layer is a regional product. It is not, and in the GCC the difference is visible in the first sentence.

Arabic that reads as translated damages credibility immediately. Regional rules handled as a localization afterthought produce output that is fluent and non-compliant, which is worse than output that is obviously wrong.

The four steps it runs

Operates natively in Arabic and English, and adds the working languages of the field force where the workforce actually needs them.
Encodes the regional rule set itself rather than the global default: ZATCA and FTA e-invoicing, Nitaqat and Emiratisation, RERA and Trakhees, SFDA and MEWA, DIFC and ADGM alongside onshore civil law.
Adapts tone, formality and timing to the market, including the working week, prayer times and local business etiquette.
Escalates anything where the Arabic and the English version would carry a materially different legal or commercial meaning.HUMAN IN THE LOOP

Where the same capability shows up

Bilingual contracts reviewed against a bilingual playbook
Customer service at genuine Arabic and English parity
Regulatory filings and submissions in the required local format
HR policy and employee communication across a multilingual workforce
Field instructions in the language the technician actually works in

Primary measure

Bilingual quality at native reviewer assessment

Secondary measure

Regional rule coverage, and parity of outcome between languages

Systems it reads and writes

Locale-aware modelsRegulator and authority portalsTerminology base and translation memoryLocal document and numbering standards

Discuss this capability Back to the stack

INT-07LAYER 04 - ACTION AND SYSTEMSSHARED ACROSS 12 FUNCTIONS

ERP, CRM & Enterprise Integration

Connect AI agents with ERP, CRM and enterprise systems under the same controls as any other user.

The leak it closes

The gap between an agent that knows the answer and a business record that changes. Without this layer, agentic AI is a very articulate advisor with no hands.

Pilots that never leave the demo stage almost always fail here. The reasoning was fine; nothing ever wrote back, so nothing ever changed and the value was never realized.

The four steps it runs

Connects to the systems of record through supported interfaces, including SAP, Oracle, Microsoft Dynamics 365, NetSuite, Odoo, Sage and Salesforce, plus the surrounding operational stack.
Reads the live record rather than an extract, so the agent reasons on what the business currently believes rather than on last night's snapshot.
Writes back through a service account scoped to exactly what the process requires and nothing beyond it.
Logs every read and write, and refuses any posting that would breach a control, a period lock or an approval threshold.HUMAN IN THE LOOP

Where the same capability shows up

Everywhere. This is the layer that converts an answer into work.
Posting validated invoices and journals
Updating pipeline, pricing and inventory records
Raising work orders, requisitions and purchase orders
Writing asset and maintenance data back to the CMMS

Primary measure

Write accuracy and control breaches, with a target of zero

Secondary measure

Integration availability and mean time to recover a broken connector

Systems it reads and writes

SAPOracleMicrosoft Dynamics 365NetSuiteOdooSageSalesforceHubSpotJiraMicrosoft FabricAPIs and middleware

Discuss this capability Back to the stack

DIP-08LAYER 03 - REASONING AND ORCHESTRATIONSHARED ACROSS 12 FUNCTIONS

Data Intelligence & Predictive Agents

Detect patterns, anomalies and trends across business data, then say what to do about them.

The leak it closes

The lag between something going wrong in the business and it becoming visible in a report. In most organizations that lag is a month, and by then the cycle is closed.

A variance discovered after the period is an explanation. The same variance discovered in the period is a decision. Every reporting-led organization is paying for the difference without seeing the invoice.

The four steps it runs

Builds the baseline from history, plan, contract terms, breed or engineering standard and comparable peers, rather than from a fixed threshold somebody set once.
Detects deviation and separates genuine signal from normal variation, so the alert volume stays credible.
Attributes the movement to a cause and quantifies the cost of not acting, rather than reporting that a number moved.
Puts the prioritised action in front of the accountable owner, then tracks whether it was taken and what it actually changed.HUMAN IN THE LOOP

Where the same capability shows up

Demand forecasting and gap-to-plan attribution
Energy drift and equipment degradation inside alarm noise
Livestock and agricultural performance against breed standard
Margin and price leakage across customers and channels
Credit and collections risk scored by predicted behaviour
Downtime and OEE loss attributed to the real cause

Primary measure

Detection lead time, from event to a named owner being told

Secondary measure

Forecast accuracy, and share of alerts that result in an action

Systems it reads and writes

Microsoft FabricSnowflakeBigQueryPower BIProcess historians and telemetryERP and CRMExternal market, commodity and weather data

Discuss this capability Back to the stack

LEG-09LAYER 04 - ACTION AND SYSTEMSSHARED ACROSS 12 FUNCTIONS

Legacy-System Computer-Use Agents

Automate approved legacy systems that have no modern API, without replacing them.

