Automation multiplies whatever a process already does. Applied to a process nobody has examined, it multiplies the mistakes faster and at greater volume.
Merjora designs and delivers AI automation for workflows where the evidence supports it: high-volume steps blocked by unstructured input, with a measurable baseline and a human review point where the consequence of an error justifies one. The scope decision comes from the assessment — some processes are better simplified than automated, and some are best left alone.
Deciding what to automate is the hard part
Exception rate is the most useful signal available. A process where most cases follow one path is a strong candidate. A process where a large share of cases deviate is telling you the design does not match reality, and automating it hard-codes the mismatch into software somebody must then maintain.
The second signal is where the bottleneck sits. If the blocking step is comprehension — reading an email, a document, a mixed-format form — AI is the right tool. If the step is deterministic, conventional automation is cheaper and more predictable.
The scope decision, made explicit
Matching the intervention to the process
Condition
Recommendation
Disputed rules, unjustifiable steps
Simplify before automating
Stable path, high volume, structured data
Conventional automation
Unstructured input blocking the step
AI with a defined review point
Branching path across several systems
Agentic workflow, with approval gates
Low volume, high variability
Leave it alone and say so
Human-in-the-loop by design
The review point is designed before the model is chosen. A step that produces a draft for approval can tolerate imperfect accuracy; a step that writes to a system of record cannot. Value thresholds, confidence thresholds and escalation paths are part of the specification, not a later hardening exercise.
Draft-and-approve for correspondence and recommendations
Confidence thresholds routing uncertain cases to a person
Value thresholds requiring approval above an agreed amount
Full audit trail of what the system did and who confirmed it
A defined fallback when the automation is unavailable
Where AI automation earns its keep
Document and invoice processing across inconsistent layouts
Inbound classification and routing for shared inboxes
Data transfer between systems that were never integrated
Drafting standard correspondence for human approval
Summarising long case histories before a person acts
Quality and exception checks that people currently perform by sampling
Measuring the return honestly
The running cost of an automation includes reviewer time, model usage, monitoring and the person who notices when it breaks. Reviewer time usually exceeds inference cost, and leaving it out is the most common reason a delivered automation fails to show the modelled saving.
What you receive
Workflow specification with inputs, outputs, rules and permissions
Review and escalation design with thresholds
Working automation integrated with the systems involved
Monitoring, audit trail and failure fallback
Baseline and re-measurement plan
Business impact
Handling time reduced on the specific steps that carried the cost
Fewer errors reaching downstream processes
Capacity released without headcount change
A documented decision trail for audit and compliance
Where this is the wrong fit
Processes with high exception rates that have not been redesigned
Workflows where nobody will own the output after go-live
Automating a control that exists for a regulatory reason without a review point
Frequently asked questions
What business processes should be automated with AI?
High-volume steps where the bottleneck is reading or interpreting unstructured input, with a measurable baseline and an acceptable review design. Deterministic, structured work is usually cheaper to automate without AI.
When should we not automate a process?
When the rules are disputed, the exception rate is high, the volume is low, or the annual cost of the process is smaller than the cost of changing it.
Do AI agents replace traditional automation?
No. Agents handle branching, multi-step work that rules engines handle badly, at the cost of more places to fail. Deterministic paths remain better served by conventional automation.
How is the return measured?
Against the baseline captured before the change, with running costs — including reviewer time — subtracted, and a re-measurement date set at go-live.