Insights
How to Start AI Projects Without a Big Bang: An Iterative Engagement Model
Most AI programmes fail because they try to transform everything at once. A board mandate, a long roadmap, twelve use cases, and a vendor that promises “enterprise AI” in one wave. Six months later: demos, drift, and no clear owner for the next decision.
There is a better pattern. Run one short cycle on one priority workflow. Prove value with humans in the loop. Then either scale that workflow — or stop it cleanly — and move to the next area.
That is how we run AI consulting in Malaysia: not as a big bang, but as an iterative engagement model.
This approach is not invented in a vacuum. It is consistent with published best practice from standards bodies, hyperscalers, and global consulting leaders — adapted into a Malaysia-practical, one-workflow-at-a-time loop. Full links are in Sources & further reading below.
In short, the model borrows:
- CRISP-DM — business understanding and data reality before modelling
- NIST AI RMF and ISO/IEC 42001 — risk and management discipline throughout
- Microsoft and Google MLOps — pilots that can become monitored production
- McKinsey Rewired and Accenture AI Refinery / Industrial AI themes — operating model, adoption, and industrialised delivery — not demos alone
What follows is the practical loop Malaysian teams can run one area at a time.
From CRISP-DM to one workflow at a time
CRISP-DM is the classic six-phase data mining cycle: business understanding → data understanding → data preparation → modeling → evaluation → deployment, with data at the centre and an outer loop that repeats. The diagram below is an original illustration of that standard process (not a copy of IBM/SPSS artwork).
CRISP-DM process cycle (illustrative). Outer arrows = iterate; centre = data reality before you scale models.
Our engagement model does not replace CRISP-DM. It packages it for consulting delivery so you run one workflow per cycle instead of a big-bang programme:
| CRISP-DM phase | Oxydata stage |
|---|---|
| Business understanding | Discover |
| Data understanding (+ early prep / quality) | Ready |
| Data preparation + modeling design | Design |
| Modeling + evaluation in a live team | Pilot |
| Evaluation decision (go / iterate / stop) | Decide |
| Deployment (or next business question) | Scale or Stop → next cycle |
flowchart LR
BU[Business understanding] --> DU[Data understanding]
DU --> DP[Data preparation]
DP --> MO[Modeling]
MO --> EV[Evaluation]
EV --> DE[Deployment]
EV -.-> BU
BU -.-> D1[Discover]
DU -.-> D2[Ready]
DP -.-> D3[Design]
MO -.-> D4[Pilot]
EV -.-> D5[Decide]
DE -.-> D6[Scale or Stop]
Why Big Bang Fails
A typical big-bang sequence:
- Leadership wants “AI transformation.”
- A long list of use cases is scored in a workshop.
- Budget is locked for a multi-workstream programme.
- Data, process, and change management lag behind the demos.
- When one workstream stalls, the whole programme looks like a failure.
Big bang optimises for announcement, not for learning. Iterative delivery optimises for evidence: what worked, what did not, and whether to continue.
Enterprises can still run two or three cycles in parallel — but only if each has its own owner and success metric. SMEs should run one cycle at a time.
The Loop: One Workflow Per Cycle
1 Discover → 2 Ready → 3 Design → 4 Pilot → 5 Decide → 6 Scale or Stop
↑__________________________________________|
Repeat the cycle for the next department or use case. Same method. New problem.
Stage 1 — Discover (1–2 weeks)
Goal: One clear problem, one owner, one success metric.
| Do | Don’t |
|---|---|
| Map the current process (who, tools, handoffs, pain) | Boil the ocean with forty use cases |
| Name one target workflow | Pick “AI for the whole company” |
| Define success in business terms (time, error rate, conversion, cost) | Define success as “we used GPT” |
| Confirm a sponsor who can decide | Run without an accountable owner |
Exit criteria: Written problem statement + success metric + owner.
Stage 2 — Ready (1–3 weeks)
Goal: Know whether data and process can support AI for this use case.
Most AI failures are data readiness failures. Check completeness, definitions, access, and PDPA constraints before you spend on models.
| Check | Why |
|---|---|
| Data exists, is accessible, and is good enough for a pilot | Models amplify mess |
| Definitions are clear (what counts as a lead, open role, active customer) | Ambiguity becomes “hallucination” later |
| PDPA / access / retention constraints noted | Malaysia reality, not an afterthought |
| Human review step identified | HITL from day one |
For deeper audits, use StringRay — Data Quality Audit or the free AI data readiness scorecard.
Exit criteria: Go / reshape / stop — with reasons. No silent “we’ll fix data later.”