The leak it closes

The excuse that a critical process cannot be automated because the system behind it is twenty years old and is not being replaced this year.

The oldest systems usually sit closest to the money. Leaving them out of scope means leaving the highest-value processes manual while the modern edges of the business get automated.

The four steps it runs

Operates the application through its own interface, under a named and permissioned account, inside an approved and isolated environment.
Follows the documented procedure step by step, keeping a screen-level record of exactly what it did and when.
Verifies the result inside the system afterwards rather than assuming that the click worked.
Stops on any unexpected screen and escalates with the screenshot, instead of continuing blind through a state it does not recognise.HUMAN IN THE LOOP

Where the same capability shows up

Legacy ERP and green-screen terminal applications
Government, regulator and authority portals with no API
Bank, payer and insurer portals
Older CMMS, PMS and practice management systems
Supplier and customer portals that only accept manual entry

Primary measure

Task success rate and unattended completion

Secondary measure

Exception rate, and hours returned to the team

Systems it reads and writes

Virtualised desktop environmentComputer use modelsCredential vault and scoped accountsSession recording and audit store

Discuss this capability Back to the stack

ADE-10LAYER 02 - KNOWLEDGE AND DATASHARED ACROSS 12 FUNCTIONS

Agentic Data Engineering

Automate data pipelines, quality checks and issue resolution, so the spine agents read from stays trustworthy.

The leak it closes

The data backlog. Broken loads, schema drift, duplicate masters and silent quality failures that a small team fixes manually, always after someone has already acted on a wrong number.

Most agentic failures are not reasoning failures. They are data failures wearing a reasoning costume. An agent reading from an unreliable spine produces confident work that is subtly wrong, which is the fastest way to lose trust permanently.

The four steps it runs

Builds and maintains ingestion from ERP, CRM, spreadsheets, PDFs, portals and operational telemetry into a governed model with lineage.
Tests every load for schema drift, completeness, duplication and business-rule violation, before anything downstream consumes it.
Diagnoses the failure and drafts the fix, whether that is a mapping change, a rule change or a chase to the supplier or distributor who sent bad data.
Applies approved fixes and escalates to the data owner anything that would change a figure already reported.HUMAN IN THE LOOP

Where the same capability shows up

Microsoft Fabric lakehouse and warehouse for finance and operations
Distributor and franchise submissions normalised to a master hierarchy
Master data across entities, customers, suppliers and assets
Consolidation and group reporting inputs
Portfolio and investor reporting data

Primary measure

Pipeline reliability and data quality pass rate

Secondary measure

Time to resolve a break, and reported figures restated

Systems it reads and writes

Microsoft FabricSnowflakeBigQueryTransformation and orchestration toolingData catalogue and lineageQuality rule library

Discuss this capability Back to the stack

GOV-11AROUND ALL FOUR - GOVERNANCESHARED ACROSS 12 FUNCTIONS

Governance, Guardrails & Observability

Prove what the agent did, why it did it, and exactly where it was stopped.

The leak it closes

The gap between a governance policy written in a document and a control that actually applies at the moment an agent acts.

Without this, every deployment is a negotiation with risk, audit and the regulator that starts from zero. With it, the conversation is about evidence rather than assurances.

The four steps it runs

Sets an autonomy level per step: assist, recommend, act-with-approval, act.
Applies technical, ethical and operational guardrails at the point of action, including data residency, permission scope, prohibited actions and value thresholds.
Records the reasoning, the sources consulted and the action taken for every run, in a form an auditor will accept as evidence.
Blocks and escalates anything outside the level, and makes tightening a level a configuration change rather than a rebuild.HUMAN IN THE LOOP

Where the same capability shows up

Every agent, in every industry. Required before any regulated deployment.
Financial controls and period locks
Clinical and patient data handling
Data residency in the UAE and KSA, your tenant, your keys
Model routing and token cost control

Primary measure

Share of agent decisions with a complete logged rationale

Secondary measure

Guardrail breaches, and time to tighten an autonomy level

Systems it reads and writes

Policy and guardrail engineImmutable audit storeIdentity and access managementModel routing layerCost and token telemetry

Discuss this capability Back to the stack

EVL-12AROUND ALL FOUR - IMPROVEMENTSHARED ACROSS 12 FUNCTIONS

Agent Memory, Evaluation & Continuous Improvement

Agents that improve because they are measured, not because the model underneath them got newer.

