Illustrative example. Norval Components is a fictional company. The people, situation and figures are invented to show the format and depth of a Sprint deliverable. They are not a real client result.
One illustrative case: a manufacturer weighing a €1.6M ERP/MES replacement. The same method applies to any significant system.
ARCHITECTURE & AI SPRINT · DECISION DOCUMENT · ILLUSTRATIVE
Norval Components: should we commit €1.6M to a new ERP and MES?
11 min read
Part A — The decision document
Decision on a page #
Takeaway: Don't sign the €1.6M contract now; fund a €380k–€520k staged programme.
Decision requested: Approve a €1.6M ERP + MES replacement (vendor quote, 24–30 months).
Our recommendation: Don't approve it now. Approve a €380k–€520k staged programme instead: a data and integration layer, shop-floor data capture and two targeted AI use cases. Re-decide the ERP at month 15 using real data.
- Capital deferred€1.0M–€1.3Mnot committed now
- 3-year TCO vs. rebuild€0.8M–€1.6M lower
- Time to first value12 weeks(live shop-floor dashboard, plant 1)
- First go/no-go gateMonth 5no more than €300k committed before Gate 1
- ConfidenceMedium–highMain uncertainty: ERP vendor API support, to be confirmed in week 1
All figures are illustrative ranges for a fictional company. All figures are planning estimates from a 2-week review, with an accuracy of ±20–30%, to be refined at each gate.
Read in 30 seconds
- 1.Don't approve the €1.6M ERP + MES now; fund a €380k–€520k staged programme instead.
- 3.Option C scores 4.2, against 3.6 for B and 1.5 for A.
- 5.First go/no-go gate at month 5, with no more than €300k committed before Gate 1.
- 6.3-year TCO €1.2M–€1.4M, against €2.2M–€2.8M for A.
- 8.Two AI use cases, both tied to money.
- 9.Five actions for Monday.
The question the board asked #
Takeaway: A 17-year-old ERP, spreadsheets on the shop floor and a €1.6M quote on the table.
“Our ERP is 17 years old, our two plants run on spreadsheets between machines and the office, and a vendor is offering us a full ERP + MES replacement for €1.6M. Customers want quotes faster and OEMs are asking for traceability. Is this the right investment, and where does AI actually help us?”
Norval Components (fictional) is a mid-sized maker of machined and stamped metal parts for automotive and industrial OEMs: two plants, around 650 people and about €95M in revenue. Quotes take 4–6 working days. Production status is collected by hand at the end of each shift. Scrap is around 4.5% (estimated from internal reports).
What we did in two weeks: 10 interviews (board, plant managers, sales, quality, IT, two key customers) · review of 6 key systems and one data sample · vendor quote teardown · 3 options costed · 1 recommendation.
Options compared #
Takeaway: C scores 4.2 out of 5: it fixes the data now and keeps A as an option.
Scores run from 1 (poor) to 5 (best). A higher score is always better, so “Risk 5” means lowest risk.
| Criterion (weight) | A. Full ERP + MES replacement | B. Keep everything, add AI on top | C. Data layer + shop-floor data + targeted AI (recommended) Recommended |
|---|---|---|---|
| Cost (25%) | 1€1.4M–€1.8M capex | 5€150k–€250k | 4€380k–€520k |
| Time to value (20%) | 118+ months | 43–4 months | 412 weeks |
| Delivery risk (25%) | 1two-plant cutover | 3low build, weak results | 4staged, gated |
| Lock-in (15%) | 2single vendor suite | 3AI tied to legacy data | 4owned data layer |
| AI leverage (15%) | 3later, after go-live | 2AI on dirty data | 5clean data by month 5 |
| Weighted score | 1.5 | 3.6 | 4.2 |
C. Data layer + shop-floor data + targeted AI (recommended) Recommended
- Cost (25%)4€380k–€520k
- Time to value (20%)412 weeks
- Delivery risk (25%)4staged, gated
- Lock-in (15%)4owned data layer
- AI leverage (15%)5clean data by month 5
Weighted score: 4.2
B. Keep everything, add AI on top
- Cost (25%)5€150k–€250k
- Time to value (20%)43–4 months
- Delivery risk (25%)3low build, weak results
- Lock-in (15%)3AI tied to legacy data
- AI leverage (15%)2AI on dirty data
Weighted score: 3.6
A. Full ERP + MES replacement
- Cost (25%)1€1.4M–€1.8M capex
- Time to value (20%)118+ months
- Delivery risk (25%)1two-plant cutover
- Lock-in (15%)2single vendor suite
- AI leverage (15%)3later, after go-live
Weighted score: 1.5
How we scored: each option is rated 1–5 on cost, time to value, delivery risk, vendor lock-in and AI leverage, weighted by what the board said matters most (cost and delivery risk 25% each, time to value 20%, vendor lock-in and AI leverage 15% each). Scores are judgements from the Sprint's evidence, not measurements.