Stage 3 — Design (a few days to 1 week)
Goal: The smallest pilot that can prove the Stage 1 metric.
| Decision | Practical rule |
|---|---|
| Scope | One workflow, one team, limited volume |
| Approach | Classical ML vs GenAI vs rules — choose for the job |
| Human in the loop | Where people approve, override, or sample-check |
| Build vs buy vs partner | Speed, IP, and risk — not dogma |
| Risk | Hallucination, bias, leakage — mapped lightly (NIST-style) |
Exit criteria: Pilot charter — in scope, out of scope, timeline, and stop rules.
Stage 4 — Pilot (2–6 weeks)
Goal: Production-like learning, not a cherry-picked demo.
| Practice | Meaning |
|---|---|
| Real users | The team that owns the process |
| Real or safe sample data | Not only perfect examples |
| Measure vs baseline | Before/after on the Stage 1 metric |
| Log failures | Edge cases become the backlog |
| Human in the loop | AI recommends; people decide where it matters |
Exit criteria: Evidence pack — what worked, what did not, cost, risk, and adoption.
Stage 5 — Decide (one workshop)
Three honest outcomes only:
- Scale — widen users, harden the pipeline, add monitoring (MLOps)
- Iterate — same use case; fix data, process, prompts, or model; run another short cycle
- Stop — kill it and free budget for the next area
Big-bang programmes rarely allow a clean stop. This model does — and that is a feature.
Stage 6 — Scale or next area
If scaling: ownership, runbooks, training, monitoring, and PDPA controls.
If not: park the learnings, pick the next priority workflow, restart at Discover.
How This Maps to Oxydata Engagements
On the AI Consulting page we describe two entry points that fit this loop:
- AI Discovery Sprint — Stages 1–2 (and light Design): process map, data readiness, priority use cases, recommended next steps
- Full Strategy & Roadmap — deeper readiness, ROI, governance, and a sequenced path from pilot to scale
Delivery of the pilot itself then follows Stages 3–6 — still one priority area at a time, then the same cycle for the next.
A One-Liner You Can Take to Leadership
We don’t do big-bang AI. We run short cycles: Discover → Ready → Design → Pilot → Decide → Scale or Stop. One workflow per cycle. Then we move to the next.
What to Ask Any Consulting Vendor
If you are choosing a partner (including us), ask:
- Do you start with one workflow — or a transformation programme?
- What is the exit criteria for discovery before build spend?
- How do you handle data readiness and PDPA?
- Where is the human in the loop?
- Can we stop a pilot with evidence, not politics?
For the full buyer’s checklist, see How to choose an AI consulting vendor in Malaysia. For a market map of firms, see Top AI consulting firms in Malaysia (2026).
Conclusion
Agile delivery matters — short cycles and feedback — but “Agile methodology” alone is not enough for AI. You need business clarity, data readiness, a bounded pilot, risk thinking, and a forced decide gate.
Run the loop. One area at a time. Use the global frameworks below for discipline; use the Malaysian context for PDPA, owners, and real processes. That is how AI projects start without wasting the first budget.
Sources & further reading
These are the public references behind the engagement model. Oxydata’s loop is our practical adaptation for Malaysian enterprises and SMEs — not a rebrand of any single firm’s proprietary method.
Standards and classic process
- NIST AI Risk Management Framework (AI RMF) — Govern, Map, Measure, Manage
- ISO/IEC 42001 — Artificial intelligence management system
- CRISP-DM — business understanding and data preparation before modelling. The cycle diagram in this article is an original Oxydata illustration of that open standard process, not IBM/SPSS product artwork.
Hyperscaler MLOps (pilot → production → monitor)
- Microsoft Azure — Machine learning operations (MLOps)
- Microsoft Azure Architecture — MLOps v2
- Google Cloud — Practitioners Guide to MLOps
- Google Cloud — MLOps: continuous delivery and automation pipelines
Global consulting leaders (pilot-to-scale / operating model)
- McKinsey — Rewired: The McKinsey Guide to Outcompeting in the Age of Digital and AI and Rewired to outcompete — digital and AI transformation that works requires capability building, not technology alone
- Accenture — AI Refinery and Industrial AI — industrialising AI from concept toward scaled delivery
Use these as benchmarks for how serious programmes are run. Use Oxydata’s six-stage loop as the executable path for one workflow at a time.
Oxydata Software helps Malaysian enterprises and SMEs start AI the right way — Discovery Sprints, data readiness with StringRay, and iterative delivery from pilot to production. Talk to us about AI consulting or explore AI Consulting Malaysia.