Methodology

Methodology Beats Tooling, and Discipline Beats Methodology

A four-phase engagement model — Discovery, Architecture, Governance, Oversight — with explicit deliverables, decision artefacts and timeboxes per phase. Built for organizations that need AI to work in production, not in demos.

Most AI Programs Fail in the Same Three Places

In a decade of enterprise AI work the failure pattern is consistent: scope unclear at kickoff, architecture chosen before requirements stabilize, governance retrofitted after the first incident. The methodology below sequences those three concerns deliberately — discovery before architecture, architecture before governance, governance before implementation. Every phase has an exit criterion; you do not advance because the calendar says so.

🎯

Scope First

A precise scope is the most expensive deliverable. Most engagements that fail did so because the scope was hand-waved at week zero.

🏗️

Architecture Before Code

The architecture decision freezes more variables than any other choice in the engagement. It deserves more attention than it usually gets.

📋

Governance Before Deployment

The cost of designing governance into a system is roughly 10% of the cost of bolting it on after an incident. Sequence accordingly.

Discovery — Decide What the AI Is For

Typical duration: 2-3 weeks. Engagement-ending exit criterion: scope and risk-classification document accepted by the named accountable owner.

What we examine

  • Which decision class the AI must inform
  • Existing baseline metric for that decision
  • Data available, with provenance and access constraints
  • Regulatory regime that applies (EU AI Act tier, sector regulation)
  • Stakeholders, named accountable owner candidate
  • Operating-model implications of accepting AI output

What you receive

  • Scope statement: decision class, intended user, residual human authority
  • Baseline metric definition with measurement method
  • Risk classification under the applicable regulatory framework
  • Data-availability assessment
  • Go / no-go recommendation with explicit reasoning

How we know it worked

The scope statement passes the "describe in one sentence" test: a stakeholder outside the project can read it and accurately predict what the AI will and will not do. If not, Discovery is not done.

Roughly one in five engagements stops at the end of Discovery because the answer the data tells us is "this AI deployment will not produce the value the sponsor expected." That outcome saves the engagement budget by an order of magnitude.

Architecture — Design the System That Will Actually Get Built

Typical duration: 3-4 weeks. Exit criterion: architecture document approved by engineering leadership, proof-of-concept demonstrably meets baseline metric.

Decisions taken in this phase

  • Model strategy: foundation model vs domain-tuned vs RAG (see AI-FAQ for the decision framework)
  • Deployment topology: SaaS, private tenant, on-premise, hybrid
  • Data architecture: retrieval sources, vector store, refresh strategy
  • Integration surface: which existing systems the AI reads from and writes to
  • Evaluation harness: how accuracy and behavior will be measured during build

Deliverables

  • Architecture document with component diagrams and data flow
  • Proof-of-concept demonstrating the chosen approach against representative data
  • Evaluation results vs the Phase 1 baseline metric
  • Cost model: build, infrastructure, ongoing operations
  • Risk register specific to the chosen architecture

How we know it worked

The proof-of-concept matches the baseline metric (or beats it) on data the engineering team has not previously seen. The cost model is defensible at executive review. The architecture document is detailed enough that an external engineering team could build the system from it without ambiguity.

Governance Design — Build Controls Before Deployment, Not After

Typical duration: 2-3 weeks, often parallel to late Phase 2. Exit criterion: governance package signed off by accountable owner and (where relevant) risk/legal.

What we design

  • Twelve-control baseline (see AI-Governance) tailored to this system
  • Risk-classification mapping under the applicable regulatory framework
  • Monitoring metrics with alerting thresholds
  • Incident-response procedure with severity definitions
  • Rollback / disable mechanism design
  • Per-system review cadence

Deliverables

  • Governance package: controls documented per the baseline
  • Inventory entry with full classification
  • Monitoring spec: metrics, thresholds, alerting paths
  • Incident-response runbook
  • Named accountable owner with documented authority

How we know it worked

The governance package answers every question a regulator or auditor would ask without anyone needing to remember context. The rollback mechanism has been demonstrated in a non-production environment. The accountable owner accepts the system in writing with the controls documented.

Implementation Oversight — Steer Without Building

Typical duration: 8-16 weeks (matches your engineering team's build timeline). Exit criterion: production-readiness sign-off across architecture, governance and operational dimensions.

The model

Implementation is done by your engineering team or a delivery partner — not by us. Slavin AI's role in Phase 4 is to ensure the system that gets built is the system that was designed. We review code-level deliverables against the architecture, run governance checkpoints at predefined gates, and certify production readiness when all controls are demonstrably in place.

Checkpoints

  • Architecture conformance review (typical: week 4 of build)
  • Governance gate 1: data and access controls (typical: week 6)
  • Evaluation gate: metric performance vs Phase 2 PoC (typical: week 10)
  • Governance gate 2: monitoring and incident response (typical: week 12)
  • Production readiness review (final)

Outputs

  • Per-checkpoint sign-off or remediation requirements
  • Production-readiness sign-off with explicit residual risks
  • Handover package to the operational team
  • Optional: 90-day post-deployment monitoring oversight

Three Ways We Engage

Same methodology, different role for us in the program.

Advisory

We run Phase 1 and Phase 3 (Discovery and Governance Design). Phase 2 is reviewed but not led by us. Phase 4 oversight is light-touch.

Best for: organizations with strong internal architecture capability who need an independent governance and risk perspective.

Typical duration: 6-10 weeks of active engagement, then quarterly review cadence.

Consultancy

We run Phases 1-3 in full and provide Phase 4 oversight throughout the build. Engineering execution stays with your team or partner.

Best for: organizations new to enterprise AI or facing a regulated deployment where architectural and governance discipline are non-negotiable.

Typical duration: 4-8 months from kickoff to production readiness.

Interim AI Architect

Acts in an interim role inside your organization — head of AI engineering, chief AI architect, or equivalent — for 6-12 months while you build the permanent function.

Best for: organizations standing up an AI capability who need senior leadership in seat immediately while permanent hiring runs.

Typical duration: 6-12 months, with explicit handover plan.

How This Methodology Connects

AI Governance

The maturity model, the twelve-control baseline and the LLM risk taxonomy the Phase 3 governance design builds on.

Read the AI Governance page →

Enterprise AI FAQ

The decision frameworks referenced in Phase 2 architecture choices: build vs buy, RAG vs fine-tuning, deployment topology, ROI measurement.

Read the AI FAQ →

Case Studies

Engagements where this methodology was applied — healthcare RAG, financial-services compliance assistant, manufacturing document AI.

Browse Case Studies →

Why this methodology — instead of AI-only delivery

AI closes the first 80% of any project well. Production is decided by the last 20% — failure modes, concurrency, data integrity, behavior under load, security, cost at scale — which AI does not model. This four-phase methodology exists to apply the judgment that catches the architecturally plausible but operationally lethal decision before it ships.

Read the full position →

Start With the First Phase

An Architecture Review is a compressed pass through Phase 1 — two hours to identify scope, baseline metric, regulatory exposure and the highest-risk architecture decision.