The MARGEN Method
A way to avoid poorly chosen AI projects.
Most AI projects start too late, on the wrong process: when someone has already decided to automate. Astor starts earlier. It identifies where margin is being lost, redesigns the workflow, and only then applies AI where the economic case justifies it.
Note on RPA
Unlike traditional RPA, which breaks when an exception appears, Astor designs systems that understand context and adapt to edge cases. Operational AI doesn't assume the process is stable: it assumes the process has real friction that a well-designed system can absorb.
Map
Key question
How does the business actually operate, not how it's assumed to operate?
Deliverable
Operational map of the current process, with real timings, handoffs, dependencies and friction points documented.
Risk it avoids
Automating a process no one fully understands. The biggest mistake is made by whoever thinks they already know how it works.
Isolate
Key question
Where is margin, time, control or speed being lost?
Deliverable
Map of operational leakage, ranked by estimated economic impact, intervention effort and execution risk.
Risk it avoids
Prioritizing interesting but economically irrelevant cases. AI on a low-volume process is a technical experiment, not a business case.
Redesign
Key question
How should the process work before applying AI?
Deliverable
Target workflow with clear ownership, explicit decision criteria and a quantified business case.
Risk it avoids
Putting AI on a broken workflow. AI amplifies what it finds; if the process is poorly designed, it automates it poorly and at greater speed.
Generate
Key question
What minimum pilot validates impact without overbuilding?
Deliverable
Limited functional system with measured baseline, acceptance KPIs and documentary traceability for every output.
Risk it avoids
Building too much before learning. Long pilots are expensive and tend to die before proving anything.
Scale
Key question
How is the pilot turned into stable operations rather than an orphaned project?
Deliverable
System in production, operational documentation, team training, governance and review cadence.
Risk it avoids
Letting the pilot die once it stops being novel. Handover without structure is the phase where most AI projects fail.
Normalize
Key question
How is the system kept alive, measured and improving?
Deliverable
Operational retainer with live KPIs, model adjustments, incident review and quarterly improvement plan.
Risk it avoids
Letting the system degrade in silence. Models age, processes change, integrations break. Without cadence, initial ROI evaporates.
Margin first. Process second. Then — and only then — AI.
This order isn't rhetorical. It's the only sequence that protects the client's investment. If a phase doesn't meet its exit criterion, the next one doesn't start. If the economic case doesn't hold at Isolate, there's no pilot. If the pilot doesn't reach its baseline at Generate, there's no system. The discipline of the method is what justifies the work.