Methodology
The RL3 AI-First Playbook
Updated on
How we decide which work stays with your team and which moves to AI, how much autonomy it gets and which controls it carries.
We redesign the process before we automate it.
Why we start with the process
Put AI on top of a badly designed process and you automate its flaws too. That's why our engineers work inside your company, next to the people doing the work, and redesign the process before automating it. The model is called Forward Deployed Engineering.
The seven steps and their deliverables
MAP
Map how the work is done today
We sit down with the people who do the work and walk through it: where information gets stuck, who has to remember what, and how long each handoff waits.
DeliverableCurrent-state flow map
DECOMPOSE
Break it into tasks
We split each block of work into small tasks and sort them: mechanical, analytical, judgement calls, or work with people.
DeliverableClassified task inventory
REDESIGN
Redesign the process
We remove redundant steps, duplicate approvals and hand-copied data, and draw the new process with AI already in it.
DeliverableRedesigned process
DELEGATE
Split the work
We give each task an autonomy level and write down what the AI receives, which tools it may use and what it must hand back.
DeliverableSpec for each AI workflow
GUARD
Set the limits
We define the worst case, which actions need approval, and how everything stops if something goes wrong.
DeliverableSafeguard matrix and stop button
EVALUATE
Measure
We test with everyday cases and edge cases, and measure quality and cost before every change.
DeliverableTest suite
LEARN
Learn from every correction
Every correction or rejection from your team is logged, and we use it to fine-tune the AI.
DeliverableCorrection log
After LEARN, we start again at MAP.
Reference frameworks
The method is aligned with the EU AI Act, including the Article 4 AI literacy duty, with ISO/IEC 42001 for AI management systems, and with the NIST AI Risk Management Framework. In practice, every workflow has its own risk inventory, documented human oversight, an audit log and a team trained to use it.
In our projects
At Efímero Clinic, Sofía confirms date, time and practitioner with the patient before saving an appointment in Flowww, and passes any urgent symptom to staff. In Compliance Brain, a true screening match stops the case the same day, and every onboarding needs the MLRO's second approval.
Where to start
Every project starts with a free assessment. We talk through one process that costs you time or money and tell you whether AI fits, at what level, and where we'd start.
Five levels of autonomy
L0
A person does it
100% human work. AI stays out of it.
L1
AI assists
The person does the work and asks the AI when they need to.
L2
AI prepares, a person reviews
AI prepares a draft and a person edits it and signs it off.
L3
AI executes
AI does the task and only asks for approval at the agreed checkpoints.
L4
AI coordinates
Several agents run a whole process and your team deals with the exceptions.
RL3 Guard
Each workflow has a risk level from R0 to R4 that caps how much autonomy the AI gets. Tasks start at the most cautious level and move up when testing supports it. If something fails, a stop button hands the work back to your team without taking the service down.
1Input
We filter what comes in, including hidden instructions meant to manipulate the AI.
2Model and tools
The AI only uses the tools and data it has been cleared for.
3Output
We check what the AI is about to say or do before it reaches a customer or your systems.
4Human oversight
Sensitive or irreversible actions go through a person and are logged.
Reference frameworks
- EU AI Act, including AI literacy (Art. 4)
- ISO/IEC 42001
- NIST AI Risk Management Framework