The leak it closes

The assumption that an agent is finished at go-live. Performance drifts, the business changes, and nobody notices until someone stops trusting the output.

A reviewer who approves everything without reading is a bigger risk than an agent that is occasionally wrong. If nothing measures edit rate, that failure is invisible until it causes something.

The four steps it runs

Captures the outcome of every run, including what a human changed before accepting it and why they changed it.
Scores acceptance rate, edit rate and override rate together, because acceptance alone can rise for the wrong reason.
Feeds confirmed corrections into the domain memory, so the same mistake is not rediscovered every cycle.
Flags drift and routes any material behaviour change back through human review before it reaches production.HUMAN IN THE LOOP

Where the same capability shows up

Every capability, after go-live
Asset and fault memory built up across maintenance cycles
Lender, payer and regulator policy memory that changes monthly
Firm precedent and prior-engagement learning
Reason-code and root-cause libraries that get sharper each period

Primary measure

Edit rate trend over time, alongside acceptance rate

Secondary measure

Override rate, and regression against golden test sets

Systems it reads and writes

Evaluation harness and golden test setsDomain memory storeRelease and approval pipelineDrift monitoring

Discuss this capability Back to the stack

Capability	Finance & accounting	Procurement, supply chain & operations	Sales & customer service	HR & legal	Risk, compliance & data
IDP-01Intelligent Document Processing	Invoices, statements, remittances	Delivery notes, BOQs, LC document sets	Order forms and signed job cards	Contracts, CVs, visa and licence documents	Certificates, filings and audit evidence
KNW-02Enterprise Knowledge & Evidence Agents	Accounting policy and treatment	Engineering standards and SOPs	Pricing rules and product knowledge	Handbook, playbooks and precedent	Regulatory requirement libraries
WFA-03Custom AI Workflow Agents	Close, reconciliation and approvals	Scheduling, dispatch and PPM recovery	Quote, renewal and follow-up chains	Onboarding and case workflows	Control testing and gap closure
MAW-04Multi-Agent Digital Workforce	Procure to pay end to end	Integrated business planning	Order to cash and transaction progression	Recruit to onboard across functions	Filing cycles across entities
VCA-05Voice & Conversational AI	Collections and supplier queries	Field and site intake on WhatsApp	Lead qualification and service desks	Employee helpdesk and candidate contact	Evidence chasing from record owners
MLR-06Multilingual & Regional AI	Arabic statutory reporting	Bilingual field instructions	Arabic and English customer parity	Bilingual policy and contracts	ZATCA, Nitaqat, RERA and SFDA rules
INT-07ERP, CRM & Enterprise Integration	Posting to ERP ledgers	Work orders and requisitions	Pipeline, pricing and inventory writes	HRIS and ATS updates	Regulator portal submissions
DIP-08Data Intelligence & Predictive Agents	Variance, margin and cash forecasting	Downtime, energy and yield anomalies	Churn, demand and price movement	Attrition and capacity signals	Risk scoring and alert triage
LEG-09Legacy-System Computer-Use Agents	Legacy ledger and banking screens	Older CMMS and plant systems	Portals with no integration	Government labour portals	Regulator and authority portals
ADE-10Agentic Data Engineering	Consolidation and group reporting	Telemetry and production data	Distributor and channel submissions	People and payroll data	Lineage for reported figures
GOV-11Governance, Guardrails & Observability	Period locks and value thresholds	Safety-critical action blocks	Commercial commitment limits	Sensitive data handling	Residency and audit evidence
EVL-12Agent Memory, Evaluation & Continuous Improvement	Controller edit rate on drafts	Technician correction of diagnoses	Seller override of recommendations	Reviewer changes to screening	Drift against golden test sets

Build the capability once. Point it at any domain.

A capability is not a feature. It is a repeatable way of doing work.

Four layers of machinery. One immune system around all of it.

Intelligent Document Processing

Enterprise Knowledge & Evidence Agents

Custom AI Workflow Agents

Multi-Agent Digital Workforce

Voice & Conversational AI

Multilingual & Regional AI

ERP, CRM & Enterprise Integration

Data Intelligence & Predictive Agents

Legacy-System Computer-Use Agents

Agentic Data Engineering

Governance, Guardrails & Observability

Agent Memory, Evaluation & Continuous Improvement

Read this grid across, not down.

Four things change. The rest is already built.

Pick one process. We will tell you which capabilities it needs.