Agentic operations
Back-office work, done by agents you can audit.
We build multi-step AI agents that take on procurement requests, accounts-payable exceptions, HR service requests and IT tickets — reading each request in Arabic or English, checking it against policy, acting in your systems and handing over to a person when the stakes rise. Each agent holds only the permissions its task needs, every action is logged, and no change reaches production without passing an evaluation first.
The problem
Why back-office queues stay full
- 01Requests arrive by e-mail, forms and chat in two languages, and someone has to read each one before anything moves.
- 02Accounts-payable exceptions — price mismatches, missing goods receipts, duplicate invoices — wait days for someone to investigate.
- 03Rule-based automation breaks on anything unexpected, so teams stop trusting it and go back to doing the work by hand.
- 04Leadership will not let software act in the ERP without clear limits, approvals and a record of what it did.
How it is built
How the system is built
An orchestrator breaks each request into steps and hands them to specialised agents. Agents act only through a tool gateway that enforces permissions, approval thresholds and logging — so the controls live in the platform, not in the prompt.
Layer 1
People and channels
- Shared mailboxes and web forms
- Microsoft Teams or Slack
- Approvers’ inbox and mobile
- Operations console for supervisors
Layer 2
Agents
- Orchestrator that plans each request
- Procurement and accounts-payable agents
- HR and IT service-desk agents
- Policy and document retrieval
Layer 3
Tool gateway and controls
- Scoped tool permissions per agent
- Approval thresholds by amount and risk
- Idempotent actions with safe retries
- Kill switch and full action logs
Layer 4
Infrastructure
- Private models on-premises or in-Kingdom cloud
- Connectors to ERP, ITSM and HR systems
- Encrypted state and audit store
- Monitoring and evaluation pipeline
Capabilities
Agentic operations
Orchestrator with specialised agents
A planner breaks each request into steps and routes them to agents built for one domain each. Narrow agents are easier to test, permission and replace than one agent that tries to do everything.
Least-privilege tool access
Each agent gets an explicit list of the tools and fields it may use — read a purchase order, draft a reply, open a ticket — and nothing else. Credentials are scoped per agent and never shown to the model.
Approval thresholds
You set the lines: an agent may update a ticket on its own, but releasing a payment above an amount you choose, or changing a supplier’s bank details, waits for a named approver.
Safe execution
Every action carries an idempotency key, so a retry never posts the same invoice twice. Failed steps retry with back-off, then escalate to a person with the full context instead of failing silently.
Action logs and a kill switch
Each step records its input, the tool called, the result and the reason. An operator can pause one agent, one workflow or all of them at once, and work in progress is parked safely.
Evaluation before every change
A test set of real, anonymised cases runs against every prompt, model or tool change before release. If accuracy or policy adherence falls below the agreed bar, the change does not ship.
How it works
How a request moves
- Request arrives
- Agent plans the steps
- Policy and limits checked
- Person approves if needed
- Systems updated and logged
Connects to
Systems it works with
- SAP S/4HANA
- Oracle Fusion and E-Business Suite
- Microsoft Dynamics 365
- ServiceNow
- Jira Service Management
- Workday
- SAP SuccessFactors
- Microsoft 365 and Exchange
- Microsoft Teams
- In-house systems through REST APIs
Where it runs
On-premises
Models and data run on hardware in your own facilities. Nothing leaves your network unless you allow it.
Private cloud in the Kingdom
Dedicated, isolated capacity in a Saudi data-centre region of a provider you select — the elasticity of cloud without a shared public AI service.
Hybrid
Sensitive workloads stay on-premises; less sensitive ones use private cloud capacity. You set the policy for what goes where.
What we measure
What we measure
Agreed with you before we build and reported against once the agents are live — targets we track together, not results we promise.
01
Straight-through rate
Share of requests completed without a person touching them, tracked by request type.
02
Turnaround time
Time from a request arriving to it being closed, compared with the manual baseline.
03
Escalation quality
Share of escalations a reviewer agrees needed a person, and share of agent actions later reversed.
04
Cost per request
Model, infrastructure and review effort per completed request.
Indicative delivery path
Indicative delivery path
- 01
2–3 weeks
Process selection and baseline
We map one or two request types end to end, measure today’s volumes and turnaround, and agree permissions, approval thresholds and measures with process owners.
- 02
6–8 weeks
Build and evaluate
We build the orchestrator, agents and tool gateway in your environment, and test them against a set of real, anonymised cases until they meet the agreed bar.
- 03
4–6 weeks
Supervised pilot
Agents work on live requests with low thresholds, so people approve most actions. We raise autonomy only where the logs show it has been earned.
- 04
Ongoing
Scale and operate
We add request types, tune thresholds and report against the measures, with change control for every model, prompt or tool update.
Durations are indicative and depend on scope, data and integrations.
Questions
Agentic operations
What stops an agent from doing something it should not?
Three layers. An agent can only call the tools on its permission list, the tool gateway routes any action above your amount or risk thresholds to a person, and every action is logged with its reason. Operators can pause any agent immediately with the kill switch.
Which processes make a good first candidate?
High-volume requests that follow written rules but need someone to read the input: purchase requisitions, three-way-match exceptions, leave and letter requests, access and password tickets. We avoid starting where the rules are unwritten or disputed — automating those only makes the disagreement faster.
Do the agents replace our ERP workflows?
No. Agents work through your existing systems and their approval chains, using the same APIs and service accounts your integrations already use. Your ERP stays the system of record; the agents prepare, check and route work into it.
Which models do you use, and where do they run?
We choose models per task — often an open-weight model such as Llama, Qwen or Mistral, or an Arabic-first model such as ALLaM — hosted on your premises or in private cloud capacity in the Kingdom. The orchestration does not depend on one model, so you can change models later; every change passes the same evaluation set first.
Other solutions
All solutionsPrivate AI platform
Open-weight and commercial models served inside your boundary, behind one governed gateway.
View solutionKnowledge assistant
Cited answers from your documents and systems, in Arabic or English — limited to what each user may open.
View solutionDocument intelligence
Invoices, contracts and applications read, checked against your records and posted — people review the rest.
View solutionCustomer service agent
Arabic and English service on web chat, WhatsApp and voice — approved answers, lookups, hand-over to people.
View solution
Have a process that should run itself?
Tell us about it. We will help you work out whether AI is the right answer, what it would take, and what it should return.

