TL;DR
Around 70% of enterprise AI pilots never reach production. The model is rarely the problem — operating model, integration, governance and change management are. The Valus four-phase approach — Discover → Design → Build → Operate — is built to kill bad ideas early and ship good ones with a real owner, real SLA and real ROI.
Phase 1 — Discover (2–4 weeks)
Goal: prove the business case, not the model.
- Map the process today: volume, cost, cycle time, error rate.
- Quantify the prize: what is a 20%, 50%, 80% improvement worth?
- Identify the decision-maker, owner and budget line for production.
- Surface integration realities: source systems, data quality, identity, security.
- Kill criteria: if the prize is <USD 500k/year or the owner is unclear, stop now.
Deliverable: a one-page business case, signed by the budget owner.
Phase 2 — Design (3–6 weeks)
Goal: design the end state, not the demo.
- Solution architecture (data, model, app, integrations, observability).
- Operating model: who runs it, who fixes it, who owns the KPIs.
- Governance: data residency, model risk, audit, escalation.
- Change plan: training, comms, incentives, sunset of the old process.
- Pricing and TCO over 3 years.
Deliverable: a target-state blueprint and a phased delivery plan.
Phase 3 — Build (6–16 weeks)
Goal: ship to production, not to a sandbox.
- Build in vertical slices — each slice runs end-to-end on real data.
- Integrate with real source systems from week one; no mocks past sprint two.
- Wire observability and evaluation before launch, not after.
- Run shadow mode against the human process for at least 2 weeks.
- Get sign-off from security, compliance and the business owner — written.
Deliverable: a production system, instrumented, with a runbook.
Phase 4 — Operate (ongoing)
Goal: keep it valuable, not just alive.
- SLOs for latency, accuracy, availability and cost.
- Monthly model and prompt review against fresh data.
- Drift detection and retraining cadence.
- Quarterly business review against the original ROI case.
- Continuous backlog of next-slice improvements.
Deliverable: a working product with a quarterly P&L conversation.
Why pilots die — and how this prevents it
| Common failure | Phase that prevents it |
| No production budget | Discover (kill if no owner) |
| Integration discovered late | Design (architect early) |
| Security blocks at launch | Design + Build (engage early) |
| Model degrades silently | Operate (SLOs + drift) |
| Users don't adopt | Design + Operate (change plan) |
| ROI never proven | Discover + Operate (baseline + QBR) |
The honest takeaway
Great AI in production is 80% operating model, 20% model. Pick a partner who treats it that way.
Take the next step
Thinking about taking an AI pilot to production? Valus runs Discover → Design → Build → Operate for enterprises across MENA — from Arabic NLP to computer vision to AI voice. We kill bad ideas early and ship good ones with real SLAs. Book a working session:
Get in touchFAQ
How long is a typical end-to-end engagement?
Discover (2–4 wks) → Design (3–6 wks) → Build (6–16 wks) → Operate (ongoing). Most enterprise programmes hit production in 4–6 months.
Can we skip Discover if we already have a pilot?
You can compress it, but don't skip the business-case and owner check. That's why most pilots die.
Who owns the model in Operate?
A named product owner on the business side, supported by an MLOps function. Shared ownership without a named owner is no ownership.
How do you handle data residency?
Designed in Phase 2 — UAE-resident infra, region-pinned models, audit trail. Not retrofitted later.
What's the biggest mistake?
Treating the pilot as the destination. The pilot is a hypothesis test; production is the product.