B is cheaper but doesn't fix the data, so the AI underdelivers. A fixes everything, but too late and at too high a risk. C fixes the data now and keeps A as an option.
Target architecture #
Takeaway: Keep the ERP, build an owned data layer in the middle, buy the shop-floor kit.
Keep the ERP as the system of record for finance and orders. Put an owned data and integration layer in the middle, so every new capability plugs into clean data instead of into the ERP.
Build = owned and built · Buy/configure = standard product · Keep = existing system, integrated
Roadmap: 18 months, three phases, three gates #
Takeaway: Three phases, three gates. If G1 fails, you stop at €300k or less.
| Phase | Months | What ships | Cost | Gate (go/no-go) |
|---|---|---|---|---|
| 1. Data foundation | 1–5 | Integration layer, master data clean-up, live dashboard in plant 1 by week 12 | €220k–€300k | G1, month 5: ERP integration stable; ≥90% of plant 1 machines reporting; part master ≥95% clean |
| 2. Shop floor | 5–10 | Plant 2 roll-out, operator tablets, scrap and downtime reasons | €100k–€140k | G2, month 10: scrap reasons captured on ≥80% of shifts; baseline OEE agreed; machine data ready for phase 3 alerts |
| 3. Targeted AI | 8–12 | Quote assistant (pilot with 3 estimators), quality alerts | €60k–€80k | G3, month 12: quote lead time ≤2 days on pilot; estimator acceptance ≥70% |
| ERP decision | 15–18 | Re-decide the ERP with 12 months of real data | Estimated then | Board decision, month 15 |
Total for phases 1–3: €380k–€520k. The phases overlap from months 8–10 by design. Phase 3 quality alerts depend on phase 2 machine data, checked at G2.
If G1 fails, stop. You've spent at most €300k, and the data clean-up is still worth having.
Cost model #
Takeaway: C's 3-year TCO is €1.2M–€1.4M, against €2.2M–€2.8M for A.
Assumption: the current ERP and spreadsheets cost €220k a year to run (licences, support, internal IT time).
| A. Full replacement | B. AI on top | C. Recommended | |
|---|---|---|---|
| Capex | €1.4M–€1.8M | €150k–€250k | €380k–€520k |
| Opex per year | €270k–€320k | €260k–€280k | €270k–€290k |
| 3-year TCO | €2.2M–€2.8M | €0.9M–€1.1M | €1.2M–€1.4M |
| Capital at risk before first proof | €0.7M–€1.1M | €150k–€250k | €220k–€300k |
C costs about €0.1M–€0.5M more than B over three years. That buys clean data, which is what makes the AI work. If you later still need A, the data layer carries over, so the C spend isn't wasted. Assumes A's first spend milestone is about 50–60% of capex.
Value side, deliberately cautious: quote lead time goes from 4–6 days to 1–2. Each 1 point of scrap reduction is worth about €0.5M–€0.6M a year at Norval's material spend (estimate, to be measured in phase 2). We don't book any value until G2 has a measured baseline.
