Decision engines
Rules and model scoring combined behind one interface, with versioning, simulation against historical cases before release, and a full record of every decision made.
Rules + ML · VersionedCrux automates work that should be automated, using the simplest mechanism that will hold — and says plainly when a process needs fixing, or a rule, rather than a model.
Automation projects rarely fail because the technology could not do it. They fail for reasons that were visible before anyone wrote code, and that a two-week assessment would have surfaced.
The corollary is uncomfortable but useful: the assessment sometimes concludes that the process should be redesigned or retired rather than automated. That answer costs a fraction of the alternative.
Each rung costs more to build, more to govern and more to explain than the one below it. The engineering discipline is choosing the lowest rung that solves the problem — not the highest one the budget allows.
Explicit logic a person wrote down. Eligibility thresholds, VAT calculation, routing by value. Boring, transparent, and correct every time. If the logic is knowable, stop here — a model adds cost and removes certainty.
Software driving a user interface because no API exists. Genuinely useful against legacy systems, and genuinely fragile: a screen layout change breaks it silently. Treat as a bridge to an integration, not a destination.
Models that learn a pattern nobody can write down: which category, which priority, which risk band. Requires labelled examples and ongoing monitoring for drift, but handles variation that rules cannot.
Reading unstructured content — contracts, invoices, correspondence — in Arabic and English, extracting structure. Confidence thresholds route uncertain extractions to a person rather than guessing.
Multi-step work across systems where the next step depends on the last. Powerful and correspondingly risky, requiring authority limits and approval gates. See our agentic AI work.
Most production automation we deliver combines rungs one and three, with four where documents are involved. Rung five is a deliberate decision, not a default.
Score every candidate on these before committing. A process scoring badly on reversibility should not be first, however attractive the volume looks.
| Criterion | Strong candidate | Weak candidate |
|---|---|---|
| Volume | Thousands of cases monthly, so a small per-case saving compounds | Dozens of cases, where the build cost never returns |
| Variability | Cases look broadly alike, with variation in data not in logic | Every case is negotiated and genuinely different |
| Exception rate | Exceptions are a clear minority and easy to identify | Exceptions are common and only recognisable by an expert |
| Reversibility | A wrong outcome is caught and corrected cheaply | A wrong outcome reaches a customer or a regulator irreversibly |
| Data availability | The inputs needed already exist in a reachable system | Key inputs live in someone's inbox or in an unscanned file |
The most common good first candidate in Saudi enterprises is document-heavy back-office work: invoice and receipt handling, supplier onboarding paperwork, claims intake, or application triage. High volume, moderate variability, and a wrong outcome that a person catches at the next step.
Selling AI means calling everything AI. But a model where a rule belongs adds cost, removes determinism, and creates an explainability problem you did not previously have. Almost every working system uses both.
Once software decides something about a person — an application, a claim, a limit — the organisation owns that decision. SDAIA's governance expectations and PDPL both bear on this, and so does the customer standing in a branch asking why.
That requirement shapes the architecture, which is the practical reason a rules layer usually survives alongside models: rules explain themselves, and models require deliberate work to do the same.
Rules and model scoring combined behind one interface, with versioning, simulation against historical cases before release, and a full record of every decision made.
Rules + ML · VersionedClassification and extraction from Arabic and English documents, with confidence thresholds routing uncertain cases to review rather than accepting them silently.
Arabic OCR · ExtractionMulti-step processes with conditional routing, retry and compensation logic, and durable state — built on Airflow, Temporal or equivalent rather than a bespoke scheduler.
Airflow · TemporalThe queue, the interface and the escalation path for exceptions. Usually the least glamorous component and the one that determines whether people trust the system.
Review queue · EscalationAutomated validation against ZATCA e-invoicing requirements, internal policy and sector rules, flagging issues before they reach a person or a regulator.
ZATCA · Policy rulesConnection to SAP, Oracle, Dynamics and Salesforce, and to Saudi platforms including Nafath for identity, ZATCA for invoicing, and sector systems where access permits.
SAP · Nafath · ZATCAThe shadow-running period is not caution for its own sake. It is the only way to find out whether the automation agrees with your people before it replaces them on a decision.
Document how the work is actually done, measure volume and exception rate, and score candidates against the five criteria. Ends with a ranked shortlist and any recommendation to fix rather than automate.
Choosing the lowest rung that solves it, defining thresholds, review paths and what "correct" means, agreed with the process owner and whoever carries the risk.
Decision logic, extraction or models, orchestration, review interface and audit logging, integrated with the systems that hold the data.
Running alongside the existing process without acting, comparing its decisions to your team's. Disagreements are the most valuable output of the entire engagement.
Acting, with review on everything consequential, until override rates stabilise and the exception patterns are understood.
Additional processes on the same framework, which is considerably faster once orchestration, review and monitoring already exist.
A first process typically reaches supervised production in 8 to 12 weeks. Subsequent processes on the same framework move considerably faster.
Rules when the logic is knowable and stable — eligibility thresholds, tax calculation, routing by value. Models when the input is unstructured or the pattern is real but nobody can write it down. Most working automation combines both, with rules governing anything that must be deterministic and auditable.
Usually because the process was never documented, so the automation encodes the existing confusion; because the exception rate was underestimated and exceptions now cost more than the original process; or because nobody owned it after launch and it decayed quietly while everyone assumed it worked.
Score candidates on volume, variability, exception rate, reversibility of a wrong outcome, and whether the data needed is actually available. High volume with low variability and reversible outcomes is the right first target. High-value irreversible decisions come later, if at all, and always with human approval.
That any automated decision can be reconstructed afterwards: which inputs, which rules, which model version, which score, which threshold. For regulated sectors and decisions affecting individuals, a reason code the customer can understand — in Arabic where the customer is Arabic-speaking — is part of the requirement, not a nicety.
Yes, with more attention than English. Arabic OCR quality varies with scan quality and font, handwriting remains difficult, and mixed Arabic-Latin content is common in Saudi commercial paperwork. Confidence thresholds and routing to human review matter more here than in an English-only pipeline.
A single well-chosen process typically reaches supervised production in 8 to 12 weeks, including shadow running. Extending across a portfolio is faster once the framework, monitoring and review workflows exist.
A two-week assessment documents how the work is actually done, measures volume and exception rates, and returns a ranked shortlist — including any process we would advise fixing or retiring instead of automating.