Methodology

Not every AI idea deserves to be built. Those that do should not start unprepared.

Every use case passes our five-layer assessment framework. It is the shared language for every decision, from portfolio prioritisation to live operation. Each stage ends at a clear decision point: proceed, fix or stop.

01

Value

Why should this be done at all?

We start from the business case: problem, benefit, measurement and strategic fit are defined in numbers.

  • Problem: which business problem, how is it solved today, what does it cost?
  • Benefit: cost reduction, revenue, risk reduction, experience; which and how much?
  • Measurement: target KPIs, expected improvement, ROI and payback period
  • Strategic fit: aligned with company strategy, a management priority?
Duration1–2 weeks
ParticipantsBusiness owner, finance, sponsor
Tools and outputs
  • Scenario portfolio workshop and value card (problem, volume, cost, KPI)
  • Comparison with the closest scenario in the agent library
Red flags
  • Benefit without numbers: “efficiency will rise” but no volume or cost
  • No sponsor; the request comes from IT and the business watches

Output A business case that speaks in numbers and a prioritised scenario portfolio

02

Readiness

Do we have the inputs we need?

Even the best idea stalls without data and organisational readiness. We check four axes before starting.

  • Data: sources available and digital; quality, freshness and ownership clear?
  • Integration: systems and APIs ready; real-time access possible?
  • Organisation: does the business unit own it; sponsor and experts identified?
  • Technology: cloud/security infrastructure, AI platform and licences ready?
Duration1–2 weeks
ParticipantsData owner, integration team, process owner
Tools and outputs
  • Readiness scorecard: data, integration, process, organisation (each axis 0–2)
  • API and data-access test; sample data set
Red flags
  • The target system has no API, only a screen
  • Process rules live only in people’s experience

Output A readiness scorecard with gaps closed before the project; no surprises

03

Feasibility

Can we actually build this?

An honest engineering assessment: AI is not always the right answer. Classic software, classic ML, agents or generative AI; which one is needed?

  • Technical fit: should AI solve this at all; which approach?
  • Complexity: how many systems, data sources and users are affected?
  • Model performance: expected accuracy, impact of a wrong output, is human approval needed?
  • Project feasibility: MVP duration, team, external dependencies, technical risks
Duration2–3 weeks
ParticipantsAginies engineering, enterprise architecture, information security
Tools and outputs
  • Approach decision: classic software, classic ML, agents, generative AI
  • Prototype and draft test corpus; wrong-output impact analysis
Red flags
  • A wrong output is irreversible and no approval point is defined
  • Scope grows at every meeting

Output A realistic MVP plan with clear scope and identified risks

04

Operations

How will it live after go-live?

AI systems are run like products; after go-live, monitoring, maintenance and release management are part of the job.

  • Operating model: product owner, process owner, operations and support defined
  • User management: training, adoption tracking, feedback loops
  • Monitoring: performance, metrics and error management in real time
  • Financial operations: monthly run cost, cost projection with growth, budget ownership
Duration1 week (plan) + continuous after go-live
ParticipantsProduct owner, process owner, operations and support
Tools and outputs
  • Operating-model template: roles, thresholds, review sample, budget owner
  • Dashboards and the monthly review calendar
Red flags
  • Nobody owns operation after go-live
  • No budget owner for the monthly run cost

Output A living, measured, continuously improving AI product

05

Governance

Is it secure, controlled and auditable?

AI must meet the same standards as every other critical system. Governance is not bolted on; it is part of the architecture.

  • Security: sensitive/personal data map, data residency, control of data flowing to models, masking
  • Compliance: privacy impact, sector regulation, internal policy requirements
  • Risk: assessment of hallucination, wrong decisions, data leakage and prompt injection
  • Control and auditability: human approval points, automation limits, kill-switch, decision history
Duration2 weeks, in parallel with the other layers
ParticipantsInformation security, compliance, legal, internal audit
Tools and outputs
  • Data map and masking rule set; approval matrix
  • Control set and evidence list; sector regulation mapping
Red flags
  • Personal data involved, no map
  • Governance postponed as “we will add it later”

Output An audit-ready, controlled, trustworthy AI operation

Transformation journey

How we progress

  1. 1

    Discovery & portfolio

    Collect scenarios with business units, prioritise with the Value layer

  2. 2

    Assessment

    Readiness + Feasibility analysis of selected scenarios, MVP design

  3. 3

    MVP & validation

    Build on Aginies, pilot with real data, validate KPIs

  4. 4

    Production & scale

    Go live with the Operations + Governance layers, grow the portfolio

Worked example

Example: how the dialogue banking assistant passed the five layers

  1. 01
    Value · Why should this be done at all?

    Simple transactions reaching the call centre and cost per transaction put in numbers; target 40% reduction. Sponsor: digital banking.

  2. 02
    Readiness · Do we have the inputs we need?

    Transaction APIs and a test environment for core banking and cards existed; OTP stack ready; policy thresholds written with information security.

  3. 03
    Feasibility · Can we actually build this?

    Approach: supervisor + four specialist agents; per-transaction OTP and approval band against wrong-output risk; test corpus of 212 scenarios.

  4. 04
    Operations · How will it live after go-live?

    Product owner digital channels, process owner operations; monthly review and threshold updates on the calendar.

  5. 05
    Governance · Is it secure, controlled and auditable?

    Personal-data map and masking rules; limit increase > 60%, dispute > 5,000 and new payee to approval; all systems in Türkiye.

Watch the scenario live: Dialogue banking: voice and text transaction assistant →

Self-assessment

Would your scenario pass all five layers?

Apply the same framework to your own scenario. The result becomes the first thing we discuss in a discovery session.

Would your scenario pass all five layers?

10 questions, 3 minutes. Answers stay in your browser; bring the result to a discovery session if you like.

Time to move from experimenting with AI to transforming with it.

In a 30-minute discovery session we take your 2–3 priority business problems, show a live demo of a similar scenario, and draft a roadmap that starts with the Value layer.

Book a 30-minute discovery session → dahi@aginies.com
  1. Your 2–3 priority problems
  2. Live demo of a similar scenario
  3. Roadmap starting with value analysis