Why a lifecycle is necessary
Enterprise AI combines deterministic software with probabilistic model behaviour. Data changes, retrieval fails, tools have side effects, permissions matter and a release can technically succeed while users ignore it. Delivery must coordinate business evidence, software engineering, data, evaluation, security, operations and adoption.
Google Cloud’s production guidance describes interacting components that require versioning, CI/CD, evaluation and monitoring. NIST’s AI Risk Management Framework treats risk management as continuous across design, development, deployment, use and evaluation. KDigital’s eight-stage sequence is our operating model built around those needs—not a sequence claimed by either external source.
Google Cloud production guidance · NIST AI RMF
Lifecycle overview
| Stage | Decision question | Minimum evidence | Gate |
|---|---|---|---|
| 1. Discover | Are we solving the right problem? | Stakeholders, workflow, baseline | Problem accepted |
| 2. Qualify | Is AI valuable, feasible and responsible? | Value, data readiness, risk | Invest / reshape / stop |
| 3. Design | What workflow and architecture govern delivery? | Target process, architecture, evidence plan | Design approved |
| 4. Prototype | Do the riskiest assumptions hold? | Thin slice, golden dataset, baseline | Assumptions validated |
| 5. Build | Is the system production-shaped? | Tested code, integrations, controls | Build complete |
| 6. Deploy | Can release be controlled and reversed? | UAT, runbook, rollback, approval | Release authorised |
| 7. Adopt | Are real users changing real work? | Enablement, usage, feedback | Adoption viable |
| 8. Optimize | Are quality, cost and impact improving? | Scorecards, traces, KPI review | Continue / scale / retire |
Discover
Understand the operational problem before selecting a model, agent or vendor.
Required evidence
Discovery brief, stakeholder map, current workflow, system and data inventory, KPI baseline, problem statement and assumption log.
Exit test
Can an independent reviewer explain whose problem is being solved and how it is measured without seeing a demo?
Qualify
Decide whether AI is the right intervention and whether the organisation can deploy it responsibly.
Required evidence
Opportunity scorecard, business case, data readiness, risk register, option analysis and go/no-go recommendation.
Exit test
Would the team still make this investment if the model or vendor changed tomorrow?
Design
Define target workflow, system boundaries, controls and evidence before implementation expands.
Required evidence
Target workflow, architecture, threat model, evaluation plan, backlog, acceptance criteria and release strategy.
Exit test
Can engineering, security, operations and business owners state what they require before release?
Prototype
Test the riskiest business and technical assumptions with the smallest useful end-to-end slice.
Required evidence
Working slice, dataset card, baseline scorecard, traces, failure taxonomy and assumption review.
Exit test
Did the prototype reduce uncertainty, or merely create presentation confidence?
Build
Convert the validated slice into a secure, integrated and maintainable production-shaped system.
Required evidence
Tested repository, API contracts, integration and access-control tests, CI/CD, documentation and updated risk register.
Exit test
Can another engineer deploy, operate and modify it without undocumented hero knowledge?
Deploy
Release through a controlled, observable and reversible process.
Required evidence
UAT, release scorecard, approvals, deployment record, dashboard, runbook, rollback proof and communication plan.
Exit test
Can the team detect harmful regression, stop unsafe behaviour and restore acceptable service?
Adopt
Make the system part of a trusted real-world workflow.
Required evidence
Enablement, champion plan, adoption dashboard, feedback log, support analysis and workflow changes.
Exit test
Are users completing the intended task, or bypassing, correcting or avoiding the system?
Optimize
Improve quality, reliability, cost and value while controlling change.
Required evidence
KPI report, expanded evaluation, reliability and cost review, risk update, backlog and scale or retire decision.
Exit test
Can the team show measured improvement without weakening safety, reliability or user control?
Six cross-cutting evidence tracks
- Business: problem, baseline, value and adoption.
- Technical: architecture, code, integrations and reliability.
- Data and AI: provenance, retrieval, model behaviour and evaluation.
- Security and governance: risk, privacy, access, oversight and audit.
- Operational: environments, release, observability, support and cost.
- Delivery: decisions, scope, ownership, communication and learning.
Decision-gate rules
Every gate needs a named accountable owner, defined evidence, explicit options and a recorded decision. A gate can continue, reshape, constrain, pause, release, roll back, scale or stop the work. Missing evidence is not the same as passing with low confidence.
Learn to own the evidence chain.
The Forward Deployed AI Engineer programme uses this lifecycle across sixteen integrated modules and a production-shaped deployment practicum.
Review the FDE programme →