Workflow Automation with a Deterministic Layer
Rules decide. AI assists. People approve.
Automation for work that is routine but not simple: fixed rules handle what must always happen the same way, and AI handles the reading and judging in between, inside limits set before anything runs.
Signal
Saudi businesses name integration with existing systems among their main obstacles to AI (42%, SAP and YouGov survey, 2025). A rules layer is where that integration is made safe.
Client problem
Fewer people, the same handovers.
- A person still reads each document or request and decides where it goes.
- Work moves between systems by copy-and-paste, and errors are found downstream.
- Exceptions take most of the effort while the straightforward majority waits in the same queue.
Approach
Define the boundary before granting the capability.
01 / MAP
Map the human task
Document how the work is done now, including the judgement calls and the cases people escalate.
02 / RULE
Write the deterministic layer
Permitted actions, approval points, limits and escalation conditions, applied before any model acts.
03 / RUN
Shadow, then release
The workflow runs alongside your team first. Authority is handed over gradually, with an audit trail and a way to stop it.
Deliverables
What you get.
- A process map separating the automatable path from the cases that stay with a person.
- A working workflow with its rules, permissions, approval gates and escalation paths.
- An audit log, monitoring and a documented procedure to pause or roll back.
Technology, grouped by purpose
Chosen for the job.
Foundation
Python services and REST integrations against your existing systems, with queues and scheduled retries.
Intelligence
Tool-using agents on LangGraph or our own Fatima Agent Harness, with models from the Anthropic or OpenAI APIs.
Governance
Least-privilege credentials, human approval steps, structured action logs and spend or volume limits.
FAQs
Questions buyers ask.
What is a deterministic layer?
A set of fixed rules that always gives the same result for the same input: approval limits, routing, required fields, policy checks. It runs before and after the AI step, so the model can only act inside what the rules allow.
What stops an agent doing something it should not?
Permissions are granted narrowly at the integration level, high-impact steps need explicit human approval, and every action is logged. The agent can only do what its credentials allow.
How do we know it works before we depend on it?
It runs in shadow mode first: it processes real work and proposes actions while your team still decides, so results are compared before any authority is handed over.
Strategy call
Bound one workflow properly.
A 30-minute conversation, no obligation. We will tell you plainly if this is not the right next step.
Related: AI Agents for Email and WhatsApp · Data Engineering and Analytics