From PoC to Production: The Four-Phase Model That Stops AI Pilots Dying in Procurement

From PoC to Production: The Four-Phase Model That Stops AI Pilots Dying in Procurement

Published: 20268 min readSummarize in ChatGPT

Valus Team

Contributor · Valus Journal

Most AI pilots never reach production — not because the model failed, but because nobody planned the operating model. A four-phase Discover → Design → Build → Operate playbook for enterprise AI.

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 touch

FAQ

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.

Ready to talk?

Let's build something worth shipping.

Get in touch