Risk register (top 5) #
Takeaway: Five risks, each with a named owner and a mitigation.
| # | Risk | Likelihood | Impact | Mitigation | Owner |
|---|---|---|---|---|---|
| 1 | ERP vendor limits API access or charges for it | Medium | High | Written confirmation and pricing in week 1; fallback to database-level read integration | CIO |
| 2 | Part master data worse than sampled | High | Medium | 6-week clean-up sprint in phase 1; G1 threshold ≥95% | Ops director |
| 3 | Operators don't log scrap reasons | Medium | High | Two-tap tablet design; shift leads co-design; tracked at G2 | Plant managers |
| 4 | Estimators distrust the AI quote drafts | Medium | Medium | Human approves every quote; pilot with 3 estimators; acceptance KPI | Sales director |
| 5 | Integration layer becomes “shadow IT” with no owner | Low | High | Named product owner; documented APIs; 20% of build time on handover | CIO |
Risks of the recommended option (C)
- Ownership: the data layer needs a named data owner in each plant and 0.5–1 FTE internally or a support contract; without them, master data degrades and G1 fails.
- OT security: connecting machines to the data layer exposes the shop-floor network, so it needs network segmentation plus an OT security review before phase 2.
- Database-level fallback: if API access fails, database-level access is read-only and only a stopgap; check the ERP licence first, as it may not allow it.
- Integration upkeep: ERP or MES upgrades can break interfaces; budget for maintenance and regression tests in opex.
AI: where it pays and where it doesn't #
Takeaway: Two AI use cases tied to money; the rest waits or doesn't happen.
| Use case | Value | Data ready? | Verdict |
|---|---|---|---|
| Quote assistant (RFQ + drawings → draft quote) | High: faster quotes win orders | Yes, 8 years of past jobs | Do: phase 3 |
| Quality anomaly alerts from machine data | High: less scrap | Phase 2, unlocks Phase 3 alerts | Do: phase 3 |
| Predictive maintenance | Medium | No, too little failure history | Later: collect data first |
| Demand forecasting | Medium | Partly; OEM call-offs already drive the plan | Later |
| Generic chatbot for customers or staff | Low | n/a | Don't |
| “AI strategy” platform licence | Unclear | n/a | Don't buy before a use case proves value |
Two AI use cases, both tied to money. Everything else waits for data or doesn't happen at all.
What we'd do on Monday #
Takeaway: Five actions the board can take next week.
- Pause signature on the €1.6M contract and tell the vendor the decision moves to month 15.
- Send the ERP vendor the API access request (template in appendix B).
- Name a product owner for the data layer: an internal person, not a supplier.
- Start the part master clean-up in plant 1, using the 300 highest-volume part numbers first.
- Book Gate 1 into the board calendar for month 5.
Assumptions and evidence
Show the 14 assumptions and evidence behind the figures
| Item | Source | Status |
|---|---|---|
| ERP is 17 years old | Systems review | Client |
| Two plants, around 650 people, about €95M revenue | Company data | Client |
| Quotes take 4–6 working days | Sales interviews | Client |
| Scrap around 4.5% | Quality reports | Client |
| Current ERP and spreadsheets cost €220k a year to run | IT and finance | Client |
| Vendor quote: €1.6M, 24–30 months | Vendor proposal | Client |
| Capex and opex ranges for options A–C | Costed in the Sprint; illustrative range | Estimate |
| 3-year TCO by option | Cost model; illustrative range | Estimate |
| 8 years of past jobs available for quoting | Systems review | Client |
| ERP vendor API access and pricing | Confirm in week 1 | To validate |
| Part master can reach ≥95% clean in phase 1 | Data sample; tested at G1 | To validate |
| Baseline OEE | Agreed at G2 | To validate |
| 1 point of scrap reduction is worth €0.5M–€0.6M a year | Material spend; measured in phase 2; illustrative range | Estimate |
| Estimator acceptance of AI drafts ≥70% | Pilot; tested at G3 | To validate |