- 01
Market
Saudi Arabia
- 02
Written for
COO, Head of Digital
- 03
Decision
Automate a back-office workflow that involves judgement
- 04
Languages
EN · AR
The split
Which half of the workflow each belongs to
A useful test: if the step would be reviewed by an auditor, it should be deterministic and logged. If it would be reviewed by a manager, a model can do it with the manager still able to see what it decided.
| Option | Typical steps | What good looks like | How failure shows up | Best for |
|---|---|---|---|---|
| Model territory | Reading an email and deciding what it is about, summarising a case, drafting a reply, classifying an exception, extracting a field. | A decision with the reasoning attached and a confidence the workflow can route on. | A wrong classification, visible in a queue and correctable by a person in seconds. | Steps a person currently reads and judges. |
| Deterministic territory | Posting a journal, issuing a refund, changing an entitlement, sending a contract, updating a customer record. | Code with tests, an audit trail, and the same result every time it runs. | An exception that stops the run, loudly, before the record changes. | Anything an auditor would ask you to reproduce. |
Method
How a workflow gets automated without becoming unauditable
Map it as it runs, not as documented
Including the workarounds. The undocumented step where someone checks a spreadsheet before approving is usually the step that matters. A step list with a decision or a transaction marked against each.
Automate one judgement step behind a review
The model proposes, a person confirms, and both are recorded. The confirmation rate over a fortnight tells you whether the review can be relaxed. Evidence for the confidence threshold rather than an opinion about it.
Keep the transaction boring
The write stays in code with tests and an audit trail. Where a model triggers it, the trigger is logged with the reasoning that produced it. A workflow an auditor can follow end to end.
Instrument the silence
The dangerous failure is a step that quietly stops firing. Alert on volume anomalies as well as errors, because a model that has started refusing everything looks healthy on an error dashboard. Failures that are noticed the same day.
A caution
Automating a broken process makes it faster, not better
A workflow with four approval steps that exist because of an incident in 2018 does not need a model reading each one; it needs three of them removed. The mapping stage frequently produces that finding, and acting on it is cheaper than any automation that could be built on top of the process as it stands.
The honest sequence is therefore: map, simplify, then automate what survives. Suppliers who skip the middle step deliver more billable work and a process that is harder to change afterwards, because the automation has encoded the steps that should have been deleted.
Our role
What an automation engagement covers
- 01
The map, including the workarounds
Produced by watching the work rather than by reading the procedure, because the two differ and the difference is where the automation would break.
- 02
The build, on both sides of the line
Model steps with reasoning and confidence, transaction steps as tested code, and an explicit boundary between them that a reviewer can see.
- 03
Monitoring that catches silence
Volume and outcome anomalies alongside errors, so a step that has stopped working is found by the system rather than by a customer.
Which approvals a process genuinely needs is your decision to make; we surface the ones that exist only by inheritance so it can be made deliberately.
Common questions
What buyers ask
Is this different from RPA?+
In where it is fragile. Screen-driven automation breaks when an interface changes and is reliable in between; model-driven steps do not break on an interface change but vary in quality on unusual input. The two are complementary, and the design question in both cases is the same: what happens when the step is wrong, and who finds out.
Can a model approve transactions if we set a limit?+
A limit is the right instrument, and the approval should still be a deterministic rule applied to the model's output rather than the model deciding. That distinction matters when somebody asks how a specific approval happened: a rule can be reproduced and shown, and a judgement cannot be reproduced identically even when it was correct.
How do we know it is still working in six months?+
By measuring the same things after launch as before it: what share of items complete without a person, what share of automated decisions get corrected, and whether volume through each step matches what it was. Most quiet failures show up as a change in distribution long before anybody logs a complaint.
In this cluster
Next step
Map one workflow end to end
One workflow mapped step by step with the model and deterministic halves separated, and the first judgement step automated behind a review.

