Seven phases from work to system.
The method is the product as much as the software is. Every Arvanord engagement — from a ready-made deployment to a $500k custom system — follows the same seven phases.
01Discover
People. Workflows. Documents. Software. Data. Problems. Objectives.
- Stakeholder and role mapping
- Current-state workflow interviews
- System and data inventory
- Objective definition with success criteria
02Map
Inputs, actions, decisions, exceptions, outputs, bottlenecks, human responsibilities.
- Inputs and outputs defined per step
- Decision points and rule sources
- Exceptions and failure modes
- Where judgement must remain human
03Architect
AI architecture, models, agents, knowledge, data, integrations, permissions, approvals, interfaces, security boundaries, evaluations.
- System architecture and data flows
- Model and knowledge strategy
- Permission and approval design
- Evaluation plan defined up front
04Forge
Engineering, integration and interface construction against the architecture.
- System build to the architecture
- Integrations with real systems
- Interfaces for real users
- Documentation as it is built
05Validate
Realistic workflows, edge cases, expected output, evaluation datasets, failure conditions, human review.
- Evaluation datasets from real cases
- Edge case and failure testing
- Human review cycles
- Acceptance against the specification
06Deploy
Rollout into production with training, monitoring and defined rollback paths.
- Production rollout plan
- User onboarding and training
- Monitoring and alerting active
- Fallback behaviour verified
07Evolve
Through Arvanord Ops: monitoring, model updates, new business rules, knowledge refresh, optimisation.
- Continuous monitoring
- Model and rule updates
- Knowledge base refresh
- Improvement roadmap
Engineering principles that apply to all seven phases.
Permissions, boundaries, approvals, logging, monitoring, fallback, evaluation, access control, versioning, testing, documentation, reliability.
Permissions
Access follows your organisation's model. Agents and assistants operate inside defined scopes.
Data boundaries
Data stays inside the boundary defined during architecture — per department, per system, per role.
Human approval
Consequential actions require a person. Approval design is part of the architecture, not an afterthought.
Logging
Actions and decisions are recorded — what the system did, when, and on which inputs.
Monitoring
Production systems are observed: quality sampling, drift detection, alerting.
Fallback behaviour
When the system is uncertain or unavailable, work routes to humans with context.
Evaluation
Behaviour is measured against specifications with datasets drawn from real work.
Access control
Authentication and roles follow existing IT policy and identity providers.
Versioning
Prompts, rules, models and knowledge change under version control.
Testing
Changes pass evaluation gates before deployment — regressions are caught, not shipped.
Documentation
Architecture, behaviour and operating procedures are documented for your team.
Reliability
Systems are engineered for operational load, failure and recovery.
About the method.
Why seven phases?
Because each phase produces something the next one needs: a discovery produces the workflow map; the map produces the architecture; the architecture produces the build plan. Skipping phases is how systems end up rebuilt.
Where does validation happen?
Continuously. Evaluations are defined in the Architect phase and run from the first build through deployment and beyond — against datasets drawn from realistic work.
What does the client need to provide?
Access, people and honesty about how work really happens — including the exceptions and the workarounds. That last part is what separates a system that ships from one that gets abandoned.
Phase 01 starts with a conversation.
Tell us how the work actually happens. We will map it, and tell you what should be engineered — and what should not.